economic systems research papers

Browse Econ Literature

• Working papers
• Software components
• Book chapters
• JEL classification

More features

• Subscribe to new research

RePEc Biblio

Author registration.

• Economics Virtual Seminar Calendar NEW!

Taylor & Francis Journals

Economic systems research.

• Publisher Info
• Serial Info

Corrections

Contact information of taylor & francis journals, serial information, impact factors.

• Simple ( last 10 years )
• Recursive ( 10 )
• Discounted ( 10 )
• Recursive discounted ( 10 )
• H-Index ( 10 )
• Euclid ( 10 )
• Aggregate ( 10 )
• By citations
• By downloads (last 12 months)

January 2024, Volume 36, Issue 1

October 2023, volume 35, issue 4, july 2023, volume 35, issue 3, april 2023, volume 35, issue 2, january 2023, volume 35, issue 1, october 2022, volume 34, issue 4, july 2022, volume 34, issue 3, april 2022, volume 34, issue 2, january 2022, volume 34, issue 1, october 2021, volume 33, issue 4, july 2021, volume 33, issue 3, april 2021, volume 33, issue 2, january 2021, volume 33, issue 1, october 2020, volume 32, issue 4, july 2020, volume 32, issue 3, april 2020, volume 32, issue 2, january 2020, volume 32, issue 1, october 2019, volume 31, issue 4, july 2019, volume 31, issue 3, april 2019, volume 31, issue 2, january 2019, volume 31, issue 1, october 2018, volume 30, issue 4, july 2018, volume 30, issue 3, april 2018, volume 30, issue 2, january 2018, volume 30, issue 1, october 2017, volume 29, issue 4, july 2017, volume 29, issue 3, april 2017, volume 29, issue 2, more services and features.

Follow serials, authors, keywords & more

Public profiles for Economics researchers

Various research rankings in Economics

RePEc Genealogy

Who was a student of whom, using RePEc

Curated articles & papers on economics topics

Upload your paper to be listed on RePEc and IDEAS

New papers by email

Subscribe to new additions to RePEc

EconAcademics

Blog aggregator for economics research

Cases of plagiarism in Economics

About RePEc

Initiative for open bibliographies in Economics

News about RePEc

Questions about IDEAS and RePEc

RePEc volunteers

Participating archives

Publishers indexing in RePEc

Privacy statement

Found an error or omission?

Opportunities to help RePEc

Get papers listed

Have your research listed on RePEc

Open a RePEc archive

Have your institution's/publisher's output listed on RePEc

Get RePEc data

Use data assembled by RePEc

Economic Systems Research

Subject Area and Category

• Economics and Econometrics

Publication type

09535314, 14695758

Information

How to publish in this journal

The set of journals have been ranked according to their SJR and divided into four equal groups, four quartiles. Q1 (green) comprises the quarter of the journals with the highest values, Q2 (yellow) the second highest values, Q3 (orange) the third highest values and Q4 (red) the lowest values.

The SJR is a size-independent prestige indicator that ranks journals by their 'average prestige per article'. It is based on the idea that 'all citations are not created equal'. SJR is a measure of scientific influence of journals that accounts for both the number of citations received by a journal and the importance or prestige of the journals where such citations come from It measures the scientific influence of the average article in a journal, it expresses how central to the global scientific discussion an average article of the journal is.

Evolution of the number of published documents. All types of documents are considered, including citable and non citable documents.

This indicator counts the number of citations received by documents from a journal and divides them by the total number of documents published in that journal. The chart shows the evolution of the average number of times documents published in a journal in the past two, three and four years have been cited in the current year. The two years line is equivalent to journal impact factor ™ (Thomson Reuters) metric.

Evolution of the total number of citations and journal's self-citations received by a journal's published documents during the three previous years. Journal Self-citation is defined as the number of citation from a journal citing article to articles published by the same journal.

Evolution of the number of total citation per document and external citation per document (i.e. journal self-citations removed) received by a journal's published documents during the three previous years. External citations are calculated by subtracting the number of self-citations from the total number of citations received by the journal’s documents.

International Collaboration accounts for the articles that have been produced by researchers from several countries. The chart shows the ratio of a journal's documents signed by researchers from more than one country; that is including more than one country address.

Not every article in a journal is considered primary research and therefore "citable", this chart shows the ratio of a journal's articles including substantial research (research articles, conference papers and reviews) in three year windows vs. those documents other than research articles, reviews and conference papers.

Ratio of a journal's items, grouped in three years windows, that have been cited at least once vs. those not cited during the following year.

Evolution of the percentage of female authors.

Evolution of the number of documents cited by public policy documents according to Overton database.

Evoution of the number of documents related to Sustainable Development Goals defined by United Nations. Available from 2018 onwards.

Leave a comment

Name * Required

Email (will not be published) * Required

* Required Cancel

The users of Scimago Journal & Country Rank have the possibility to dialogue through comments linked to a specific journal. The purpose is to have a forum in which general doubts about the processes of publication in the journal, experiences and other issues derived from the publication of papers are resolved. For topics on particular articles, maintain the dialogue through the usual channels with your editor.

Follow us on @ScimagoJR Scimago Lab , Copyright 2007-2024. Data Source: Scopus®

Cookie settings

Cookie Policy

Legal Notice

Privacy Policy

Featured Topics

View the  Entrepreneurship Working Group  page.

Econometrics

Microeconomics, macroeconomics, international economics, financial economics, public economics, health, education, and welfare, labor economics, industrial organization, development and growth, environmental and resource economics, regional and urban economics, more from nber.

In addition to working papers , the NBER disseminates affiliates’ latest findings through a range of free periodicals — the NBER Reporter , the NBER Digest , the Bulletin on Retirement and Disability , the Bulletin on Health , and the Bulletin on Entrepreneurship  — as well as online conference reports , video lectures , and interviews .

Economic Systems Research

( API-Link )

Impact Factor : 1.800 (based on Web of Science 2023)

• # 125 / 348 (Q2) in Economics

Partner: • University Press Alert

Economic Systems and Institutions

• December 2022

• University of Miskolc

Discover the world's research

• 25+ million members
• 160+ million publication pages
• 2.3+ billion citations

• Fem Leg Stud

• Daniel Newman
• Carole Pateman
• Lucas Fiuza

• Emanuel International Monetary Fund
• Lincoln International Monetary Fund
• Christopher International Monetary Fund
• Gary Bowden
• David Goldblatt
• Am Acad Arts Sci Bull
• Robert D. Putnam
• Fabrizio Zilibotti
• Philippe Aghion
• Peter Howitt

• Recruit researchers
• Join for free
• Login Email Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google Welcome back! Please log in. Email · Hint Tip: Most researchers use their institutional email address as their ResearchGate login Password Forgot password? Keep me logged in Log in or Continue with Google No account? Sign up

• Federal Reserve Facebook Page
• Federal Reserve Instagram Page
• Federal Reserve YouTube Page
• Federal Reserve Flickr Page
• Federal Reserve LinkedIn Page
• Federal Reserve Threads Page
• Federal Reserve Twitter Page
• Subscribe to RSS
• Subscribe to Email
• Recent Postings
• Publications

Board of Governors of the Federal Reserve System

The Federal Reserve, the central bank of the United States, provides the nation with a safe, flexible, and stable monetary and financial system.

Economic Research

Finance and economics discussion series.

Tobias J. Moskowitz, Chase P. Ross , Sharon Y. Ross , Kaushik Vasudevan

Michael B. Gordy and Alexander J. McNeil

Christopher Gust and David Lopez-Salido

International Finance Discussion Papers

Baris Kaymak and Immo Schott

Christoph E. Boehm and T. Niklas Kroner

Martin Bodenstein , Pablo Cuba-Borda , Nils Goernemann , Ignacio Presno

Trang Hoang and Gordon Lewis

François de Soyres , Grace Lofstrom, Mitch Lott, Chris Machol, Zina Saijid

Data, Models and Tools

FRB/US Model :

A large-scale estimated general equilibrium model of the U.S. economy for FRB/US Model  

Yield Curve Models and Data :

Models of daily yield curves, and a dynamic term structure model of Treasury yields

Survey of Consumer Finances (SCF) :

Information on families' balance sheets, pensions, income, and demographic characteristics

Economic Research Data:

View selected research data from the Federal Reserve Board's working papers and notes series.

Estimated Dynamic Optimization (EDO) Model :

A medium-scale New Keynesian dynamic stochastic general equilibrium (DSGE) model of the U.S. economy.

Research Resources

• Staff Publications and Working Papers 2022-2023 (PDF)
• Research Support
• Seminars and Workshops
• Visiting Scholars
• Alumni List
• Other Research
• Federal Reserve Board Research Data Center

Meet the Researchers

By Division

Consumer and Community Affairs Financial Stability International Finance Monetary Affairs Office of Board Members Research and Statistics Reserve Bank Operations and Payment Systems Supervision and Regulation

By Field of Interest

Finance International Economics Macroeconomics Mathematical and Quantitative Methods Microeconomics  

All Researchers

• Overview of economist positions
• Frequently asked questions
• Overview of research assistant positions

Disclaimer: The economic research that is linked from this page represents the views of the authors and does not indicate concurrence either by other members of the Board's staff or by the Board of Governors. The economic research and their conclusions are often preliminary and are circulated to stimulate discussion and critical comment.

The Board values having a staff that conducts research on a wide range of economic topics and that explores a diverse array of perspectives on those topics. The resulting conversations in academia, the economic policy community, and the broader public are important to sharpening our collective thinking.

Numbers, Facts and Trends Shaping Your World

Read our research on:

Full Topic List

Regions & Countries

Publications

• Our Methods
• Short Reads
• Tools & Resources

Read Our Research On:

Economic Systems

Biden, trump supporters both say the u.s. economic system unfairly favors powerful interests.

Trump and Biden supporters share a fair amount of common ground when it comes to criticisms of the U.S. economic system.

Economic ratings are poor – and getting worse – in most countries surveyed

Majorities of adults in 18 of 24 countries surveyed this spring rate their nation’s economic situation poorly.

Most U.S. bank failures have come in a few big waves

After two of the largest U.S. banks collapsed in March, some have started to wonder if a new widespread banking crisis is coming.

Black Americans view capitalism more negatively than positively but express hope in Black businesses

In an August 2022 survey, 54% of Black adults said they had a very or somewhat negative impression of capitalism, up from 40% in May 2019.

Trust in America: How do Americans view economic inequality?

Amid rising inequality, many Americans feel that the U.S. economic system is unfair and generally favors powerful special interests.

Citizens in Advanced Economies Want Significant Changes to Their Political Systems

Dissatisfaction with the functioning of democracy is linked to concerns about the economy, the pandemic and social divisions.

East Germany has narrowed economic gap with West Germany since fall of communism, but still lags

Despite improvements in recent decades, the former East Germany trails the former West on several important economic measures.

10 key takeaways about public opinion in Europe 30 years after the fall of communism

Read key takeaways from a new survey that explores European attitudes three decades after the fall of the Berlin Wall in 1989.

European Public Opinion Three Decades After the Fall of Communism

Thirty years ago, a wave of optimism swept across Europe as walls and regimes fell, and long-oppressed publics embraced open societies, open markets and a more united Europe. Three decades later, a new Pew Research Center survey finds that few people in the former Eastern Bloc regret the monumental changes of 1989-1991.

In Their Own Words: Behind Americans’ Views of ‘Socialism’ and ‘Capitalism’

For many, “socialism” is a word that evokes a weakened work ethic, stifled innovation and excessive reliance on the government. For others, it represents a fairer, more generous society.

REFINE YOUR SELECTION

• Pew Research Center Staff (4)
• Drew DeSilver (2)
• Jacob Poushter (2)
• Janell Fetterolf (2)
• Richard Wike (2)
• Alexandra Castillo (1)
• Andrew Kohut (1942-2015) (1)
• Christine Huang (1)
• J.J. Moncus (1)
• John Gramlich (1)
• Jordan Lippert (1)
• Kat Devlin (1)
• Katie Simmons (1)
• Khadijah Edwards (1)
• Kim Arias (1)
• Laura Clancy (1)
• Laura Silver (1)
• Rakesh Kochhar (1)
• Ruth Igielnik (1)
• Shannon Greenwood (1)
• Shannon Schumacher (1)
• Ted Van Green (1)

Research Teams

• Politics (9)
• Race and Ethnicity (3)
• Social Trends (3)
• Data Labs (2)
• Global Migration and Demography (2)
• Internet and Technology (2)
• Journalism (2)
• Methods (2)
• Pew Research Center (2)
• Religion (2)
• Science (2)

901 E St. NW, Suite 300 Washington, DC 20004 USA (+1) 202-419-4300 | Main (+1) 202-857-8562 | Fax (+1) 202-419-4372 |  Media Inquiries

Research Topics

• Email Newsletters

ABOUT PEW RESEARCH CENTER  Pew Research Center is a nonpartisan fact tank that informs the public about the issues, attitudes and trends shaping the world. It conducts public opinion polling, demographic research, media content analysis and other empirical social science research. Pew Research Center does not take policy positions. It is a subsidiary of  The Pew Charitable Trusts .

© 2024 Pew Research Center

Examining the complex causal relationships between the digital economy and urban tourist destination competitiveness

• Published: 02 August 2024
• Volume 57 , article number  152 , ( 2024 )

Cite this article

• Yan Zhang 1 &
• Jiekuan Zhang 1  

Explore all metrics

Tourism competitiveness has always been a crucial aspect of tourism research. With the emergence of the digital economy, it is important to understand how this new form of economic activity impacts tourism competitiveness. This paper utilizes the configurational theory of systems thinking to examine the complex causal impact of the digital economy on tourism competitiveness. The paper finds that none of the digital economy variables are necessary for tourism competitiveness. There are two basic paths for the digital economy to drive high tourist destination competitiveness: dual-driven model of digital infrastructure and finance occurring in less developed regions and digital regulation-led dual-driven model of digital innovation and finance occurring in developed regions. The reliability of the findings is confirmed by rigorous robustness tests. This research furnishes critical insights into the digital economy and tourism competitiveness theories and applications.

This is a preview of subscription content, log in via an institution to check access.

Access this article

Subscribe and save.

• Get 10 units per month
• Download Article/Chapter or eBook
• 1 Unit = 1 Article or 1 Chapter
• Cancel anytime

Price includes VAT (Russian Federation)

Instant access to the full article PDF.

Rent this article via DeepDyve

Institutional subscriptions

Afolabi JA (2023) Advancing digital economy in Africa: The role of critical enablers. Technol Soc 75:102367

Article   Google Scholar  

Aldrich H, Ruef M (2006) Organizations Evolving. Sage Publications, London

Book   Google Scholar  

Balcerzak P, Bernard MP (2017) Digital economy in Visegrad countries. Multiple-criteria decision analysis at regional level in the years 2012 and 2015. J Compet 9(2):5–18

Google Scholar  

Baños-Pino JF, Boto-García D, Valle ED, Zapico E (2023) Is visitors’ expenditure at destination influenced by weather conditions? Curr Issue Tour 26(10):1554–1572

Bigne E, Ruiz C, Curras-Perez R (2019) Destination appeal through digitalized comments. J Bus Res 101:447–453

Braumoeller BF (2017) Aggregation bias and the analysis of necessary and sufficient conditions in fsQCA. Sociol Methods Res 46(2):242–251

Bulchand-Gidumal J, William Secin E, O’Connor P, Buhalis D (2023) Artificial intelligence’s impact on hospitality and tourism marketing: exploring key themes and addressing challenges. Curr Issue Tour. https://doi.org/10.1080/13683500.2023.2229480

Cardona M, Kretschmer T, Strobel T (2013) ICT and productivity: conclusions from the empirical literature. Inf Econ Policy 25(3):109–125

Chen Z, Kahn ME, Liu Y, Wang Z (2018) The consequences of spatially differentiated water pollution regulation in China. J Environ Econ Manag 88:468–485

Chen Q, Qi Y, Zhang G (2023a) Digital economy, government intellectual property protection, and entrepreneurial activity in China. Technol Anal Strateg Manag. https://doi.org/10.1080/09537325.2023.2264950

Chen S, Chan ICC, Xu S, Law R, Zhang M (2023b) Metaverse in tourism: drivers and hindrances from stakeholders’ perspective. J Travel Tour Mark 40(2):169–184

Cheng X, Xue T, Yang B, Ma B (2023) A digital transformation approach in hospitality and tourism research. Int J Contemp Hosp Manag. https://doi.org/10.1108/IJCHM-06-2022-0679

Crilly D, Zollo M, Hansen MT (2012) Faking it or muddling through? understanding decoupling in response to stakeholder pressures. Acad Manag J 55(6):1429–1448

Crouch GI (2011) Destination competitiveness: An analysis of determinant attributes. J Travel Res 50(1):27–45

De Crescenzo V, Ribeiro-Soriano DE, Covin JG (2020) Exploring the viability of equity crowdfunding as a fundraising instrument: a configurational analysis of contingency factors that lead to crowdfunding success and failure. J Bus Res 115:348–356

Ding H (2022) What kinds of countries have better innovation performance?–a country-level fsQCA and NCA study. J Innov Knowl 7(4):100215

Ditta-Apichai M, Gretzel U, Kattiyapornpong U (2023) Platform empowerment: Facebook’s role in facilitating female micro-entrepreneurship in tourism. J Sustain Tour. https://doi.org/10.1080/09669582.2023.2215479

Dolnicar S (2019) A review of research into paid online peer-to-peer accommodation: launching the annals of tourism research curated collection on peer-to-peer accommodation. Ann Tour Res 75:248–264

Du Y, Liu Q, Cheng J (2020) What kind of ecosystem for doing business will contribute to city-level high entrepreneurial activity? A research based on institutional configurations. J Manag World 36(9):141–154 ( In Chinese )

Dul J (2016) Necessary condition analysis (NCA) logic and methodology of “necessary but not sufficient” causality. Organ Res Methods 19(1):10–52

Dul J, Van der Laan E, Kuik R (2020) A statistical significance test for necessary condition analysis. Organ Res Methods 23(2):385–395

Dwyer L, Kim C (2003) Destination competitiveness: determinants and indicators. Curr Issue Tour 6(5):369–414

Filimonau V, Naumova E (2020) The blockchain technology and the scope of its application in hospitality operations. Int J Hosp Manag 87:102383

Fiss PC (2011) Building better causal theories: A fuzzy set approach to typologies in organization research. Acad Manag J 54(2):393–420

Frenzel F, Giddy J, Frisch T (2022) Digital technology, tourism and geographies of inequality. Tour Geogr 24(6–7):923–933

Furnari S, Crilly D, Misangyi VF, Greckhamer T, Fiss PC, Aguilera RV (2021) Capturing causal complexity: Heuristics for configurational theorizing. Acad Manag Rev 46(4):778–799

Geng Y, Zheng Z, Ma Y (2023) Digitization, perception of policy uncertainty, and corporate green innovation: A study from China. Econ Anal Policy 80:544–557

Gomez-Vega M, Herrero-Prieto LC, López MV (2022) Clustering and country destination performance at a global scale: determining factors of tourism competitiveness. Tour Econ 28(6):1605–1625

Gon M (2021) Local experiences on Instagram: Social media data as source of evidence for experience design. J Destin Mark Manag 19:100435

González-Rodríguez MR, Díaz-Fernández MC, Pulido-Pavón N (2023) Tourist destination competitiveness: an international approach through the travel and tourism competitiveness index. Tour Manag Perspect 47:101127

Gössling S (2021) Tourism, technology and ICT: a critical review of affordances and concessions. J Sustain Tour 29(5):733–750

Gössling S, Larson M, Pumputis A (2021) Mutual surveillance on Airbnb. Ann Tour Res 91:103314

Guedes A, Faria S, Gouveia S, Rebelo J (2023) The effect of virtual proximity and digital adoption on international tourism flows to Southern Europe. Tour Econ 29(6):1643–1661

Guizzardi A, Mariani MM (2021) Introducing the dynamic destination satisfaction method: an analytical tool to track tourism destination satisfaction trends with repeated cross-sectional data. J Travel Res 60(5):965–980

Guo F, Wang J, Wang F, Kong X, Cheng Z (2020) Measuring China’s digital financial inclusion: index compilation and spatial characteristics. China Econ Quart 19(4):1401–1418 ( In Chinese )

Guo X, Wang Y, Tao J, Guan H (2023) Identifying unique attributes of tourist attractions: an analysis of online reviews. Curr Issue Tour. https://doi.org/10.1080/13683500.2023.2165904

Huang B, Li H, Liu J, Lei J (2023a) Digital technology innovation and the high-quality development of Chinese enterprises: Evidence from enterprise’s digital patents. Econ Res J 58(03):97–115 ( in Chinese )

Huang X, Zhang S, Zhang J, Yang K (2023b) Research on the impact of digital economy on regional green technology innovation: moderating effect of digital talent aggregation. Environ Sci Pollut Res 30(29):74409–74425

Ivars-Baidal JA, Vera-Rebollo JF, Perles-Ribes J, Femenia-Serra F, Celdrán-Bernabeu MA (2023) Sustainable tourism indicators: what’s new within the smart city/destination approach? J Sustain Tour 31(7):1556–1582

Kumar S, Kumar D, Nicolau JL (2024) How does culture influence a Country’s travel and tourism competitiveness? a longitudinal frontier study on 39 countries. Tour Manage 100:104822

Lee JS, Park S (2023) A cross-cultural anatomy of destination image: an application of mixed-methods of UGC and survey. Tour Manage 98:104746

Li F (2020) The digital transformation of business models in the creative industries: a holistic framework and emerging trends. Technovation 92:102012

Li M, Meng M, Chen Y (2024) The impact of the digital economy on green innovation: the moderating role of fiscal decentralization. Econ Chang Restruct 57:37

Liu H, Cui W, Zhang M (2022) Exploring the causal relationship between urbanization and air pollution: evidence from China. Sustain Cities Soc 80:103783

Liu J, Chen Y, Liang FH (2023) The effects of digital economy on breakthrough innovations: evidence from Chinese listed companies. Technol Forecast Soc Chang 196:122866

Ma L, Ouyang M (2023) Spatiotemporal heterogeneity of the impact of digital inclusive finance on tourism economic development: evidence from China. J Hosp Tour Manag 56:519–531

Ma S, Wei W, Li J (2023) Has the digital economy improved the ecological environment? empirical evidence from China. Environ Sci Pollut Res 30(40):91887–91901

Martínez-Caro E, Cegarra-Navarro JG, Alfonso-Ruiz FJ (2020) Digital technologies and firm performance: the role of digital organisational culture. Technol Forecast Soc Chang 154:119962

Mertzanis C, Papastathopoulos A (2021) Epidemiological susceptibility risk and tourist flows around the world. Ann Tour Res 86:103095

Nambisan S, Wright M, Feldman M (2019) The digital transformation of innovation and entrepreneurship: progress, challenges and key themes. Res Policy 48(8):103773

Nathan RJ, Victor V, Tan M, Fekete-Farkas M (2020) Tourists’ use of Airbnb app for visiting a historical city. Inf Technol Tour 22(2):217–242

Navío-Marco J, Ruiz-Gómez LM, Sevilla-Sevilla C (2018) Progress in information technology and tourism management: 30 years on and 20 years after the internet-Revisiting Buhalis & Law’s landmark study about eTourism. Tour Manage 69:460–470

Ndubisi NO, Nair S (2023) International tourism: Inimitable VS imitable core tourism resources and destination image. J Destin Mark Manag 27:100756

Neumann P, Mason CW (2023) The influence of transportation and digital technologies on backcountry tourism and recreation in British Columbia. Can Tour Geogr 25(4):1166–1185

Niu G, Jin X, Wang Q, Zhou Y (2022) Broadband infrastructure and digital financial inclusion in rural China. China Econ Rev 76:101853

Ozili PK (2018) Impact of digital finance on financial inclusion and stability. Borsa Istanbul Rev 18(4):329–340

Pan W, Xie T, Wang Z, Ma L (2022) Digital economy: An innovation driver for total factor productivity. J Bus Res 139:303–311

Pandey N, Pal A (2020) Impact of digital surge during Covid-19 pandemic: a viewpoint on research and practice. Int J Inf Manage 55:102171

Pencarelli T (2020) The digital revolution in the travel and tourism industry. Inf Technol Tour 22(3):455–476

Ragin CC (2009) Redesigning social inquiry: Fuzzy sets and beyond. University of Chicago Press, Chicago

Song H, Jiao E, Park S (2023) Sectoral productivity and destination competitiveness. Ann Tour Res 103:103645

Sturgeon TJ (2021) Upgrading strategies for the digital economy. Glob Strateg J 11(1):34–57

Tang R (2023) Can digital economy improve tourism economic resilience? evidence from China. Tour Econ. https://doi.org/10.1177/13548166231206241

Tang R (2023b) Digital economy and total factor productivity of tourism enterprises in China: the perspective of market competition theory. Curr Issue Tour. https://doi.org/10.1080/13683500.2022.2159338

Tham A, Sigala M (2020) Road block (chain): bit (coin)s for tourism sustainable development goals? J Hosp Tour Technol 11(2):203–222

Vis B, Dul J (2018) Analyzing relationships of necessity not just in kind but also in degree: complementing fsQCA with NCA. Sociol Methods Res 47(4):872–899

Vu K, Hartley K (2022) Drivers of growth and catch-up in the tourism sector of industrialized economies. J Travel Res 61(5):1156–1172

Wang J, Dong K, Dong X, Taghizadeh-Hesary F (2022a) Assessing the digital economy and its carbon-mitigation effects: The case of China. Energy Econ 113:106198

Wang J, Dong X, Dong K (2022b) How digital industries affect China’s carbon emissions? Analysis of the direct and indirect structural effects. Technol Soc 68:101911

Wilson J, Garay-Tamajon L, Morales-Perez S (2022) Politicising platform-mediated tourism rentals in the digital sphere: Airbnb in Madrid and Barcelona. J Sustain Tour 30(5):1080–1101

Woods O, Shee SY (2021) “ Doing it for the’gram”? the representational politics of popular humanitarianism. Ann Tour Res 87:103107

World Economic Forum (2022) Travel & Tourism Development Index 2021: Rebuilding for a Sustainable and Resilient Future . Derived from https://www3.weforum.org/docs/WEF_Travel_Tourism_Development_2021.pdf . Accessed on September 4, 2023.

Wu F, Hu H, Lin H, Ren X (2021) Enterprise digital transformation and capital market performance: Empirical evidence from stock liquidity. J Manag World 37(07):130–144 ( in Chinese )

Wu D, Li H, Wang Y (2023) Measuring sustainability and competitiveness of tourism destinations with data envelopment analysis. J Sustain Tour 31(6):1315–1335

Xu J, Au T (2023) Destination competitiveness since 2010: Research themes, approaches, and agenda. Tour Rev 78(3):665–696

Zhang J (2023a) A multidimensional perspective on the relationship between tourism and green growth. J Sustain Tour. https://doi.org/10.1080/09669582.2023.2276039

Zhang Z (2023b) The impact of the artificial intelligence industry on the number and structure of employments in the digital economy environment. Technol Forecast Soc Chang 197:122881

Zhang JK (2024) Does innovative city construction promote tourist destination competitiveness? an analysis of a quasi-natural experiment. J Destin Mark Manag 33:100916

Zhang J, Zhang Y (2021) A qualitative comparative analysis of tourism and gender equality in emerging economies. J Hosp Tour Manag 46:284–292

Zhang J, Lyu Y, Li Y, Geng Y (2022a) Digital economy: an innovation driving factor for low-carbon development. Environ Impact Assess Rev 96:106821

Zhang W, Liu X, Wang D, Zhou J (2022b) Digital economy and carbon emission performance: evidence at China’s city level. Energy Policy 165:112927

Zhou H, Wang S, Zhao T (2023) Digital technology patents in China: an integrated analysis of patent distribution and transactions. Technol Anal Strateg Manag. https://doi.org/10.1080/09537325.2023.2250016

Download references

Author information

Authors and affiliations.

School of Tourism Management, Guilin Tourism University, No.26, Liangfeng Road, Yanshan District, Guilin, Guangxi, China

Yan Zhang & Jiekuan Zhang

You can also search for this author in PubMed   Google Scholar

Corresponding author

Correspondence to Jiekuan Zhang .

Ethics declarations

Conflict of interest, additional information, publisher's note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

See Table  6 and Fig.  5 .

Entropy-weighted TOPSIS technique

Rights and permissions

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted manuscript version of this article is solely governed by the terms of such publishing agreement and applicable law.

Reprints and permissions

About this article

Zhang, Y., Zhang, J. Examining the complex causal relationships between the digital economy and urban tourist destination competitiveness. Econ Change Restruct 57 , 152 (2024). https://doi.org/10.1007/s10644-024-09739-1

Download citation

Received : 17 February 2024

Accepted : 16 July 2024

Published : 02 August 2024

DOI : https://doi.org/10.1007/s10644-024-09739-1

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Digital economy
• Tourist destination competitiveness
• Configurational theory
• Fuzzy-set qualitative comparative analysis
• Necessary condition analysis
• Find a journal
• Publish with us
• Track your research

Integrating an abandoned farmland simulation model (AFSM) using system dynamics and CLUE-S for sustainable agriculture

• Yang, Dazhi
• Wang, Yanwei

Addressing the challenges of achieving United Nations Sustainable Development Goal SDG2 is complex, especially considering the profound impacts of population growth, environmental degradation, and disruptions from the COVID-19 pandemic and armed conflicts on the global food supply system. Unjustifiable farmland abandonment poses a critical obstacle to food security, demanding a comprehensive knowledge framework for sensible decision-making in abandoned farmland reclamation. However, current research lacks systematic solutions. This research seeks to address the challenges of farmland abandonment through a comprehensive knowledge framework. Integrating remote sensing surveillance, mechanism analysis, and scenario projection, the primary objective is to develop the abandoned farmland simulation model (AFSM) by combining abandoned farmland identification, the system dynamics (SD), and CLUE-S. The AFSM facilitates explicit spatiotemporal abandonment simulation, contributing to a deeper understanding of the dynamic evolution and mechanisms involved in farmland abandonment. The AFSM is formulated through a logical sequence of "spatiotemporal data acquisition – quantity prognostication – spatial simulation." The process begins by tracking annual land use changes, identifying spatio-temporal changes in abandoned farmland. Subsequently, the research employs the SD model to establish an elucidative framework for abandonment mechanisms, facilitating the quantitative analysis of factors influencing farmers' decisions regarding farmland abandonment. Finally, the CLUE-S model is utilized to prognosticate the spatial abandonment trend. All modules of the model have passed the precision test and inspection. The AFSM outcomes reveal that the abandonment rate in the city of Jingdezhen fluctuated between 5.03% and 13.21% from 2003 to 2020, averaging 8.37%. The model systematically quantifies the multiple socio-economic variables that impact abandonment decisions. Key parameter groups delineate distinct development scenarios and make predictions. By 2032, the abandonment rate under the scenario of farmland protection is projected to be 10.88%, markedly lower than the inertial development and economic priority scenarios by 15.24% and 13.43%, respectively. The concentration of abandoned farmland is primarily observed in elevated, steep-sloped areas, exhibiting a "climbing mountain and slope" trend. The AFSM emerges as a crucial tool for policymakers, farmland managers, and researchers, empowering them to make well-informed decisions concerning abandoned farmland reclamation. The explicit spatio-temporal simulation, complemented by a clarifying framework for abandonment mechanisms, enables the quantification of the influence of various factors on farmland abandonment decision-making and realizes the coupling of quantitative and spatial simulation under various scenarios.

• Abandoned farmland;
• Scenario simulation;
• Food security;
• Spatio-temporal information;
• Mechanism explanation

INDIA-MIDDLE EAST-EUROPE CORRIDOR (IMEC): RHETORIC, REALITIES AND IMPLICATIONS FOR PAKISTAN

• Raza Ali Khan PhD Scholar at the Department of International Relations, National University of Modern Languages (NUML), Islamabad.

This research article explores the India-Middle East-Europe Corridor (IMEC) to ascertain whether it is a myth or a reality. It also examines its transparency, timeliness, and primary objectives. Against the backdrop of instability in the Middle East, mainly arising from the ongoing Gaza situation, the study assesses the regional and global implications of the IMEC, considering arguments both in favour and against the initiative. Nevertheless, the problem has a similar impact on all such ventures as China’s Belt and Road Initiative (BRI) since the successful realisation is contingent upon the stable and enabling regional environment. This exploration suffers from the challenge of access to authentic sources. Given the contemporary nature of the subject matter, this study will primarily rely on secondary literature. Employing an Intergovernmentalism theoretical approach, this research delves into the dynamics of IMEC, discussing its intended objectives and challenges. The central argument contends that a meticulous analysis of the corridor's cause and effect on Pakistan's economic and geostrategic interests is imperative. The IMEC initiative exhibits the potential of a relatively cheap system, enhancing contemporary transportation infrastructure. However, the study asserts that the IMEC presents an obvious threat to the China-Pakistan Economic Corridor (CPEC) and the BRI, potentially compromising Pakistan's economic and strategic interests. The paper adopts a nuanced perspective, emphasising that IMEC and BRI contrast in strategic planning and development approaches. Should the IMEC materialise according to its envisaged plan, it will yield substantial diplomatic, strategic, and economic benefits for India. The article concludes by underscoring the challenges faced by IMEC, including coordination complexities, financial constraints for some participating nations, and potential geopolitical tensions affecting corridor operations .

Bibliography Entry

Khan, Raza Ali. 2024. "India-Middle East-Europe Corridor (IMEC): Rhetoric, Realities and Implications for Pakistan."  Margalla Papers  28 (1): 75-92.

Author Biography

Raza ali khan, phd scholar at the department of international relations, national university of modern languages (numl), islamabad..

Raza Ali Khan is an author and PhD Scholar at the Department of International Relations, National University of Modern Languages (NUML), Islamabad.

How to Cite

• Endnote/Zotero/Mendeley (RIS)

Copyright (c) 2024 Raza Ali Khan

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International License .

License Terms

Make a Submission

e-ISSN: 2789-7028

ISSN-L: 1999-2297

HEC Recognition

------------------------------

Announcements

Call for papers - 2024 (issue-ii), scan qr code, information.

• For Readers
• For Authors
• For Librarians

Current Issue

Social Media

Important Links

• NDU Journal
• ISSRA Papers
• Journal of Contemporary Studies
• Strategic Thought

Quick Links

• Publication Policy
• Publication Ethics
• Guidelines for Authors
• Article Processing Chart
• Subscription

Institute for Strategic Studies, Research and Analysis, National Defence University, Islamabad 

Copyright © 2023 National Defence University, Islamabad, Pakistan

There appears to be a technical issue with your browser

This issue is preventing our website from loading properly. Please review the following troubleshooting tips or contact us at [email protected] .

Analysis: A New Era of Financial Warfare Has Begun

Create an FP account to save articles to read later.

ALREADY AN FP SUBSCRIBER? LOGIN

Downloadable PDFs are a benefit of an FP subscription.

Subscribe Now

World Brief

• Editors’ Picks
• Africa Brief

China Brief

• Latin America Brief

South Asia Brief

Situation report.

• Flash Points
• War in Ukraine
• Israel and Hamas
• U.S.-China competition
• U.S. election 2024
• Biden's foreign policy
• Trade and economics
• Artificial intelligence
• Asia & the Pacific
• Middle East & Africa

What Iran Might Do Next

Ones and tooze, foreign policy live.

Summer 2024 Issue

Print Archive

FP Analytics

• In-depth Special Reports
• Issue Briefs
• Power Maps and Interactive Microsites
• FP Simulations & PeaceGames
• Graphics Database

Catalysts for Change

Fp @ unga79, ai for healthy cities, her power @ unga79, fp tech forum @ unga79.

By submitting your email, you agree to the Privacy Policy and Terms of Use and to receive email correspondence from us. You may opt out at any time.

Your guide to the most important world stories of the day

Essential analysis of the stories shaping geopolitics on the continent

The latest news, analysis, and data from the country each week

Weekly update on what’s driving U.S. national security policy

Evening roundup with our editors’ favorite stories of the day

One-stop digest of politics, economics, and culture

Weekly update on developments in India and its neighbors

A curated selection of our very best long reads

A New Era of Financial Warfare Has Begun

The west’s latest actions against russia carry risks for the global system and could provoke china..

• United States

Russia’s War in Ukraine

Understanding the conflict two years on.

More on this topic

Washington and the West have begun a new phase of financial warfare against Russia and China—a powerful but also potentially risky escalation that, if people aren’t careful, could eventually give Moscow and Beijing exactly the outcome they are believed to be looking for.

How so? Because the unprecedented actions taken at the G-7 summit in June to hand over to Ukraine billions of dollars in profits earned on frozen Russian assets—along with new actions taken against Chinese banks—could begin to undermine the legitimacy of the U.S.-dominated international financial system, some experts say. And that could make Russian President Vladimir Putin and especially Chinese President Xi Jinping, who is said to want to create an alternative renminbi-based financial system, very happy in the end.

At a time when many nations are unsure about whether to do business with Russia and are falling into the debt-enforced embrace of China, the G-7 action sends a message: What was once sacrosanct in international finance may be no longer. A number of sovereign wealth funds, central banks, corporations, and private investors—especially from the smaller countries of the global south that are most vulnerable to sanctions—may well want to hedge against full investment in dollar- and euro-based holdings.

“This decision crosses the Rubicon,” said Ryan Martínez Mitchell, a law professor at the Chinese University of Hong Kong, by “weakening the norm of sovereign immunity for foreign central banks.”

“Any shift away from a U.S. dollar-based global financial system is not a near-term prospect, but decisions like these do probably add to the constituency that would welcome that kind of future,” Mitchell said. Others agree. “There were many forces pushing for a search for alternatives to [the U.S.] dollar, and this move will give an additional push to those efforts,” said Harold James, a financial historian at Princeton University. “I believe we are at a tipping point in which two worries coincide: one about the likely fiscal path of the U.S. and an unsustainably large burden; the second about seizure of assets, with secondary sanctions possibly being applied to countries that are in a supply chain with China and then indirectly with Russia.”

The “tipping point,” James warns, could come in the form of many countries, even U.S. allies, beginning to move their assets away from the dollar and euro. According to Raghuram Rajan of the University of Chicago, a former governor of the Reserve Bank of India, nations are disturbed by the idea that Russia’s $300 billion in central bank reserves have been inaccessible for more than two years. “Some central banks have started diversifying reserves a little more as a result, including into gold,” Rajan said.

James added: “One sign that I find very telling is how Central European countries, the Czech Republic and Poland, both of which feel very close to the U.S. and who weren’t interested in gold reserves when they felt secure—indeed, the Czech Republic sold its gold reserves the day they entered NATO in March 1999—are now buying large amounts of gold.”

Putin himself spoke triumphantly of this trend in his notorious interview with renegade U.S. newscaster Tucker Carlson in February. Washington’s decision “to use the dollar as a tool of the foreign-policy struggle is one of the biggest strategic mistakes made by the U.S. political leadership,” Putin said , pointing to America’s fiscal profligacy. “Even the U.S. allies are now downsizing their dollar reserves.” At another point, Putin warned other countries that they “could be next in line for expropriation by the United States and the West.”

Wary of the risks of sending a destabilizing message, the G-7 did stop short of actually seizing the Russian assets at its summit in Italy. Instead, it adopted a complex scheme to transfer so-called windfall profits on earnings from frozen Russian central bank securities—the earnings of some $3 billion to $4 billion a year come from investments by Euroclear, the financial services company in Belgium that holds the Russian assets—to supply finance to Ukraine.

It was unprecedented all the same. As a senior Biden administration official described it: “Never before in history has a multilateral coalition immobilized the sovereign assets of an aggressor country and then found a way to unlock the value of those assets for the benefit of the aggrieved party as it fights for its freedom. That’s what happened at this G-7.”

However it’s done, making money off other nations’ assets—even aggressor nations, such as Russia, in total violation of global norms—is a risky precedent. “Once a new sanction becomes seen as effective, its usage tends to proliferate,” said Jon Bateman, a senior fellow at the Carnegie Endowment for International Peace. “In recent years, creative new uses of export control powers—such as the Entity List and the Foreign Direct Product Rule—have ping-ponged between Chinese and Russian targets, with each country serving as a proving ground for actions later taken against the other.”

Nor did the G-7 leaders stop there. They also indicated that new measures were being considered that might gradually cut Beijing out of the international financial system. While saying in a communiqué that they “recognize the importance of China in global trade” and affirming that they “are not trying to harm China or thwart its economic development,” the leaders obliquely threatened Chinese banks “and other entities in China” with measures to “restrict access to our financial systems.” That could ratchet up the war—and the risks to the system—dramatically.

China has already been quietly insulating itself from financial retaliation over its support of Russia in the past two years, said Hung Tran, a former deputy director at the International Monetary Fund, in a June 21 interview. “The major Chinese banks have been very cautious even in reducing their exposure and dealings with Russia. In place of that, smaller institutions not having any business with any U.S. entity have been set up to handle trade with Russia so that basically Russia-China trade is settled in renminbi and rubles.”

The senior administration official justified the decision to increase pressure on China by saying that “some of China’s actions to support the Russian war machine are now not just threatening Ukraine’s existence but European security and trans-Atlantic security.” The official added that among other “unrivaled policy distortions coming out of China”—meaning its unfair trade practices—Beijing was now openly supplying dual-use components and other economic aid to Russia. “There was unanimous agreement that the Russian military has been sustained by transforming its entire economy into a war machine and because China and other countries have been willing to serve” that effort, the official said.

In a blunt statement during his visit to Beijing in April, Secretary of State Antony Blinken reiterated these accusations, declaring that China was “powering Russia’s brutal war of aggression against Ukraine” as “the top supplier of machine tools, microelectronics, nitrocellulose, which is critical to making munitions and rocket propellants, and other dual-use items that Moscow is using to ramp up its defense industrial base.”

The actions taken at the G-7 summit may well have been necessary. Nearly two and a half years into the war, support for aid from the United States and Europe is flagging, Kyiv’s forces are exhausted, Russia’s economy is still looking fairly robust, and a new anti-Western alignment is hardening between Moscow, Beijing, Tehran, and most recently North Korea. “We are stepping up our collective efforts to disarm and defund Russia’s military industrial complex,” the G-7 leaders said in their communiqué.

This latest approach to squeezing Russia started slowly, even painfully, amid a great deal of tension between the United States and European governments about just how tough to get with Moscow. Immediately following Putin’s invasion of Ukraine in February 2022, none of those governments had a problem imposing the usual economic sanctions—import and export restrictions and the like—and quickly. They took a major step further when they froze Russia’s central bank assets—an unprecedented move against such a large country—in addition to real estate properties, stocks, bonds, and various investments held by Russian oligarchs.

But actually seizing those bank assets was seen as a step too far, especially by the Europeans, who fought off an effort led by the U.S. Congress, and ultimately backed by the Biden administration, to pursue full seizure. That meant tampering with the international financial system itself—the complex postwar network of norms, codes, and laws that has underwritten the greatest surge of prosperity in recorded history and enriched the West. That felt a little too much like playing with elemental fire because it meant threatening the idea of sovereign immunity that is central to the system and because it meant posing increased risks to the holding of dollar- and euro-denominated assets. And having established this precedent, what about China? What effect will the G-7’s warnings have on Xi?

The shot fired in the communiqué could deter Xi from doing even more to isolate China’s ailing economy than he already has—specifically by invading or blockading Taiwan. Or, alternatively, it could mean the beginning of the end of the postwar global economic system if Xi decides to move against Taiwan anyway. Indeed, he could easily gamble that the United States wouldn’t dare do to China what it’s doing to Russia for exactly that reason.

If the United States and West were to respond to an invasion or blockade of Taiwan by freezing and leveraging Chinese assets, the result could be a freeze-up of the whole financial system and a devastating blow to the global economy. In the case of Russia, Washington needed to undergo many months of negotiation with the European Union because the vast majority of Russian assets are held in Europe and there was only about $300 billion or so to freeze. The same is not true of Chinese assets, which are huge and spread all over the world. Under the International Emergency Economic Powers Act, Washington would be able to freeze some $800 billion in Chinese Treasury bill holdings entirely on its own, which is only a portion of some $3 trillion in Chinese-owned sovereign assets overseas. But Beijing could easily retaliate against that nearly $6 trillion in Western investment in China.

As Tran argues, the threat of a kind of financial MAD, or mutual assured destruction, is far too great. In “terms of balance sheet exposures, China has about $3.4 trillion of identifiable international assets at risk of possible sanctions and up to $5.8 trillion of liabilities to, or assets in China of, international investors and companies largely from Western countries. China therefore has plenty of room to take retaliatory actions,” Tran wrote in a 2022 post for the Atlantic Council titled “Wargaming a Western Freeze of China’s Foreign Reserves.”

The deep cross-integration between China and the West is what has led both sides to avoid a complete decoupling of economies, reflecting what former U.S. Treasury Secretary Larry Summers once called a “financial balance of terror.” As a result, “there will be more resistance to imposing the scope of sanctions we have imposed on Russia because Western economies are far more intertwined with China’s than they were with Russia’s,” said William Reinsch, a former U.S. commerce undersecretary now at the Center for Strategic and International Studies.

Reinsch notes there is an important “qualitative difference” as well: “The Russian assets being used are those seized from oligarchs who have supported/enabled Putin. There are some Chinese oligarchs, but their relationship with their own government is much different, as is their role in the economy. If you go beyond oligarchs, you get very quickly to seizing sovereign assets, which I doubt the West would do and for which the consequences would be significant.”

But according to some China experts, the latest moves might only spur Xi to further decouple his economy. The “dimmer” that peaceful reunification with Taiwan seems, “the more incentives Beijing would have to reduce vulnerabilities to sanctions in case of a militarized conflict,” said Zongyuan Zoe Liu, a fellow at the Council on Foreign Relations and columnist for Foreign Policy . “China has been diversifying its foreign exchange reserves since the 2000s. While previously the primary motivation was to search for higher returns and strategic assets, now it is also to reduce vulnerabilities to sanctions.”

And while Xi’s dream of a renminbi-based system still “has a long way to go”—the yuan is a distant fifth in global reserve currency holdings—escalating Western moves “may ultimately weaken international law protections for everyone, not only their intended targets,” Mitchell wrote recently for the Quincy Institute. As a result, “intensified weaponization of Western currencies could indeed boost China’s yuan efforts, and, more significantly, provide a major stimulus to plans for a BRICS basket reserve currency. The move would simultaneously improve Beijing’s reputation as an apparently more responsible actor with respect to foreign assets, while also perversely incentivizing it to further experiment with its own nascent unilateral sanctions regime.”

Russia is much more willing than China to blow up the international system. But that doesn’t mean Xi won’t decide he can afford to see that happen as well. As Tran argues, Beijing has been pursuing a “dual-track” strategy of working within the current Western-led trading system “but also wanting to find alternative ways to do this trade without being exposed to dollar sanctions.” Further sanctions could only push Xi further in the radical direction of trying to set up an alternative renminbi-based financial trading system.

“Both sides are kind of upping their ante,” Tran said.

Michael Hirsh is a columnist for Foreign Policy. He is the author of two books:  Capital Offense: How Washington’s Wise Men Turned America’s Future Over to Wall Street  and  At War With Ourselves: Why America Is Squandering Its Chance to Build a Better World . Twitter:  @michaelphirsh

Join the Conversation

Commenting on this and other recent articles is just one benefit of a Foreign Policy subscription.

Already a subscriber? Log In .

Subscribe Subscribe

View Comments

Join the conversation on this and other recent Foreign Policy articles when you subscribe now.

Not your account? Log out

Please follow our comment guidelines , stay on topic, and be civil, courteous, and respectful of others’ beliefs.

Change your username:

I agree to abide by FP’s comment guidelines . (Required)

Confirm your username to get started.

The default username below has been generated using the first name and last initial on your FP subscriber account. Usernames may be updated at any time and must not contain inappropriate or offensive language.

Is This a Revolution? Or Are People Just Very Ticked Off?

In a new book, Fareed Zakaria explores how much the times are a-changin’. At risk, he says, is the entire global system.

The G-7 Must Prepare Now for Trump

This summer’s summit needs to be much more than just a 50th anniversary celebration.

Sign up for Editors' Picks

A curated selection of fp’s must-read stories..

You’re on the list! More ways to stay updated on global news:

The Stubborn Legend of a Western ‘Coup’ in Ukraine

The 1960s novella that got ai (mostly) right, inside ukraine’s fight for hearts and minds, how olympic officials first drew gender lines, iran vows ‘revenge’ on israel as haniyeh buried in qatar, editors’ picks.

• 1 The Dragon Warrior
• 2 What the Olympics Mean to China

The Stubborn Legend of a Western 'Coup' in Ukraine

Review: buzzati's 'the singularity': the sci-fi novella that got ai (mostly) right, ukraine's campaign to win over the world, olympic gender disputes go back to 1930s rules, hamas's haniyeh buried in qatar, middle east prepares for military escalation, more from foreign policy, the trump-vance unilateralist delusion.

The Republican ticket has unified—around a fundamentally unrealistic foreign policy.

Harris Candidacy Gives Democrats a Chance to Pivot on Gaza

What she can do to right U.S. foreign policy and bring back voters.

China and the U.S. Are Careening Toward a South China Sea Crisis

A failure of deterrence risks pushing the region into deeper conflict.

The Kamala Harris Doctrine

Everything we know about the presumptive Democratic nominee’s foreign-policy views.

Why Americans and Israelis Don’t See Eye to Eye on Iran

The dragon warrior, what the olympics mean to china, how bad will political violence in the u.s. get, don’t call what’s happening in venezuela a civil war.

Sign up for World Brief

FP’s flagship evening newsletter guiding you through the most important world stories of the day, written by Alexandra Sharp . Delivered weekdays.

• Academic Medicine
• Acid Base, Electrolytes, Fluids
• Allergy and Clinical Immunology
• American Indian or Alaska Natives
• Anesthesiology
• Anticoagulation
• Art and Images in Psychiatry
• Artificial Intelligence
• Assisted Reproduction
• Bleeding and Transfusion
• Caring for the Critically Ill Patient
• Challenges in Clinical Electrocardiography
• Climate and Health
• Climate Change
• Clinical Challenge
• Clinical Decision Support
• Clinical Implications of Basic Neuroscience
• Clinical Pharmacy and Pharmacology
• Complementary and Alternative Medicine
• Consensus Statements
• Coronavirus (COVID-19)
• Critical Care Medicine
• Cultural Competency
• Dental Medicine
• Dermatology
• Diabetes and Endocrinology
• Diagnostic Test Interpretation
• Drug Development
• Electronic Health Records
• Emergency Medicine
• End of Life, Hospice, Palliative Care
• Environmental Health
• Equity, Diversity, and Inclusion
• Facial Plastic Surgery
• Gastroenterology and Hepatology
• Genetics and Genomics
• Genomics and Precision Health
• Global Health
• Guide to Statistics and Methods
• Hair Disorders
• Health Care Delivery Models
• Health Care Economics, Insurance, Payment
• Health Care Quality
• Health Care Reform
• Health Care Safety
• Health Care Workforce
• Health Disparities
• Health Inequities
• Health Policy
• Health Systems Science
• History of Medicine
• Hypertension
• Images in Neurology
• Implementation Science
• Infectious Diseases
• Innovations in Health Care Delivery
• JAMA Infographic
• Law and Medicine
• Leading Change
• Less is More
• LGBTQIA Medicine
• Lifestyle Behaviors
• Medical Coding
• Medical Devices and Equipment
• Medical Education
• Medical Education and Training
• Medical Journals and Publishing
• Mobile Health and Telemedicine
• Narrative Medicine
• Neuroscience and Psychiatry
• Notable Notes
• Nutrition, Obesity, Exercise
• Obstetrics and Gynecology
• Occupational Health
• Ophthalmology
• Orthopedics
• Otolaryngology
• Pain Medicine
• Palliative Care
• Pathology and Laboratory Medicine
• Patient Care
• Patient Information
• Performance Improvement
• Performance Measures
• Perioperative Care and Consultation
• Pharmacoeconomics
• Pharmacoepidemiology
• Pharmacogenetics
• Pharmacy and Clinical Pharmacology
• Physical Medicine and Rehabilitation
• Physical Therapy
• Physician Leadership
• Population Health
• Primary Care
• Professional Well-being
• Professionalism
• Psychiatry and Behavioral Health
• Public Health
• Pulmonary Medicine
• Regulatory Agencies
• Reproductive Health
• Research, Methods, Statistics
• Resuscitation
• Rheumatology
• Risk Management
• Scientific Discovery and the Future of Medicine
• Shared Decision Making and Communication
• Sleep Medicine
• Sports Medicine
• Stem Cell Transplantation
• Substance Use and Addiction Medicine
• Surgical Innovation
• Surgical Pearls
• Teachable Moment
• Technology and Finance
• The Art of JAMA
• The Arts and Medicine
• The Rational Clinical Examination
• Tobacco and e-Cigarettes
• Translational Medicine
• Trauma and Injury
• Treatment Adherence
• Ultrasonography
• Users' Guide to the Medical Literature
• Vaccination
• Venous Thromboembolism
• Veterans Health
• Women's Health
• Workflow and Process
• Wound Care, Infection, Healing
• Download PDF
• Share X Facebook Email LinkedIn
• Permissions

The Costs of Long COVID

• 1 Department of Economics and Kennedy School of Government, Harvard University, Cambridge, Massachusetts
• Research Letter BNT162b2 Vaccination and Long COVID After Infections Not Requiring Hospitalization in Health Care Workers Elena Azzolini, MD, PhD; Riccardo Levi, MSc; Riccardo Sarti, MSc; Chiara Pozzi, PhD; Maximiliano Mollura, MSc; Alberto Mantovani, MD; Maria Rescigno, PhD JAMA
• Viewpoint Addressing the Long-term Effects of COVID-19 Rachel L. Levine, MD JAMA
• Medical News & Perspectives The US Now Has a Research Plan for Long COVID—Is It Enough? Jennifer Abbasi JAMA

More than 6 million people have died from COVID-19 worldwide, including nearly 1 million in the US. 1 But mortality is not the only adverse consequence of COVID-19. Many survivors suffer long-term impairment, officially termed postacute sequelae of SARS-CoV-2 infection and commonly called long COVID .

Long COVID—typically defined as symptoms lasting more than 30 days after acute COVID infection—has received some public attention, but it is not nearly as intense as it is for acute COVID-19 infection. Support groups are devoted to the condition, and Congress has allocated more than $1 billion to the National Institutes of Health to study it. But the relatively meager attention that has been paid to long COVID is unfortunate because its health and economic consequences are likely to be every bit as substantial as those due to acute illness.

People who have more severe COVID-19 are more likely to experience long COVID, but severe acute disease is not a prerequisite. Long COVID has been found in people with only mild initial illness. The most common symptom of long COVID is fatigue. 2 More severe cases involve damage to a variety of organ systems (the lungs, heart, nervous system, kidneys, and liver have all been implicated), along with mental health impairment. Researchers have hypothesized that physiological pathways may involve direct consequences of the viral infection along with inflammatory or autoimmune responses.

Because many prevalence estimates are based on convenience samples of members of COVID-19 support groups or people who had severe acute disease, the population prevalence of long COVID is not entirely known. 3 British population data suggest that 22% to 38% of people with the infection will have at least 1 COVID-19 symptom 12 weeks after initial symptom onset, and 12% to 17% will have 3 or more symptoms. 2

Rates this high translate to an enormous number of people with long COVID. The US Centers for Disease Control and Prevention estimates that as of May 5, 2022, the US has had roughly 81 million cases of COVID-19 and 994 187 COVID deaths. Even the lower-end estimate of 12% of people with 3 or more symptoms of long COVID implies that 9.6 million people in the US may have developed long COVID—roughly 10 times the number of COVID-19 deaths. It is not known how long people with long COVID will be symptomatic, but recovery in the first year of long COVID for affected individuals may be very slow. 4

Reduced health is not the only consequence of long COVID. People with the condition work and earn less than they would have otherwise. One survey found that 44% of people with long COVID were out of the labor force and 51% worked fewer hours. 5 In the economy as a whole, more than 1 million people may be out of the workforce at any given time because of long COVID. 6

This reduction in labor supply is a direct earning loss. If 1 million people are out of the labor force because of long COVID, the lost income would be more than $50 billion annually. People out of the workforce because of long COVID disproportionately worked in service jobs , including health care, social care, and retail. 7 The widely noted shortage of workers in these sectors is driving up both wages and prices. Part of the recent surge in inflation in the US may thus be related to long COVID.

People who are no longer able to work may also apply for Social Security Disability Insurance. To date, there has been no sustained increase in disability insurance applications since the onset of COVID-19. This is good news, though it bears watching as disability centers continue reopening from their COVID-19 shutdowns.

Increased medical spending is another consequence of long COVID. The medical costs for treating long COVID have not been estimated, but costs have been estimated for similar conditions. If treatment of long COVID is similar to treatment of myalgic encephalomyelitis (chronic fatigue syndrome), these estimated costs could be about $9000 per person annually. 8

In an October 2020 analysis , we estimated 9 the then-nascent COVID-19 pandemic might result in $2.6 trillion of cost as a result of long COVID. Unfortunately, our estimate seems very much on target.

The massive cost of long COVID has several policy implications. Investing in treatments for long COVID is obviously a high priority. According to a recent report from the Rockefeller Foundation, progress to date has been “achingly slow” and that needs to change. 10 Experimenting with ways to make employment easier for people with long-term complications is also a high priority. People with chronic fatigue may be better able to work at home or with frequent breaks than they can with a time-delimited office day and a long commute. By speeding up the transition to telework, enhanced employment opportunities for those with long COVID may be possible.

In addition, the economic cost of long COVID reinforces the value of comprehensive actions to prevent and treat new infections. Mask mandates are unpopular in many areas and a substantial share of the public resists being vaccinated—though each action should still be encouraged. But additional progress might also be made through expanding rapid COVID-19 test capability, global surveillance to detect new SARS-CoV-2 variants, and immediate action should any such variants be detected. Such measures have associated costs, but no matter how large these costs are, they pale compared with the potential benefits.

Published: May 12, 2022. doi:10.1001/jamahealthforum.2022.1809

Open Access: This is an open access article distributed under the terms of the CC-BY License . © 2022 Cutler DM. JAMA Health Forum .

Corresponding Author: David M. Cutler, PhD, Department of Economics, Harvard University, 1805 Cambridge St, Cambridge, MA 02138 ( [email protected] ).

Conflict of Interest Disclosures: Dr Cutler reported being an expert witness in multidistrict litigation regarding opioids and JUUL products.

See More About

Cutler DM. The Costs of Long COVID. JAMA Health Forum. 2022;3(5):e221809. doi:10.1001/jamahealthforum.2022.1809

Manage citations:

© 2024

Artificial Intelligence Resource Center

Best of JAMA Network 2022

Browse and subscribe to JAMA Network podcasts!

Others Also Liked

• Open access
• Published: 29 July 2024

Predicting hospital length of stay using machine learning on a large open health dataset

• Raunak Jain 1 ,
• Mrityunjai Singh 1 ,
• A. Ravishankar Rao 2 &
• Rahul Garg 1  

BMC Health Services Research volume  24 , Article number:  860 ( 2024 ) Cite this article

227 Accesses

Metrics details

Governments worldwide are facing growing pressure to increase transparency, as citizens demand greater insight into decision-making processes and public spending. An example is the release of open healthcare data to researchers, as healthcare is one of the top economic sectors. Significant information systems development and computational experimentation are required to extract meaning and value from these datasets. We use a large open health dataset provided by the New York State Statewide Planning and Research Cooperative System (SPARCS) containing 2.3 million de-identified patient records. One of the fields in these records is a patient’s length of stay (LoS) in a hospital, which is crucial in estimating healthcare costs and planning hospital capacity for future needs. Hence it would be very beneficial for hospitals to be able to predict the LoS early. The area of machine learning offers a potential solution, which is the focus of the current paper.

We investigated multiple machine learning techniques including feature engineering, regression, and classification trees to predict the length of stay (LoS) of all the hospital procedures currently available in the dataset. Whereas many researchers focus on LoS prediction for a specific disease, a unique feature of our model is its ability to simultaneously handle 285 diagnosis codes from the Clinical Classification System (CCS). We focused on the interpretability and explainability of input features and the resulting models. We developed separate models for newborns and non-newborns.

The study yields promising results, demonstrating the effectiveness of machine learning in predicting LoS. The best R 2 scores achieved are noteworthy: 0.82 for newborns using linear regression and 0.43 for non-newborns using catboost regression. Focusing on cardiovascular disease refines the predictive capability, achieving an improved R 2 score of 0.62. The models not only demonstrate high performance but also provide understandable insights. For instance, birth-weight is employed for predicting LoS in newborns, while diagnostic-related group classification proves valuable for non-newborns.

Our study showcases the practical utility of machine learning models in predicting LoS during patient admittance. The emphasis on interpretability ensures that the models can be easily comprehended and replicated by other researchers. Healthcare stakeholders, including providers, administrators, and patients, stand to benefit significantly. The findings offer valuable insights for cost estimation and capacity planning, contributing to the overall enhancement of healthcare management and delivery.

Peer Review reports

Introduction

Democratic governments worldwide are placing an increasing importance on transparency, as this leads to better governance, market efficiency, improvement, and acceptance of government policies. This is highlighted by reports from the Organization for Economic Co-operation and Development (OECD) an international organization whose mission it is to shape policies that foster prosperity, equality, opportunity and well-being for all [ 1 ]. Openness and transparency have been recognized as pillars for democracy, and also for fostering sustainable development goals [ 2 ], which is a major focus of the United Nations ( https://sustainabledevelopment.un.org/sdg16 ).

An important government function is to provide for the healthcare needs of its citizens. The U.S. spends about $3.6 trillion a year on healthcare, which represents 18% of its GDP [ 3 ]. Other developed nations spend around 10% of their GDP on healthcare. The percentage of GDP spent on healthcare is rising as populations age. Consequently, research on healthcare expenditure and patient outcomes is crucial to maintain viable national economies. It is advantageous for nations to combine investigations by the private sector, government sector, non-profit agencies, and universities to find the best solutions. A promising path is to make health data open, which allows investigators from all sectors to participate and contribute their expertise. Though there are obvious patient privacy concerns, open health data has been made available by organizations such as New York State Statewide Planning and Research Cooperative System (SPARCS) [ 4 ].

Once the data is made available, it needs to be suitably processed to extract meaning and insights that will help healthcare providers and patients. We favor the creation and use of an open-source analytics system so that the entire research community can benefit from the effort [ 5 , 6 , 7 ]. As a concrete demonstration of the utility of our system and approach, we revealed that there is a growing incidence of mental health issues amongst adolescents in specific counties in New York State [ 8 ]. This has resulted in targeted interventions to address these problems in these communities [ 8 ]. Knowing where the problems lie allows policymakers and funding agencies to direct resources where needed.

Healthcare in the U.S. is largely provided through private insurance companies and it is difficult for patients to reliably understand what their expected healthcare costs are [ 9 , 10 ]. It is ironic that consumers can readily find prices of electronics items, books, clothes etc. online, but cannot find information about healthcare as easily. The availability of healthcare information including costs, incidence of diseases, and the expected length of stay for different procedures will allow consumers and patients to make better and more informed choices. For instance, in the U.S., patients can budget pre-tax contributions to health savings accounts, or decide when to opt for an elective surgery based on the expected duration of that procedure.

To achieve this capability, it is essential to have the underlying data and models that interpret the data. Our goal in this paper is twofold: (a) to demonstrate how to design an analytics system that works with open health data and (b) to apply it to a problem of interest to both healthcare providers and patients. Significant advances have been made recently in the fields of data mining, machine-learning and artificial intelligence, with growing applications in healthcare [ 11 ]. To make our work concrete, we use our machine-learning system to predict the length of stay (LoS) in hospitals given the patient information in the open healthcare data released by New York State SPARCS [ 4 ].

The LoS is an important variable in determining healthcare costs, as costs directly increase for longer stays. The analysis by Jones [ 12 ] shows that the trends in LoS, hospital bed capacity and population growth have to be carefully analyzed for capacity planning and to ensure that adequate healthcare can be provided in the future. With certain health conditions such as cardiovascular disease, the hospital LoS is expected to increase due to the aging of the population in many countries worldwide [ 13 ]. During the COVID-19 pandemic, hospital bed capacity became a critical issue [ 14 ], and many regions in the world experienced a shortage of healthcare resources. Hence it is desirable to have models that can predict the LoS for a variety of diseases from available patient data.

The LoS is usually unknown at the time a patient is admitted. Hence, the objective of our research is to investigate whether we can predict the patient LoS from variables collected at the time of admission. By building a predictive model through machine learning techniques, we demonstrate that it is possible to predict the LoS from data that includes the Clinical Classifications Software (CCS) diagnosis code, severity of illness, and the need for surgery. We investigate several analytics techniques including feature selection, feature encoding, feature engineering, model selection, and model training in order to thoroughly explore the choices that affect eventual model performance. By using a linear regression model, we obtain an R 2 value of 0.42 when we predict the LoS from a set of 23 patient features. The success of our model will be beneficial to healthcare providers and policymakers for capacity planning purposes and to understand how to control healthcare costs. Patients and consumers can also use our model to estimate the LoS for procedures they are undergoing or for planning elective surgeries.

Stone et al. [ 15 ] present a survey of techniques used to predict the LoS, which include statistical and arithmetic methods, intelligent data mining approaches and operations-research based methods. Lequertier et al. [ 16 ] surveyed methods for LoS prediction.

The main gap in the literature is that most methods focus on analyzing trends in the LoS or predicting the LoS only for specific conditions or restrict their analysis to data from specific hospitals. For instance, Sridhar et al. [ 17 ] created a model to predict the LoS for joint replacements in rural hospitals in the state of Montana by using a training set with 127 patients and a test set with 31 patients. In contrast, we have developed our model to predict the LoS for 285 different CCS diagnosis codes, over a set of 2.3 million patients over all hospitals in New York state. The CCS diagnosis code refers to the code used by the Clinical Classifications Software system, which encompasses 285 possible diagnosis and procedure categories [ 18 ]. Since the CCS diagnosis codes are too numerous to list, we give a few examples that we analyzed, including but not limited to abdominal hernia, acute myocardial infarction, acute renal failure, behavioral disorders, bladder cancer, Hodgkins disease, multiple sclerosis, multiple myeloma, schizophrenia, septicemia, and varicose veins. To the best of our knowledge, we are not aware of models that predict the LoS on such a variety of diagnosis codes, with a patient sample greater than 2 million records, and with freely available open data. Hence, our investigation is unique from this point of view.

Sotodeh et al. [ 19 ] developed a Markov model to predict the LoS in intensive care unit patients. Ma et al. [ 20 ] used decision tree methods to predict LoS in 11,206 patients with respiratory disease.

Burn et. al. examined trends in the LoS for patients undergoing hip-replacement and knee-replacement in the U.K. [ 21 ]. Their study demonstrated a steady decline in the LoS from 1997–2012. The purpose of their study was to determine factors that contributed to this decline, and they identified improved surgical techniques such as fast-track arthroplasty. However, they did not develop any machine-learning models to predict the LoS.

Hachesu et al. examined the LoS for cardiac disease patients [ 22 ] and found that blood pressure is an important predictor of LoS. Garcia et al. determined factors influencing the LoS for undergoing treatment for hip fracture [ 23 ]. B. Vekaria et al. analyzed the variability of LoS for COVID-19 patients [ 24 ]. Arjannikov et al. [ 25 ] used positive-unlabeled learning to develop a predictive model for LoS.

Gupta et al. [ 26 ] conducted a meta-analysis of previously published papers on the role of nutrition on the LoS of cancer patients, and found that nutrition status is especially important in predicting LoS for gastronintestinal cancer. Similarly, Almashrafi et al. [ 27 ] performed a meta-analysis of existing literature on cardiac patients and reviewed factors affecting their LoS. However, they did not develop quantitative models in their work. Kalgotra et al. [ 28 ] use recurrent neural networks to build a prediction model for LoS.

Daghistani et al. [ 13 ] developed a machine learning model to predict length of stay for cardiac patients. They used a database of 16,414 patient records and predicted the length of stay into three classes, consisting of short LoS (< 3 days), intermediate LoS ( 3–5 days) and long LoS (> 5 days). They used detailed patient information, including blood test results, blood pressure, and patient history including smoking habits. Such detailed information is not available in the much larger SPARCS dataset that we utilized in our study.

Awad et al. [ 29 ] provide a comprehensive review of various techniques to predict the LoS. Though simple statistical methods have been used in the past, they make assumptions that the LoS is normally distributed, whereas the LoS has an exponential distribution [ 29 ]. Consequently, it is preferable to use techniques that do not make assumptions about the distribution of the data. Candidate techniques include regression, classification and regression trees, random forests, and neural networks. Rather than using statistical parametric techniques that fit parameters to specific statistical distributions, we favor data-driven techniques that apply machine-learning.

In 2020, during the height of the COVID-19 pandemic, the Lancet, a premier medical journal drew widespread rebuke [ 30 , 31 , 32 ] for publishing a paper based on questionable data. Many medical journals published expressions of concern [ 33 , 34 ]. The Lancet itself retracted the questionable paper [ 35 ], which is available at [ 36 ] with the stamp “retracted” placed on all pages. One possible solution to prevent such incidents from occurring is for top medical journals to require authors to make their data available for verification by the scientific community. Patient privacy concerns can be mitigated by de-identifying the records made available, as is already done by the New York State SPARCS effort [ 4 ]. Our methodology and analytics system design will become more relevant in the future, as there is a desire to prevent a repetition of the Lancet debacle. Even before the Lancet incident, there was declining trust amongst the public related to medicine and healthcare policy [ 37 ]. This situation continues today, with multiple factors at play, including biased news reporting in mainstream media [ 38 ]. A desirable solution is to make these fields more transparent, by releasing data to the public and explaining the various decisions in terms that the public can understand. The research in this paper demonstrates how such a solution can be developed.

Requirements

We describe the following three requirements of an ideal system for processing open healthcare data

Utilize open-source platforms to permit easy replicability and reproducibility.

Create interpretable and explainable models.

Demonstrate an understanding of how the input features determine the outcomes of interest.

The first requirement captures the need for research to be easily reproduced by peers in the field. There is growing concern that scientific results are becoming hard for researchers to reproduce [ 39 , 40 , 41 ]. This undermines the validity of the research and ultimately hurts the fields. Baker termed this the “reproducibility crisis”, and performed an analysis of the top factors that lead to irreproducibility of research [ 39 ]. Two of the top factors consist of the unavailability of raw data and code.

The second requirement addresses the need for the machine-learning models to produce explanations of their results. Though deep-learning models are popular today, they have been criticized for functioning as black-boxes, and the precise working of the model is hard to discern. In the field of healthcare, it is more desirable to have models that can be explained easily [ 42 ]. Unless healthcare providers understand how a model works, they will be reluctant to apply it in their practice. For instance, Reyes et al. determined that interpretable Artificial Intelligence systems can be better verified, trusted, and adopted in radiology practice [ 43 ].

The third requirement shows that it is important for relevant patient features to be captured that can be related to the outcomes of interest, such as LoS, total cost, mortality rate etc. Furthermore, healthcare providers should be able to understand the influence of these features on the performance of the model [ 44 ]. This is especially critical when feature engineering methods are used to combine existing features and create new features.

In the subsequent sections, we present our design for a healthcare analytics system that satisfies these requirements. We apply this methodology to the specific problem of predicting the LoS.

We have designed the overall system architecture as shown in Fig.  1 . This system is built to handle any open data source. We have shown the New York SPARCS as one of the data sources for the sake of specificity. Our framework can be applied to data from multiple sources such as the Center for Medicare and Medicaid Services (CMS in the U.S.) as shown in our previous work [ 6 ]. We chose a Python-based framework that utilizes Pandas [ 45 ] and Scikit learn [ 46 ]. Python is currently the most popular programming language for engineering and system design applications [ 47 ].

Shows the system architecture. We use Python-based open-source tools such as Pandas and Scikit-Learn to implement the system

In Fig.  2 , we provide a detailed overview of the necessary processing stages. The specific algorithms used in each stage are described in the following sections.

Shows the processing stages in our analytics pipeline

Recent research has shown that it is highly desirable for machine learning models used in the healthcare domain to be explainable to healthcare providers and professionals [ 48 ]. Hence, we focused on the interpretability and explainability of input features in our dataset and the models we chose to explore. We restricted our investigation to models that are explainable, including regression models, multinomial logistic regression, random forests, and decision trees. We also developed separate models for newborns and non-newborns.

Brief description of the dataset

During our investigation, we utilized open-health data provided by the New York State SPARCS system. The data we accessed was from the year 2016, which was the most recent year available at the time. This data was provided in the form of a CSV file, containing 2,343,429 rows and 34 columns. Each row contains de-identified in-patient discharge information. The dataset columns contained various types of information. They included geographic descriptors related to the hospital where care was provided, demographic descriptors such as patient race, ethnicity, and age, medical descriptors such as the CCS diagnosis code, APR DRG code, severity of illness, and length of stay. Additionally, payment descriptors were present, which included information about the type of insurance, total charges, and total cost of the procedure.

Detailed descriptions of all the elements in the data can be found in [ 49 ]. The CCS diagnosis code has been described earlier. The term “DRG” stands for Diagnostic Related Group [ 49 ], which is used by the Center for Medicare and Medicaid services in the U.S. for reimbursement purposes [ 50 ].

The data includes all patients who underwent inpatient procedures at all New York State Hospitals [ 51 ]. The payment for the care can come from multiple sources: Department of Corrections, Federal/State/Local/Veterans Administration, Managed Care, Medicare, Medicaid, Miscellaneous, Private Health Insurance, and Self-Pay. The dataset sourced from the New York State SPARCS system, encompassing a wider patient population beyond Medicare/Medicaid, holds greater value compared to datasets exclusively composed of Medicare/Medicaid patients. For instance, Gilmore et al. analyzed only Medicare patients [ 52 ].

We examine the distribution of the LoS in the dataset, as shown in Fig.  3 . We note that the providers of the data have truncated the length of stay to 120 days. This explains the peak we see at the tail of the distribution.

Distribution of the length of stay in the dataset

Data pre-processing and cleaning

We identified 36,280 samples, comprising 1.55% of the data where there were missing values. These were discarded for further analysis. We removed samples which have Type of Admission = ‘Unknown’ (0.02% samples). So, the final data set has 2,306,668 samples. ‘Payment Typology 2’, and ‘Payment Typology 3’, have missing values (> = 50% samples), which were replaced by a ‘None’ string.

We note that approximately 10% of the dataset consists of rows representing newborns. We treat this group as a separate category. We found that the ‘Birth Weight’ feature had a zero value for non-newborn samples. Accordingly, to better use the ‘Birth Weight’ feature, we partitioned the data into two classes: newborns and non-newborns. This results in two classes of models, one for newborns and the second for all other patients. We removed the ‘Birth Weight’ feature in the input for the non-newborn samples as its value was zero for those samples.

The column ‘Total Costs’ (and in a similar way, ‘Total Charges’) are usually proportional to the LoS, and it would not be fair to use these variables to predict the LoS. Hence, we removed this column. We found that the columns 'Discharge Year', 'Abortion Edit Indicator'' are redundant for LoS prediction models, and we removed them. We also removed the columns ‘CCS Diagnosis Description’, ‘CCS Procedure Description’, ‘APR DRG Description’, ‘APR MDC Description’, and ‘APR Severity of Illness Description’ as we were given their corresponding numerical codes as features.

Since the focus of this paper is on the prediction of the LoS, we analyzed the distribution of LoS values in the dataset.

We developed regression models using all the LoS values, from 1–120. We also developed classification models where we discretized the LoS into specific bins. Since the distribution of LoS values is not uniform, and is heavily clustered around smaller values, we discretized the LoS into a small number of bins, e.g. 6 to 8 bins.

We utilized 10% of the data as a holdout test-set, which was not seen during the training phase. For the remaining 90% of the data, we used tenfold cross-validation in order to train the model and determine the best parameters to use.

Feature encoding

Many variables in the dataset are categorical, e.g., the variable “APR Severity of Illness Description” has the values in the set [Major, Minor, Moderate, Extreme]. We used distribution-dependent target encoding techniques and one-hot techniques to improve the model performance [ 53 ]. We replaced categorical data with the product of mean LoS and median LoS for a category value. The categorical feature can then better capture the dependence distribution of LoS with the value of the categorical feature.

For the linear regression model [ 54 ], we sampled a set of 6 categorical features, [‘Type of Admission’, ‘Patient Disposition’, ‘APR Severity of Illness Code’, ‘APR Medical Surgical Description’, ‘APR MDC Code’] which we target encoded with the mean of the LoS and the median of the LoS. We then one-hot encoded every feature (all features are categorical) and for each such one-hot encoded feature, created a new feature for each of the features in the sampled set, by replacing the ones in the one-hot encoded feature with the value of the corresponding feature in the sampled set. For example, we one-hot encoded ‘Operating Certificate Number’, and for samples where ‘Operating Certificate Number’ was 3, we created 6 features, each where samples having the value 3 were assigned the target encoded values of the sampled set features, and the other samples were assigned zero. We used such techniques to exploit the linear relation between LoS and each feature.

According to the sklearn documentation [ 55 ], a random forest regressor is “a meta estimator that fits a number of decision tree regressors on various sub-samples of the dataset and uses averaging to improve the predictive accuracy and control over-fitting”. The random forest regressor leverages ensemble learning based on many randomized decision trees to make accurate and robust predictions for regression problems. The averaging of many trees protects against single trees overfitting the training data.

The random forest classifier is also an ensemble learning technique and uses many randomized decision trees to make predictions for classification problems. The 'wisdom of crowds' concept suggests that the decision made by a larger group of people is typically better than an individual. The random forest classifier uses this intuition, and allows each decision tree to make a prediction. Finally, the most popular predicted class is chosen as the overall classification.

For the Random Forest Regressor [ 56 , 57 ] and Random Forest Classifier [ 58 ], we only used a similar distribution dependent target encoding as a random forest classifier/ regressor is unsuitable for sparse one-hot encoded columns.

Multinomial logistic regression is a type of regression analysis that predicts the probabilities of the different possible outcomes of a categorically distributed dependent variable, given a set of independent variables. It allows for more than two discrete outcomes, extending binomial logistic regression for binary classification to models with multiple class membership. For the multinomial logistic regression model [ 59 ], we used only one-hot encoding, and not target encoding, as the target value was categorical.

Finally, we experimented with combinations of target encoding and one-hot encoding. We can either use target encoding, or one-hot encoding, or both. When both encodings are employed, the dimensionality of the data increases to accommodate the one-hot encoded features. For each combination of encodings, we also experimented with different regression models including linear regression and random forest regression.

Feature importance, selection, and feature engineering

We experimented with different feature selection methods. Since the focus of our work is on developing interpretable and explainable models, we used SHAP analysis to determine relevant features.

We examine the importance of different features in the dataset. We used the SHAP value (Shapley Additive Explanations), a popular measure for feature importance [ 60 ]. Intuitively, the SHAP value measures the difference in model predictions when a feature is used versus omitted. It is captured by the following formula.

where \({{\varnothing }}_{i}\) is the SHAP value of feature \(i\) , \(p\) is the prediction by the model, n is the number of features and S is any set of features that does not include the feature \(i\) . The specific model we used for the prediction was the random forest regressor where we target-encoded all features with the product of the mean and the median of the LoS, since most of the features were categorical.

Classification models

One approach to the problem is to bin the LoS into different classes, and train a classifier to predict which class an input sample falls in. We binned the LoS into roughly balanced classes as follows: 1 day, 2 days, 3 days, 4–6 days, > 6 days. This strategy is based on the distribution of the LoS as shown earlier in Figs.  3 and  4 .

A density plot of the distribution of the length of stay. The area under the curve is 1. We used a kernel density estimation with a Gaussian kernel [ 61 ] to generate the plot

We used three different classification models, comprising the following:

Multinomial Logistic Regression

Random Forest Classifier

CatBoost classifier [ 62 ].

We used a Multinomial Logistic Regression model [ 59 ] trained and tested using tenfold cross validation to classify the LoS into one of the bins. The multinomial logistic regression model is capable of providing explainable results, which is part of the requirements. We used the feature engineering techniques described in the previous section.

We used a Random Forest Classifier model trained and tested using tenfold cross validation to classify the LoS into one of the bins. We used a maximum depth of 10 so as to get explainable insights into the model.

Finally, we used a CatBoost Classifier model trained and tested using tenfold cross validation to classify the LoS into one of the bins.

Regression models

We used three different regression models with the feature engineering techniques mentioned above ( Feature encoding section). These comprise:

Linear regression

Catboost regression

Random forest regression

The linear regression was implemented using the nn.Linear() function in the open source library PyTorch [ 63 ]. We used the ‘Adam’ optimization algorithm [ 64 ] in mini-batch settings to train the model weights for linear regression.

We investigated CatBoost regression in order to create models with minimal feature sets, whereby models with a low number of input features would provide adequate results. Accordingly, we trained a CatBoost Regressor [ 65 ] in order to determine the relationship between combinations of features and the prediction accuracy as determined by the R 2 correlation score.

The random forest regression was implemented using the function RandomForestRegressor() in scikit learn [ 55 ].

Model performance measures

For the regression models, we used the following metrics to compare the model performance.

The R 2 score and the p -value. We use a significance level of α = 0.05 (5 %) for our statistical tests.  If the p -value is small, i.e. less than α = 0.05, then the R 2 score is statistically significant.

For classifier models, we used the following metrics to compare the model performance.

True positive rate, false negative rate, and F1 score [ 66 ].

We computed the Brier score using Brier’s original calculation in his paper [ 67 ]. In this formulation, for R classes the Brier score B can vary between 0 and R, with 0 being the best score possible.

where \({\widehat{y}}_{i,c}\) is the class probability as per the model and \({I}_{i,c}=1\) if the i th sample belongs to class c and \({I}_{i,c}=0\) if it does not belong to class c .

We used the Delong test [ 68 ] to compare the AUC for different classifiers.

These metrics will allow other researchers to replicate our study and provide benchmarks for future improvements.

In this section we present the results of applying the techniques in the Methods section.

Descriptive statistics

We provide descriptive statistics that help the reader understand the distributions of the variables of interest.

Table 1 summarizes basic statistical properties of the LoS variable.

Figure  5 shows the distribution of the LoS variable for newborns.

This figure depicts the distribution of the LoS variable for newborns

Table 2 shows the top 20 APR DRG descriptions based on their frequency of occurrence in the dataset.

Figure  6 shows the distribution of the LoS variable for the top 20 most frequently occurring APR DRG descriptions shown in Table  2 .

A 3-d plot showing the distribution of the LoS for the top-20 most frequently occuring APR DRG descriptions. The x-axis (horizontal) depicts the LoS, the y-axis shows the APR DRG codes and the z-axis shows the density or frequency of occurrence of the LoS

We experimented with different encoding schemes for the categorical variables and for each encoding we examined different regression techniques. Our results are shown in Table 3 . We experimented with the three encoding schemes shown in the first column. The last row in the table shows a combination of one-hot encoding and target encoding, where the number of columns in the dataset are increased to accommodate one-hot encoded feature values for categorical variables.

Feature importance, selection and feature engineering

We obtained the SHAP plots using a Random Forest Regressor trained with target-encoded features.

Figures  7  and 8 show the SHAP values plots obtained for the features in the newborn partition of the dataset. We find that the features, “APR DRG Code”, “APR Severity of Illness Code”, “Patient Disposition”, “CCS Procedure Code”, are very useful in predicting the LoS. For instance, high feature values for “APR Severity of Illness Code”, which are encoded by red dots have higher SHAP values than the blue dots, which correspond to low feature values.

SHAP Value plot for newborns

1-D SHAP plot, in order of decreasing feature importance: top to bottom (for non-newborns)

A similar interpretation can be applied to the features in the non-newborn partition of the dataset. We note that “Operating Certificate Number” is among the top-10 most important features in both the newborn and non-newborn partitions. This finding is discussed in the Discussion section.

From Fig.  9 , we observe that as the severity of illness code increases from 1–4, there is a corresponding increase in the SHAP values.

A 2-D plot showing the relationship between SHAP values for one feature, “APR Severity of Illness Code”, and the feature values themselves (non-newborns)

To further understand the relationship between the APR Severity of Illness code and the LoS, we created the plot in Fig.  10 . This shows that the most frequently occurring APR Severity of Illness code is 1 (Minor), and that the most frequently occurring LoS is 2 days. We provide this 2-D projection of the overall distribution of the multi-dimensional data as a way of understanding the relationship between the input features and the target variable, LoS.

A density plot showing the relationship between APR Severity of Illness Code and the LoS. The color scale on the right determines the interpretation of colors in the plot. We used a kernel density estimation with a Gaussian kernel [ 61 ] to generate the plot

Similarly, Fig.  11 shows the relationship between the birth weight and the length of stay. The most common length of stay is two days.

A density plot showing the distribution of the birth weight values (in grams) versus the LoS. The colorbar on the right shows the interpretation of color values shown in the plot. We used a kernel density estimation with a Gaussian kernel [ 61 ] to generate the plot

Classification

We obtained a classification accuracy of 46.98% using Multinomial Logistic Regression with tenfold cross-validation in the 5-class classification task for non-newborn cases. The confusion matrix in Fig.  12 shows that the highest density of correctly classified samples is in or close to the diagonal region. The regions where out model fails occurs between adjacent classes as can be inferred from the given confusion matrix.

Confusion matrix for classification of non-newborns. The number inside each square along the diagonal represents the number of correctly classified samples. The color is coded so lighter colors represent lower numbers

For the newborn cases, we obtained a classification accuracy of 60.08% using Random Forest Classification model with tenfold cross-validation in the 5-class classification task. The confusion matrix in Fig.  13 shows that the majority of data samples lie in or close to the diagonal region. The regions where our model does not do well occurs between adjacent classes as can be inferred from the given confusion matrix,

Confusion matrix for classification of newborns. The number inside each square along the diagonal represents the number of correctly classified samples. The color is coded so lighter colors represent lower numbers

The density plot in Fig.  14 shows the relationship between the actual LoS and the predicted LoS. For a LoS of 2 days, the centroid of the predicted LoS cluster is between 2 and 3 days.

Shows the density plot of the predicted length of stay versus actual length of stay for the classifier model for non-newborns. We used a kernel density estimation with a Gaussian kernel [ 61 ] to generate the plot

A quantitative depiction of our model errors is shown in Fig.  15 . The values in Fig.  15 are interpreted as follows. Referring to the column for LoS = 2, the top row shows that 51% of the predicted LoS values for an actual stay of 2 days is also 2 days (zero error), and that 23% of the predicted values for LoS equal to 2 days have an error of 1 day and so on. The relatively high values in the top row indicates that the model is performing well, with an error of less than 1 day. There are relatively few instances of errors between 2 and 3 days (typically less than 10% of the values show up in this row). The only exception is for the class corresponding to LoS great than 8 days. The truncation of the data to produce this class results in larger model errors specifically for this class.

Shows the distribution of correctly predicted LoS values for each class used in our model. Along the columns, we depict the different classes used in the model, consisting of LoS equal to 1, 2, 3 …8, and more than 8. Each row depicts different errors made in the prediction. For instance, the top row depicts an error of less than or equal to one day between the actual LoS and the predicted Los. The second row from the top depicts an error which is greater than 1 and less than or equal 2 days. And so on for the other rows, for non-newborns

Figures  16 and 17 show the scatter plots for the linear regression models. The exact line represents a line with slope 1, and a perfect model would be one that produced all points lying on this line.

Scatter plot showing an instance of a linear regression fit to the data (newborns). The R 2 score is 0.82. The blue line represents an exact fit, where the predicted LoS equals the actual LoS (slope of the line is 1)

Scatter plot for linear regression. (non-newborns). The R 2 score is 0.42. The blue line represents an exact fit, where the predicted LoS equals the actual LoS (slope of the line is 1)

Figure  18 shows a density plot depicting the relationship between the predicted length of stay and the actual length of stay.

Shows the density plot of the predicted length of stay versus actual length of stay for the classifier model for non-newborns. We used a kernel density estimation with a Gaussian kernel [ 40 ] to generate the plot. The best fit regression line to our predictions is shown in green, whereas the blue line represents the ideal fit (line of slope 1, where actual LoS and predicted LoS are equal)

Most of the existing literature on LoS stay prediction is based on data for specific disease conditions such as cancer or cardiac disease. Hence, in order to understand which CCS diagnosis codes produce good model fits, we produced the plot in Fig.  19 .

This figure shows the three CCS diagnosis codes that produced the top three R 2 scores using linear regression. These are 101, 100 and 109. The three CCS Diagnosis codes that produced the lowest R 2 scores are 159, 657, and 659

We provide the following descriptions in Tables  4  and 5 for the 3 CCS Diagnosis Codes in Fig.  19 with the top R 2 Scores using linear regression.

Similarly, the following table shows the 3 CCS Diagnosis Codes in Fig.  19 for the lowest R 2 Scores using linear regression.

Models with minimal feature sets

We trained a CatBoost Regressor [ 65 ] on the complete dataset in order to determine the relationship between combinations of features and the prediction accuracy as determined by the R 2 correlation score. This is shown in Fig.  20

The labels for each row on the left show combinations of different input features. A CatBoost regression model was developed using the selected combination of features. The R 2 correlation scores for each model is shown in the bar graph

We can infer from Fig.  20 that only four features (‘'APR MDC Code', 'APR Severity of Illness Code', 'APR DRG Code', 'Patient Disposition') are sufficient for the model to reach very close to its maximum performance. We obtain similar concurring results when using other regression models for the same experiment.

Classification trees

We used a random forest tree approach to generate the trees in Figs.  21 and 22 .

A random forest tree that represents a best-fit model to the data for newborns. With 4 levels of the decision tree, the R 2 score is 0.65

A random forest tree using only a tree of depth 3 that represents a best-fit model to the data for non-newborns. The R 2 score is 0.28. We can generate trees with greater depth that better fit the data, but we have shown only a depth of 3 for the sake of readability in the printed version of this paper. Otherwise, the tree would be too large to be legible on this page. The main point in this figure is to showcase the ease of interpretation of the working of the model through rules

We used tenfold cross validation to determine the regression scores. The results are summarized in Tables  6 and 7 .

We computed the multi-class classifier metrics for logistic regression, using one-hot encoding for non-newborns. The results are presented in Table  8 . The first row represents the accuracy of the classifier when Class 0 is compared against the rest of the classes. A similar interpretation applies to the other rows in the table, ie one-versus-rest. The macro average gives the balanced recall and precision, and the resulting F1 score. The weighted average gives a support (number of samples) weighted average of the individual class metric. The overall accuracy is computed by dividing the total number of accurate predictions, which is 49,686 out of a total number of 105,932 samples, which yields a value of 0.47.

For the category of non-newborns, Fig.  23  provides a graphical plot that visualizes the ROC curves for the different multiclass classifiers we developed.

This figure applies to data concerning non-newborns. We show the multiclass ROC curves for the performance of the catboost classifier for the different classes shown. The area under the ROC curve is 0.7844

In Table  9 we compare the performance of our multiclass classifier using logistic regression developed on 2016 SPARCS data against 2017 SPARCS data.

In order to compare the performance of the different classifiers, we computed the AUC measures reported in Table  10 . Figure 24 visualizes the data in Table 10 and Fig. 25 visualizes the data in Table 11 . In Tables 12 and 13 we report the results of computing the Delong test for non-newborns and newborns respectively. In Tables 14 and 15 we report the results of computing the Brier scores for non-new borns and newborns respectively.

A bar chart that depicts the data in Table  10 for non-newborns

A bar chart that depicts the data in Table  11

Model parameters

In Table  16 we present the parameter and hyperparameter values used in the different models.

Additional results shown in the Appendix/Supplementary material

Due to space restrictions, we show additional results in the Appendix/Supplementary Material. These results are in tabular form and describe the R 2 scores for different segmentations of the variables in the dataset, e.g. according to age group, severity of illness code, etc.

The most significant result we obtain is shown in Figs.  21 and 22 , which provides an interpretable working of the decision trees using random forest modeling. Figure  21 for newborns shows that the birth weight features prominently in the decision tree, occurring at the root node. Low birth weights are represented on the left side of the tree and are typically associated with longer hospital stays. Higher birth weights occur on the right side of the tree, and the node in the bottom row with 189,574 samples shows that the most frequently occurring predicted stay is 2.66 days. Figure  22 for non-newborns shows that the features of “APR DRG Code”, “APR Severity of Illness Code” and “Patient Disposition” are the most important top-level features to predict the LoS. This provides a relatively simple rule-based model, which can be easily interpreted by healthcare providers as well as patients. For instance, the right-most branch of the tree classifies the input data into a relatively high LoS (46 days) when the branch conditions APR DRG Code is greater than 813.55 and the APR Severity of Illness Code is less than 91.

The results in Fig.  19 and Table  4 show that if we restrict our model to specific CCS Diagnosis descriptions such as “coronary atherosclerosis and other heart disease”, we obtain a good R 2 Score of 0.62. The objective of our work is not to cherry-pick CCS Diagnosis codes that produce good results, but rather to develop a single model for the entire SPARCS dataset to obtain a birds-eye perspective. For future work, we can explicitly build separate models for each CCS Diagnosis code, and that could have relevance to specific medical specialties, such as cardiovascular care.

Similarly, the results in Fig.  19 and Table  5 show that there are CCS Diagnosis codes corresponding to schizophrenia and mood disorders that produce a poor model fit. Factors that contribute to this include the type of data in the SPARCS dataset, where information about patient vitals, medications, or a patient’s income level is not provided, and the inherent variability in treating schizophrenia and mood disorders. Baeza et al. [ 69 ] identified several variables that affect the LoS in psychiatric patients, which include psychiatric admissions in the previous years, psychiatric rating scale scores, history of attempted suicide, and not having sufficient income. Such variables are not provided in the SPARCS dataset. Hence a policy implication is to collect and make such data available, perhaps as a separate dataset focused on mental health issues, which have proven challenging to treat.

Figures  16 and 17 show that a better regression fit is obtained when a specific CCS Diagnosis code is used to build the model, such as “Newborn” in Fig.  16 . To put these results in context, we note that it is difficult to obtain a high R 2 value for healthcare datasets in general, and especially for large numbers of patient samples that span multiple hospitals. For instance, Bertsimas [ 70 ] reported an R 2 value of 0.2 and Kshirsagar [ 71 ] reported an R 2 value of 0.33 for similar types of prediction problems as studied in this paper.

Further details for a segmentation of R 2 scores by the different variable categories are shown in the Appendix/Supplementary Material section. For instance, the table corresponding to Age Groups shows that there is close agreement between the mean of the predicted LoS from our model and the actual LoS. Furthermore, the mean LoS increases steadily from 4.8 days for Age group 0–17 to 6.4 days for ages 70 or older. A discussion of these tables is outside the scope of this paper. However, they are being provided to help other researchers form hypotheses for further investigations or to find supporting evidence for ongoing research.

Table 3 shows that the best encoding scheme is to combine target encoding with one-hot encoding and then apply linear regression. This produces an R 2 score of 0.42 for the non-newborn data, which is the best fit we could obtain. This table also shows that significant improvements can be obtained by exploring the search space which consists of different strategies of feature encoding and regression methods. There is no theoretical framework which determines the optimum choice, and the best method is to conduct an experimental search. An important contribution of the current paper is to explore this search space so that other researchers can use and build upon our methodology.

The distribution of errors in Fig.  15 shows that the truncation we employed at a LoS of 8 days produces artifacts in the prediction model as all stays of greater than 8 days are lumped into one class. Nevertheless, the distribution of LoS values in Fig.  4 shows that a relatively small number of data samples have LoS greater than 8 days. In the future, we will investigate different truncation levels, and this is outside the scope of the current paper. By using our methodology, the truncation level can also be tuned by practitioners in the field, including hospital administrators and other researchers.

Our results in Fig.  7 show that certain features are not useful in predicting the LoS. The SHAP plot shows that features such as race, gender, and ethnicity are not useful in predicting the LoS. It would have been interesting if this were not the case, as that implies that there is systemic bias based on race, gender or ethnicity. For instance, a person with a given race may have a smaller LoS based on their demographic identity. This would be unacceptable in the medical field. It is satisfying to see that a large and detailed healthcare dataset does not show evidence of bias.

To place this finding in context, racial bias is an important area of research in the U.S., especially in fields such as criminology and access to financial services such as loans. In the U.S., it is well known that there is a disproportional imprisonment of black and Hispanic males [ 72 ]. Researchers working on criminal justice have determined that there is racial bias in the process of sentencing and granting parole, with blacks being adversely affected [ 73 ]. This bias is reinforced through any algorithms that are trained on the underlying data. There is evidence that banks discriminate against applicants for loans based on their race or gender [ 74 ].

This does not appear to be the case in our analysis of the SPARCS data. Though we did not specifically investigate the issue of racial bias in the LoS, the feature analysis we conducted automatically provides relevant answers. Other researchers including those in the U.K [ 21 ] have also determined that gender does not have an effect on LoS or costs. Hence the results in the current paper are consistent with the findings of other researchers in other countries working on entirely different datasets.

From Table  6 we see that in the case of data concerning non-newborns, the catboost regression performs the best, with an R 2 score of 0.432. The p -value is less than 0.01, indicating that the correlation between the actual and predicted values of LoS through catboost regression is statistically significant. Similarly, the p -values for linear regression and random forest regression indicate that these models produce predictions that are statistically significant, i.e. they did not occur by random chance.

From Table  7 that refers to data from newborns, the linear regression performs the best, with an R 2 score of 0.82. The p -value is less than 0.01, indicating that the correlation between the actual and predicted values of LoS through linear regression is statistically significant. Similarly, the p -values for random forest regression and catboost regression indicate that these models produce predictions that are statistically significant.

We examine the performance of classifiers on non-newborn data, as shown in Tables  10 and 12 . The Delong test conducted in Table  12 shows that there is a statistically significant difference between the AUCs of the pairwise comparisons of the models. Hence, we conclude that the catboost classifier performs the best with an average AUC of 0.7844. We also note that there is a marginal improvement in performance when we use the catboost classifier instead of the random forest classifier. Both the catboost classifier and the random forest classifier perform better than logistic regression. We conclude that the best performing model for non-newborns is the catboost classifier, followed by the random forest classifier, and then logistic regression.

In the case of newborn data, we examine the performance of the classifiers as shown in Tables  11 and 13 . From Table 13 , we note that the p -values in all the rows are less than 0.05, except for the binary class “one vs. rest for class 3”, random forests vs. catboost. Hence, for this particular comparison between the random forest classifier and the catboost classifier for “one vs. rest for class 3”, we cannot conclude that there is a statistically significant difference between the performance of these two classifiers. From Table  11 we observe that the AUCs of these two classifiers are very similar. We also note that only about 10% of the dataset consists of newborn cases.

From Table  14 we note that the Brier score for the catboost classifier is the lowest. A lower Brier score indicates better performance. According to the Brier scores for the non-newborn data, the catboost classifier performs the best, followed by the random forest classifier and then logistic regression. Table 15 shows that for newborns, the random forest classifier performs the best, followed by the catboost classifier and logistic regression. The performance of the random forest classifier and catboost classifier are very similar.

From a practical perspective, it may make sense to use a catboost classifier on both newborn and non-newborn data as it simplifies the processing pipeline. The ultimate decision rests with the administrators and implementers of these decision systems in the hospital environment.

Burn et al. observe [ 21 ] that though the U.S. has reported similar declines in LoS as in the U.K, the overall costs of joint replacement have risen. The U.K. government created policies to encourage the formation of specialist centers for joint replacement, which have resulted in reduction in the LoS as well as delivering cost reductions. The results and analysis presented in our current paper can help educate patients and healthcare consumers about trends in healthcare costs and how they can be reduced. An informed and educated electorate can press their elected representatives to make changes to the healthcare system to benefit the populace.

Hachesu et al. examined the LoS for cardiac disease patients [ 22 ] where they used data from around 5000 patients and considered 35 input variables to build a predictive model. They found that the LoS was longer in patients with high blood pressure. In contrast, our method uses data from 2.5 million patients and considers multiple disease conditions simultaneously. We also do not have access to patient vitals such as blood pressure measurements, due to the limitation of the existing New York State SPARCS data.

Garcia et al. [ 23 ] conducted a study of elderly patients (age greater than 60) to understand factors governing the LoS for hip fracture treatment. They used 660 patient records and determined that the most significant variable was the American Society of Anesthesiologists (ASA) classification system. The ASA score ranges from 1–5 and captures the anesthesiologist’s impression of a patient’s health and comorbidities at the time of surgery. Garcia et al. showed a monotonically increasing relationship between the ASA score and the LoS. However, they did not build a specific predictive model. Their work shows that it is possible to find single variables with significant information content in order to estimate the LoS. The New York SPARCS dataset that we used does not contain the ASA score. Hence a policy implication of our research is to alert the healthcare authorities include such variables such as the ASA score where relevant in the datasets released in the future. The additional storage required is very small (one additional byte per patient record).

Arjannikov et al. [ 25 ] developed predictive models by binarizing the data into two categories, e.g. LoS <  = 2 days or LoS > 2 days. In our work, we did not employ such a discretization. In contrast, we used continuous regression techniques as well as classification into more than two bins. It is preferable to stay as close to the actual data as possible.

Almashrafi et al. [ 27 ] and Cots et al. [ 75 ] observed that larger hospitals tended to have longer LoS for patients undergoing cardiac surgery. Though we did not specifically examine cardiac surgery outcomes, our feature analysis indicated that the hospital operating certificate number had lower relevance than other features such as DRG codes. Nevertheless, the SHAP plots in Fig.  7 and Fig.  8 show that the hospital operating certificate number occurs within the top 10 features in order of SHAP values. We will investigate this relationship in more detail in future research, as it requires determining the size of the hospital from the operating certificate number and creating an appropriate machine-learning model. The Appendix contains results that show certain operating certificate numbers that produce a good model fit to the data.

A major focus of our research is on building interpretable and explainable models. Based on the principle of parsimony, it is preferable to utilize models which involve fewer features. This will provide simpler explanations to healthcare professionals as well as patients. We have shown through Fig.  20 that a model with five features performs just as well as a model with seven features. These features also make intuitive sense and the model’s operation can be understood by both patients and healthcare providers.

Patients in the U.S. increasingly have to pay for medical procedures out-of-pocket as insurance payments do not cover all the expenses, leading to unexpectedly large bills [ 76 ]. Many patients also do not possess health insurance in the U.S., with the consequence that they get charged the highest [ 77 ]. Kullgreen et.al. observe that patients in the U.S. need to be discerning healthcare consumers [ 78 ], as they can optimize the value they receive from out-of-pocket spending. In addition to estimating the cost of medical procedures, patients will also benefit from estimating the expected duration for a procedure such as joint replacement. This will allow them to budget adequate time for their medical procedures. Patients and consumers will benefit from obtaining estimates from an unbiased open data source such as New York State SPARCS and the use of our model.

Other researchers have developed specific LoS models for particular health conditions, such as cardiac disease [ 22 ], hip replacement [ 21 ], cancer [ 26 ], or COVID-19 [ 24 ]. In addition, researchers typically assume a prior statistical distribution for the outcomes, such a Weibull distribution [ 24 ]. However, we have not made any assumptions of specific prior statistical distributions, nor have we restricted our analysis to specific diseases. Consequently, our model and techniques should be more widely applicable, especially in the face of rapidly changing disease trajectories worldwide.

Our study is based exclusively on freely available open health data. Consequently, we cannot control the granularity of the data and must use the data as-is. We are unable to obtain more detailed patient information such as their physiological variables such as blood pressure, heartrate variability etc. at the time of admittance and during their stay. Hospitals, healthcare providers, and insurers have access to this data. However, there is no mandate for them to make this available to researchers outside their own organizations. Sometimes they sell de-identified data to interested parties such as pharmaceutical companies [ 79 ]. Due to the high costs involved in purchasing this data, researchers worldwide, especially in developing countries are at a disadvantage in developing AI algorithms for healthcare.

There is growing recognition that medical researchers need to standardize data formats and tools used for their analysis, and share them openly. One such effort is the organization for Observational Health Data Sciences and Informatics (OHDSI) as described in [ 80 ].

Twitter has demonstrated an interesting path forward, where a small percentage of its data was made available freely to all users for non-commercial purposes through an API [ 81 ]. Recently, Twitter has made a larger proportion of its data available to qualified academic researchers [ 82 ]. In the future, the profit motives of companies need to be balanced with considerations for the greater public good. An advantage of using the Twitter model is that it spurs more academic research and allows universities to train students and the workforce of the future on real-world and relevant datasets.

In the U.S., a new law went into effect in January 2021 requiring hospitals to make pricing data available publicly. The premise is that having this data would provide better transparency into the working of the healthcare system in the U.S. and lead to cost efficiencies. However, most hospitals are not in compliance with this law [ 83 ]. Concerted efforts by government officials as well as pressure by the public will be necessary to achieve compliance. If the eventual release of such data is not accompanied by a corresponding interest shown by academicians, healthcare researchers, policymakers, and the public it is likely that the very premise of the utility of this data will be called into question. Furthermore, merely dumping large quantities of data into the public domain is unlikely to benefit anyone. Hence research efforts such as the one presented in this paper will be valuable in demonstrating the utility of this data to all stakeholders.

Our machine-learning pipeline can easily be applied to new data that will be released periodically by New York SPARCS, and also to hospital pricing data [ 83 ]. Due to our open-source methodology, other researchers can easily extend our work and apply it to extract meaning from open health data. This improves reproducibility, which is an essential aspect of science. We will make our code available on Github to interested researchers for non-commercial purposes.

Limitations of our models

Our models are restricted to the data available through New York State SPARCS, which does not provide detailed information about patient vitals. More detailed physiological data is available through the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC) framework [ 84 ], though for a smaller number of patients. We plan to extend our methodology to handle such data in the future. Another limitation of our study is that it does not account for patient co-morbidities. This arises from the de-identification process used to release the SPARCS data, where patient information is removed. Hence we are unable to analyze multiple hospital admissions for a given patient, possibly for different conditions. The main advantage of our approach is that it uses large-scale population data (2.3 million patients) but at a coarse level of granularity, where physiological data is not available. Nevertheless, our approach provides a high-level view of the operation of the healthcare system, which provides valuable insights.

There is growing interest in using data analytics to increase government transparency and inform policymaking. It is expected that the meaning and insights gained from such evidence-based analysis will translate to better policies and optimal usage of the available infrastructure. This requires cooperation between computer scientists, domain experts, and policy makers. Open healthcare data is especially valuable in this context due to its economic significance. This paper presents an open-source analytics system to conduct evidence-based analysis on openly available healthcare data.

The goal is to develop interpretable machine learning models that identify key drivers and make accurate predictions related to healthcare costs and utilization. Such models can provide actionable insights to guide healthcare administrators and policy makers. A specific illustration is provided via a robust machine learning pipeline that predicts hospital length of stay across 285 disease categories based on 2.3 million de-identified patient records. The length of stay is directly related to costs.

We focused on the interpretability and explainability of input features and the resulting models. Hence, we developed separate models for newborns and non-newborns, given differences in input features. The best performing model for non-newborn data was catboost regression, which used linear regression and achieved an R 2 score of 0.43. The best performing model for newborns and non-newborns respectively was linear regression, which achieved an R 2 score of 0.82. Key newborn predictors included birth weight, while non-newborn models relied heavily on the diagnostic related group classification. This demonstrates model interpretability, which is important for adoption. There is an opportunity to further improve performance for specific diseases. If we restrict our analysis to cardiovascular disease, we obtain an improved R 2 score of 0.62.

The presented approach has several desirable qualities. Firstly, transparency and reproducibility are enabled through the open-source methodology. Secondly, the model generalizability facilitates insights across numerous disease states. Thirdly, the technical framework can easily integrate new data while allowing modular extensions by the research community. Lastly, the evidence generated can readily inform multiple key stakeholders including healthcare administrators planning capacity, policy makers optimizing delivery, and patients making medical decisions.

Availability of data and materials

Data is publicly available at the website mentioned in the paper, https://www.health.ny.gov/statistics/sparcs/

There is an “About Us” tab in the website which contains all the contact details. The authors have nothing to do with this website as it is maintained by New York State.

Gurría A. Openness and Transparency - Pillars for Democracy, Trust and Progress. OECD.org. Available: https://www.oecd.org/unitedstates/opennessandtransparency-pillarsfordemocracytrustandprogress.htm . Accessed 28 June 2024.

Jetzek T. The Sustainable Value of Open Government Data: Uncovering the Generative Mechanisms of Open Data through a Mixed Methods Approach. lCopenhagen Business School, Institut for IT-Ledelse Department of IT Management. 2015.

Move fast and heal things: How health care is turning into a consumer product. The Economist. 2022.  https://www.economist.com/business/how-health-care-is-turning-into-a-consumer-product/21807114 . Accessed 28 June 2024.

New York State Department Of Health, Statewide Planning and Research Cooperative System (SPARCS).  https://www.health.ny.gov/statistics/sparcs/ . Accessed 5 Oct 2022.

Rao AR, Chhabra A, Das R, Ruhil V. A framework for analyzing publicly available healthcare data. In 2015 17th International Conference on E-health Networking, Application & Services (IEEE HealthCom). 2015: IEEE, pp. 653–656.

Rao AR, Clarke D. A fully integrated open-source toolkit for mining healthcare big-data: architecture and applications. In IEEE International Conference on Healthcare Informatics ICHI, Chicago. 2016: IEEE, pp. 255–261.

Rao AR, Garai S, Dey S, Peng H. PIKS: A Technique to Identify Actionable Trends for Policy-Makers Through Open Healthcare Data. SN Computer Science. 2021;2(6):1–22.

Article   Google Scholar  

Rao AR, Rao S, Chhabra R. Rising mental health incidence among adolescents in Westchester, NY. Community Ment Health J. 2021:1–1. 

Boylan J F. My $145,000 Surprise Medical Bill. New York Times. 2020.  https://www.nytimes.com/2020/02/19/opinion/surprise-medical-bill.html . Accessed 28 June 2024.

Peterson K, Bykowicz J. Congress Debates Push to End Surprise Medical Billing. Wall Street J. 2020.  https://www.wsj.com/articles/congress-debates-push-to-end-surprise-medical-billing-11589448603 . Accessed 28 June 2024.

Wang S, Zhang J, Fu Y, Li Y. ACM TIST Special Issue on Deep Learning for Spatio-Temporal Data: Part 1. 12th ed. NY: ACM New York; 2021. p. 1–3.

Google Scholar  

Jones R. lining length of stay and future bed numbers. BJHCM. 2015;21(9):440–1.

Daghistani TA, Elshawi R, Sakr S, Ahmed AM, Al-Thwayee A, Al-Mallah MH. Predictors of in-hospital length of stay among cardiac patients: a machine learning approach. Int J Cardiol. 2019;288:140–7.

Article   PubMed   Google Scholar  

Sen-Crowe B, Sutherland M, McKenney M, Elkbuli A. A closer look into global hospital beds capacity and resource shortages during the COVID-19 pandemic. J Surg Res. 2021;260:56–63.

Article   CAS   PubMed   Google Scholar  

Stone K, Zwiggelaar R, Jones P, Mac Parthaláin N. A systematic review of the prediction of hospital length of stay: Towards a unified framework. PLOS Digital Health. 2022;1(4):e0000017.

Article   PubMed   PubMed Central   Google Scholar  

Lequertier V, Wang T, Fondrevelle J, Augusto V, Duclos A. Hospital length of stay prediction methods: a systematic review. Med Care. 2021;59(10):929–38.

Sridhar S, Whitaker B, Mouat-Hunter A, McCrory B. Predicting Length of Stay using machine learning for total joint replacements performed at a rural community hospital. PLoS ONE. 2022;17(11);e0277479.

Article   CAS   PubMed   PubMed Central   Google Scholar  

CCS (Clinical Classifications Software) - Synopsis. https://www.nlm.nih.gov/research/umls/sourcereleasedocs/current/CCS/index.html . Accessed 13 Jan 2022.

Sotoodeh M, Ho JC. Improving length of stay prediction using a hidden Markov model. AMIA Summits on Translational Science Proceedings. 2019;2019:425.

PubMed Central   Google Scholar  

Ma F, Yu L, Ye L, Yao DD, Zhuang W. Length-of-stay prediction for pediatric patients with respiratory diseases using decision tree methods. IEEE J Biomed Health Inform. 2020;24(9):2651–62.

Burn E, et al. Trends and determinants of length of stay and hospital reimbursement following knee and hip replacement: evidence from linked primary care and NHS hospital records from 1997 to 2014. BMJ Open. 2018;8(1);e019146.

Hachesu PR, Ahmadi M, Alizadeh S, Sadoughi F. Use of data mining techniques to determine and predict length of stay of cardiac patients. Healthcare informatics research. 2013;19(2):121–9.

Garcia AE, et al. Patient variables which may predict length of stay and hospital costs in elderly patients with hip fracture. J Orthop Trauma. 2012;26(11):620–3.

Vekaria B, et al. Hospital length of stay for COVID-19 patients: Data-driven methods for forward planning. BMC Infect Dis. 2021;21(1):1–15.

Arjannikov T, Tzanetakis G. An empirical investigation of PU learning for predicting length of stay. In 2021 IEEE 9th International Conference on Healthcare Informatics (ICHI). 2021: IEEE, pp. 41–47.

Gupta D, Vashi PG, Lammersfeld CA, Braun DP. Role of nutritional status in predicting the length of stay in cancer: a systematic review of the epidemiological literature. Ann Nutr Metab. 2011;59(2–4):96–106.

Almashrafi A, Elmontsri M, Aylin P. Systematic review of factors influencing length of stay in ICU after adult cardiac surgery. BMC Health Serv Res. 2016;16(1):318.

Kalgotra P, Sharda R. When will I get out of the hospital? Modeling Length of Stay using Comorbidity Networks. J Manag Inf Syst. 2021;38(4):1150–84.

Awad A, Bader-El-Den M, McNicholas J. Patient length of stay and mortality prediction: a survey. Health Serv Manage Res. 2017;30(2):105–20.

Editorial-Board. The Lancet, HCL and Trump. Wall Street J. 2020.  https://www.wsj.com/articles/the-lancet-hcl-and-trump-11591226880 . Accessed 28 June 2024.

Servick  K, Enserink M. A mysterious company’s coronavirus papers in top medical journals may be unraveling. Science. 2020.  https://www.science.org/content/article/mysterious-company-s-coronavirus-papers-top-medical-journals-may-be-unraveling . Accessed 28 June 2024.

Gabler E, Rabin RC. The Doctor Behind the Disputed Covid Data. New York Times. 2020.  https://www.nytimes.com/2020/07/27/science/coronavirus-retracted-studies-data.html . Accessed 28 June 2024.

Lancet-Editors. Expression of concern: Hydroxychloroquine or chloroquine with or without a macrolide for treatment of COVID-19: a multinational registry analysis. 2020;395:10240. https://www.science.org/content/article/mysterious-company-s-coronavirus-papers-topmedical-journals-may-be-unraveling . Accessed 28 June 2024.

Editorial-Board. Expression of Concern: Mehra MR et al. Cardiovascular Disease, Drug Therapy, and Mortality in Covid-19. N Engl J Med. 2020.  https://www.nejm.org/doi/full/10.1056/NEJMoa2007621 . Accessed 28 June 2024.

Hopkins JS, Gold R. Authors Retract Studies That Found Risks of Using Antimalaria Drugs Against Covid-19. Wall Street J. 2020. https://www.wsj.com/articles/authors-retract-study-that-found-risks-of-using-antimalaria-drug-against-covid-19-11591299329 . Accessed 28 June 2024.

https://www.thelancet.com/pdfs/journals/lancet/PIIS0140-6736(20)31180-6.pdf . Accessed 9 Jan 2022.

Wolfensberger M, Wrigley A. Trust in Medicine. Cambridge University Press. 2019. ISBN-13: 978-1108487191.

Bhattacharya J, Nicholson T. A Deceptive Covid Study, Unmasked. Wall Street J. 2022. https://www.wsj.com/articles/deceptive-covid-study-unmasked-abc-misleading-omicron-north-carolina-students-duke-mask-test-to-stay-11641933613 . Accessed 28 June 2024.

Baker M. 1,500 scientists lift the lid on reproducibility. Nature. 2016;533(7604):452–4.

Begley CG, Ioannidis JP. Reproducibility in science: improving the standard for basic and preclinical research. Circ Res. 2015;116(1):116–26.

Eisner D. Reproducibility of science: Fraud, impact factors and carelessness. J Mol Cell Cardiol. 2018;114:364–8.

Wang F, Kaushal R, Khullar D. Should health care demand interpretable artificial intelligence or accept “black box” medicine? Am College Phys. 2020;172:59–60.

Reyes M, et al. On the interpretability of artificial intelligence in radiology: challenges and opportunities. Radiol Art Intell. 2020;2(3):e190043.

Savadjiev P, et al. Demystification of AI-driven medical image interpretation: past, present and future. Eur Radiol. 2019;29(3):1616–24.

McKinney W. Python for data analysis: Data wrangling with Pandas, NumPy, and IPython. " O’Reilly Media, Inc. 2012.

Pedregosa F, et al. Scikit-learn: Machine learning in Python. J Machine Learn Res. 2011;12:2825–30.

Cass S. The top programming languages: Our latest rankings put Python on top-again-[Careers]. IEEE Spectr. 2020;57(8):22–22.

Tjoa E, Guan C. A survey on explainable artificial intelligence (xai): Toward medical xai," IEEE Transactions on Neural Networks and Learning Systems. 2020.

https://www.health.ny.gov/statistics/sparcs/docs/sparcs_data_dictionary.xlsx . Accessed 28 June 2024.

Design and development of the Diagnosis Related Group (DRG). https://www.cms.gov/icd10m/version37-fullcode-cms/fullcode_cms/Design_and_development_of_the_Diagnosis_Related_Group_(DRGs).pdf . Accessed 5 Oct 2022.

ARTICLE 28, Hospitals, Public Health (PBH) CHAPTER 45. 2023. Available: https://www.nysenate.gov/legislation/laws/PBH/A28 . Accessed 28 June 2024.

Gilmore‐Bykovskyi A, et al. Disparities in 30‐day readmission rates among Medicare enrollees with dementia. J Am Geriatr Soc. 2023.

Rodríguez P, Bautista MA, Gonzalez J, Escalera S. Beyond one-hot encoding: Lower dimensional target embedding. Image Vis Comput. 2018;75:21–31.

Montgomery DC, Peck EA, Vining GG. Introduction to linear regression analysis. 6th ed. John Wiley & Sons; 2021. ISBN-13 978-1119578727.

Random forest regressor in sklearn. Available: https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html . Accessed 28 June 2024.

Breiman L. Random forests. Mach Learn. 2001;45:5–32.

Svetnik V, Liaw A, Tong C, Culberson JC, Sheridan RP, Feuston BP. Random forest: a classification and regression tool for compound classification and QSAR modeling. J Chem Inf Comput Sci. 2003;43(6):1947–58.

Liaw A, Wiener M. Classification and regression by randomForest. R news. 2002;2(3):18–22.

Böhning D. Multinomial logistic regression algorithm. Ann Inst Stat Math. 1992;44(1):197–200.

Vaid A, et al. Machine Learning to Predict Mortality and Critical Events in a Cohort of Patients With COVID-19 in New York City: Model Development and Validation. J Med Internet Res. 2020;22(11);e24018.

Density Estimation.  https://scikit-learn.org/stable/modules/density.html . Accessed 5 Oct 2022.

CatBoost, a high-performance open source library for gradient boosting on decision trees. Available:  https://catboost.ai/  and https://catboost.ai/en/docs/concepts/python-usages-examples . Accessed 28 June 2024.

PyTorch documentation for torch.nn, the basic building blocks for graphs. Available: https://pytorch.org/docs/stable/nn.html . Accessed 28 June 2024.

Kingma DP, Ba J. Adam: A method for stochastic optimization," arXiv preprint arXiv:1412.6980. 2014.

Prokhorenkova L, Gusev G, Vorobev A, Dorogush AV, Gulin A. CatBoost: unbiased boosting with categorical features," arXiv preprint arXiv:1706.09516. 2017.

Tharwat A. Classification assessment methods. Applied computing and informatics. 2020;17(1):168–92.

Brier GW. Verification of forecasts expressed in terms of probability. Mon Weather Rev. 1950;78(1):1–3.

DeLong ER, DeLong DM, Clarke-Pearson DL. Comparing the areas under two or more correlated receiver operating characteristic curves: a nonparametric approach. Biometrics. 1988:837–45.

Baeza FL, da Rocha NS, Fleck MP. Predictors of length of stay in an acute psychiatric inpatient facility in a general hospital: a prospective study. Brazilian Journal of Psychiatry. 2017;40:89–96.

Bertsimas D, et al. Algorithmic prediction of health-care costs. Oper Res. 2008;56(6):1382–92.

Kshirsagar R. Accurate and Interpretable Machine Learning for Transparent Pricing of Health Insurance Plans," presented at the AAAI 2021 Conference. 2021.

Ulmer J, Painter-Davis N, Tinik L. Disproportional imprisonment of Black and Hispanic males: Sentencing discretion, processing outcomes, and policy structures. Justice Q. 2016;33(4):642–81.

Angwin J, J. Larso J, Mattu S, Kirchner L. Machine bias: There’s software used across the country to predict future criminals. And it’s biased against blacks. ProPublica (2016). Google Scholar. 2016;23.

Steil JP, Albright L, Rugh JS, Massey DS. The social structure of mortgage discrimination. Hous Stud. 2018;33(5):759–76.

Cots F, Mercadé L, Castells X, Salvador X. Relationship between hospital structural level and length of stay outliers: Implications for hospital payment systems. Health Policy. 2004;68(2):159–68.

Evans M, McGinty T. Hospital Prices Are Arbitrary. Just Look at the Kingsburys’ $100,000 Bill. Wall Street J. 2021.  https://www.wsj.com/articles/hospital-prices-arbitrary-healthcare-medical-bills-insurance-11635428943 . Accessed 28 June 2024.

Evans M. Hospitals Often Charge Uninsured People the Highest Prices, New Data Show. Wall Street J. 2021. https://www.wsj.com/articles/hospitals-often-charge-uninsured-people-the-highest-prices-new-data-show-11625584448 . Accessed 28 June 2024.

Kullgren JT, et al. A survey of Americans with high-deductible health plans identifies opportunities to enhance consumer behaviors. Health Aff. 2019;38(3):416–24.

Wetsman N. Hospitals are selling treasure troves of medical data — what could go wrong? The Verge. 2021. Available: https://www.theverge.com/2021/6/23/22547397/medical-records-health-data-hospitals-research . Accessed 28 June 2024.

Hripcsak G, et al. Observational Health Data Sciences and Informatics (OHDSI): Opportunities for Observational Researchers. Stud Health Technol Inform. 2015;216:574–8.

PubMed   PubMed Central   Google Scholar  

Gabarron E, Dorronzoro E, Rivera-Romero O, Wynn R. Diabetes on Twitter: a sentiment analysis. J Diabetes Sci Technol. 2019;13(3):439–44.

Statt N. Twitter is opening up its full tweet archive to academic researchers for free. The Verge. 2021. Available: https://www.theverge.com/2021/1/26/22250203/twitter-academic-research-public-tweet-archive-free-access . Accessed 28 June 2024. 

Evans M, Mathews AW, McGinty T. Hospitals Still Not Fully Complying With Federal Price-Disclosure Rules. Wall Street J. 2021.  https://www.wsj.com/articles/hospital-price-public-biden-11640882507 .

Johnson AE, et al. MIMIC-III, a freely accessible critical care database. Scientific data. 2016;3(1):1–9.

Download references

Acknowledgements

We are grateful to the New York State SPARCS program for making the data available freely to the public. We greatly appreciate the feedback provided by the anonymous reviewers which helped in improving the quality of this manuscript.

No external funding was available for this research.

Author information

Authors and affiliations.

Indian Institute of Technology, Delhi, India

Raunak Jain, Mrityunjai Singh & Rahul Garg

Fairleigh Dickinson University, Teaneck, NJ, USA

A. Ravishankar Rao

You can also search for this author in PubMed   Google Scholar

Contributions

Raunak Jain, Mrityunjai Singh, A. Ravishankar Rao, and Rahul Garg contributed equally to all stages of preparation of the manuscript.

Corresponding author

Correspondence to A. Ravishankar Rao .

Ethics declarations

Ethics approval and consent to participate.

Not applicable as no human subjects were used in our study.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.

Additional information

Publisher’s note.

Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.

Supplementary Information

Supplementary material 1., rights and permissions.

Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by/4.0/ . The Creative Commons Public Domain Dedication waiver ( http://creativecommons.org/publicdomain/zero/1.0/ ) applies to the data made available in this article, unless otherwise stated in a credit line to the data.

Reprints and permissions

About this article

Cite this article.

Jain, R., Singh, M., Rao, A.R. et al. Predicting hospital length of stay using machine learning on a large open health dataset. BMC Health Serv Res 24 , 860 (2024). https://doi.org/10.1186/s12913-024-11238-y

Download citation

Received : 19 June 2023

Accepted : 24 June 2024

Published : 29 July 2024

DOI : https://doi.org/10.1186/s12913-024-11238-y

Share this article

Anyone you share the following link with will be able to read this content:

Sorry, a shareable link is not currently available for this article.

Provided by the Springer Nature SharedIt content-sharing initiative

• Machine learning
• Artificial intelligence
• Health informatics
• Open-source software
• Healthcare analytics

BMC Health Services Research

ISSN: 1472-6963

• General enquiries: [email protected]