Forecasting Mortality: How Hospital Algorithms Shape End-of-Life Care

By: Parmis Mokhtari-Dizaji

Edited By: Stephen Shiwei Wang

Hospitals across the United States are increasingly adopting predictive algorithms designed to estimate a patient’s likelihood of dying within a given period. These “mortality risk” algorithms analyze electronic health records, laboratory results, and physiological measurements to generate statistical models that predict patient outcomes. Researchers note that such models can help clinicians “anticipate more accurately a patient’s timing of death and assess their possibility of recovery,” allowing medical teams to identify patients who may require earlier clinical attention.1 In practice, these tools are often framed as planning aids that help clinicians determine a patient’s course of care.2

Yet a statistical prediction of death is not the same as a clinical diagnosis.3 Once mortality risk scores appear in a patient’s medical chart, they can shape how clinicians interpret symptoms, prioritize treatment decisions, and allocate their attention within busy hospital settings. As mortality prediction models become embedded in everyday clinical workflows, they raise important policy questions about oversight, transparency, and how algorithmic forecasts should be governed once they influence routine clinical practice.

The Expansion of Mortality Prediction

Efforts to estimate mortality risk have existed in medicine for decades. Clinicians have long relied on scoring systems designed to measure illness severity among critically ill patients, such as the Acute Physiology and Chronic Health Evaluation (APACHE) and Sequential Organ Failure Assessment (SOFA) scores. These scoring systems combine physiological measurements and laboratory values to estimate illness severity and a patient’s associated mortality risk. For example, the APACHE score classifies intensive-care patients “prognostically by risk of death” using physiological indicators collected at admission, while the SOFA score evaluates organ dysfunction across multiple systems to track whether a patient’s condition is improving or deteriorating during treatment in intensive care.4,5 

As these scoring systems remain widely used in critical care, advances in artificial intelligence, particularly machine learning algorithms, have enabled new approaches to mortality prediction based on large-scale clinical data.6 The growing availability of digital clinical data has made these methods possible, as electronic health records (EHRs) now contain extensive physiological and treatment information collected during routine care.7 In one study using electronic health record data, researchers found that “data-driven models used on an extended dataset can outperform conventional models for prognosis.”8 Comparisons with traditional severity scoring systems similarly suggest that machine-learning approaches may achieve higher accuracy in predicting mortality in intensive care settings.9 However, many existing prediction models have been tested only within the datasets from which they were developed rather than on independent external datasets, as “few studies validated predictions using independent data,” raising concerns that their reported accuracy may not generalize to all clinical settings.10

As these systems become more sophisticated, hospitals are increasingly incorporating mortality prediction models directly into EHR platforms. In some cases, these models are “integrated directly into the transactional, user-facing database layer of the EHR,” making results “immediately available to clinical teams” and “displayed prominently in each patient’s chart.”11 National survey data suggest that such integration has become widespread: by 2024, approximately 71 percent of U.S. hospitals reported using predictive AI embedded within their EHR systems.12 As a result, predictive models become part of the information clinicians encounter while reviewing patient charts and making clinical decisions. This integration raises important questions about how these predictions influence clinical judgment in practice.

When Predictions Shape Treatment Decisions

Although mortality prediction models are intended to assist clinicians rather than replace their judgment, their presence can still influence how care is delivered in practice. Risk scores embedded within EHR systems can shape how physicians interpret a patient’s condition, particularly in fast-paced hospital environments where clinicians must evaluate multiple patients with competing needs. These models operate within broader clinical decision-support systems designed to “improve healthcare delivery by enhancing medical decisions with targeted clinical knowledge, patient information, and other health information.”13 

In some cases, predictive models may improve care by helping clinicians identify patients whose conditions are deteriorating sooner than expected. Earlier recognition of high-risk patients may also support more timely conversations about treatment goals that might otherwise occur later in a patient’s hospitalization.14 However, clinicians have also expressed concern that mortality predictions may unintentionally shape clinical judgment if algorithmic forecasts are interpreted too definitively in practice. Interviews with physicians using predictive models reveal similar worries. In one study examining the implementation of artificial intelligence in emergency medicine, a physician warned that mortality predictions could be interpreted “as giving the patient a death sentence” if clinicians rely too heavily on the model’s output.15 In such cases, questions of responsibility become increasingly complex, as algorithmic predictions may influence treatment decisions while clinicians remain responsible for interpreting and acting on the model’s output.

Policy and Governance Challenges

As mortality prediction models become more integrated into clinical practice, questions about how they should be governed have become increasingly important. The growing use of mortality prediction algorithms has outpaced the development of governance frameworks capable of overseeing their clinical impact. Many of these systems have been implemented in healthcare settings without formal evaluation requirements in place, leaving oversight responsibilities largely to the institutions deploying them.16

This governance gap is also reflected in the uncertain regulatory status of many predictive clinical tools. The Food and Drug Administration (FDA) has explained that Section 3060(a) of the 21st Century Cures Act amended the Federal Food, Drug, and Cosmetic Act by “removing certain software functions from the definition of device.”17 As a result, some predictive algorithms used in hospitals may influence clinical decisions while still falling outside the regulatory review typically required for other medical technologies. Under current FDA guidance, one key dividing line is whether a clinician can independently review the basis for an algorithm-based recommendation rather than rely primarily on the program’s output. As predictive algorithms become more complex and opaque, however, this standard becomes increasingly difficult to satisfy. 

Beyond questions of regulatory classification or technical performance, the growing use of predictive algorithms in healthcare also raises broader ethical concerns. Because machine learning models are trained on historical medical data, they may reproduce biases already present in healthcare systems, potentially shaping clinical recommendations in ways that clinicians and patients struggle to detect. Research on algorithmic decision-making has shown that predictive algorithms can produce “problematic decisions that reflect biases inherent in the data used to train them.”18 The study also cautions that “remaining ignorant about the construction of machine-learning systems or allowing them to be constructed as black boxes could lead to ethically problematic outcomes.” Empirical evidence suggests that this oversight gap may already affect how predictive tools are evaluated in practice. A national analysis of 2,425 U.S. hospitals found that while 65 percent reported using AI-assisted predictive models, fewer than half had evaluated those systems for potential bias.19 Although 61 percent of hospitals reported testing models for accuracy, only 44 percent conducted similar evaluations for bias, raising concerns about whether predictive algorithms are being systematically assessed for fairness or equity.20

Governing Predictive Medicine

The policy challenge, therefore, is not simply technical accuracy but institutional responsibility. Because many predictive models are developed internally by hospitals or technology vendors and integrated directly into electronic health record platforms, they may fall outside traditional regulatory pathways designed for medical devices. In practice, this can leave hospitals responsible for overseeing these systems themselves, placing institutions “in the position of self-regulation without clear guidance” regarding how predictive clinical decision-support tools should be validated or monitored.21

These challenges raise important questions about how predictive models should be evaluated, monitored, and deployed. As hospitals rely more heavily on predictive analytics, clearer standards for algorithm validation, independent auditing, and clinical oversight will become increasingly necessary. National surveys suggest that many hospitals are beginning to adopt multidisciplinary governance structures for predictive AI oversight, with approximately three-quarters reporting that multiple institutional entities share responsibility for evaluating these systems.22 While such collaborative oversight models may improve accountability, the variation in governance practices across institutions suggests that clearer national standards may still be needed.

As predictive systems become a routine part of medical infrastructure, ensuring that their use remains accountable, transparent, and clinically responsible will be essential for maintaining trust in data-driven medical decision-making. Without clear governance, mortality prediction systems risk quietly shaping life-and-death decisions while remaining largely outside the structures designed to ensure accountability in modern healthcare.

Work Cited​​

  1. Olang, Orkideh, et al. 2025. “Artificial Intelligence-Based Models for Prediction of Mortality in ICU Patients: A Scoping Review.” Journal of Intensive Care Medicine 40 (12): 1240–1246. https://doi.org/10.1177/08850666241277134. 
  2. Courtright, Katherine R., et al. 2019. “Electronic Health Record Mortality Prediction Model for Targeted Palliative Care Among Hospitalized Medical Patients: A Pilot Quasi-experimental Study.” Journal of General Internal Medicine 34 (9): 1841–1847. https://doi.org/10.1007/s11606-019-05169-2. 
  3. Sutton, Richard T., et al. 2020. “An Overview of Clinical Decision Support Systems: Benefits, Risks, and Strategies for Success.” npj Digital Medicine 3: 17. https://doi.org/10.1038/s41746-020-0221-y. 
  4. Knaus, William A., et al. 1985. “APACHE II: A Severity of Disease Classification System.” Critical Care Medicine 13 (10): 818–829. 
  5. Vincent, Jean-Louis, et al. 1996. “The SOFA (Sepsis-Related Organ Failure Assessment) Score to Describe Organ Dysfunction/Failure.” Intensive Care Medicine 22 (7): 707–710. https://doi.org/10.1007/BF01709751. 
  6. Steele, Andrew J., et al. 2018. “Machine Learning Models in Electronic Health Records Can Outperform Conventional Survival Models for Predicting Patient Mortality in Coronary Artery Disease.” PLOS ONE 13 (8): e0202344. https://doi.org/10.1371/journal.pone.0202344. 
  7. Shillan, Duncan, et al. 2019. “Use of Machine Learning to Analyse Routinely Collected Intensive Care Unit Data: A Systematic Review.” Critical Care 23 (1): 284. https://doi.org/10.1186/s13054-019-2564-9. 
  8. Steele et al., “Machine Learning Models in Electronic Health Records.”
  9. Chou, Ruey-Hsing, et al. 2024. “Machine-learning Models Are Superior to Severity Scoring Systems for the Prediction of the Mortality of Critically Ill Patients in a Tertiary Medical Center.” Journal of the Chinese Medical Association 87 (4): 369–376. https://doi.org/10.1097/JCMA.0000000000001066. 
  10. Shillan et al., “Use of Machine Learning to Analyse Routinely Collected Intensive Care Unit Data.”
  11. Mou, Zongyang, et al. 2022. “Electronic Health Record Machine Learning Model Predicts Trauma Inpatient Mortality in Real Time: A Validation Study.” Journal of Trauma and Acute Care Surgery 92 (1): 74–80. https://doi.org/10.1097/TA.0000000000003431. 
  12. Chang, Wei, et al. 2025. Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023–2024. Washington, DC: Office of the Assistant Secretary for Technology Policy. https://www.ncbi.nlm.nih.gov/books/NBK618497/. 
  13. Sutton et al., “An Overview of Clinical Decision Support Systems.”
  14. Alabbasi, Abdullah, Muhanad Alzahrani, Faris Sultan, and Mohammed Sayes. 2026. “Artificial Intelligence-Powered Predictive Tools to Improve End-of-Life Decision-Making: Mini-Review.” BMJ Supportive & Palliative Care, February 18. https://doi.org/10.1136/spcare-2025-006066. 
  15. Petersson, Lena, Kalista Vincent, Petra Svedberg, Jens M. Nygren, and Ingrid Larsson. 2023. “Ethical Considerations in Implementing AI for Mortality Prediction in the Emergency Department: Linking Theory and Practice.” Digital Health 9: 1–16. https://doi.org/10.1177/20552076231206588. 
  16. Weissman, Gary E. 2021. “FDA Regulation of Predictive Clinical Decision-Support Tools: What Does It Mean for Hospitals?” Journal of Hospital Medicine 16 (4): 244–246. https://doi.org/10.12788/jhm.3450. 
  17. U.S. Food and Drug Administration. 2019. Changes to Existing Medical Software Policies Resulting from Section 3060 of the 21st Century Cures Act. https://www.fda.gov/media/109622/download. 
  18. Char, Danton S., Nigam H. Shah, and David Magnus. 2018. “Implementing Machine Learning in Health Care—Addressing Ethical Challenges.” New England Journal of Medicine 378 (11): 981–983. https://doi.org/10.1056/NEJMp1714229. 
  19. Nong, Paige, et al. 2025. “Current Use and Evaluation of Artificial Intelligence and Predictive Models in US Hospitals.” Health Affairs. https://doi.org/10.1377/hlthaff.2024.00842. 
  20. Ibid.
  21. Weissman, “FDA Regulation of Predictive Clinical Decision-Support Tools.”
  22. Chang et al., Hospital Trends in the Use, Evaluation, and Governance of Predictive AI.

Parmis Mokhtari-Dizaji

Parmis Mokhtari-Dizaji is a sophomore at Cornell University, pursuing a B.S. in Public Policy at the Jeb E. Brooks School of Public Policy, with minors in Applied Economics, Law & Society, and Design Innovation & Strategy. Her interests center on migration and resettlement policy, sustainability, and human-centered design, as well as technological innovation for social and environmental good. At Cornell, she serves as a project manager and analyst for Social Enterprise at Cornell, where she leads consulting projects that advance social impact through business and emerging technologies. She also serves as the research and program coordinator for the Cornell Law-Med Project, contributing to asylum research that bridges medical and legal evidence, and works with the Global Design Initiative for Refugee Children to develop participatory design solutions for displaced youth. Additionally, Parmis designs with the Medium Design Collective Incubates Subteam and writes for the Cornell Policy Review on humanitarian governance, innovation, and digital regulation.
View all posts

Related Posts

Scroll to Top