Performance of Machine Learning Models in Predicting Hospital Readmissions and Subsequent Mortality of Patients ≥65 Years

Introduction: In an earlier retrospective cohort study of 129,433 patients aged ≥65 years admitted to one or more of four hospitals in Calgary, Alberta, between 2013 and 2021, we identified a large number (139,829) of second through 39th readmissions. Furthermore, we also identified that mortality rates rose with repeated readmissions: within six months after the fifth readmission (12,271), the mortality rate was 60%. This article reports the next step in our research, which is to conduct a literature search to identify which machine learning methods best predict which patients are at the highest risk of readmission and death. Objectives: Our goal was to identify which machine learning methods have both the highest sensitivity in predicting the likelihood of readmission and the highest specificity in predicting a low likelihood of readmission and death. We also wished to identify which machine learning model best identifies readmissions in the subsets of patients who receive the most frequent diagnoses at discharge so that discipline-centred preventive teams can optimise care and achieve the lowest possible readmission rates. Methods: We conducted searches in Medline and Embase on 2 September 2026 with the search terms Patient Readmission/or exp Hospitalization/or rehospitalisation.mp. AND Machine Leaning OR AI AND age 65 years or older. The search was then conducted separately for cardiac, COPD, pneumonia, diabetes, stroke, abdominal surgery and hip and knee surgery. A separate search was conducted for Patient Readmission/or exp Hospitalization/or rehospitalisation.mp. AND Systematic Review OR meta-analysis AND Machine Leaning OR AI and 65 years and older. We pragmatically selected AUC scores ≥ 0.800 to identify those machine learning methods which achieved AUCs with high sensitivity in identifying the cohorts of patients that were most likely to require readmission, were at high risk of mortality, and thus were most in need of additional interventions to prevent readmission. Results: The machine learning models with AUCs ≥ 0.800 were Gradient Boosting, XGBoost, CatBoost, LightGBM, GBDT, Support Vector Machine, Graph Convolutional Network, Polynomial-SVM, Linear-SVM and Sigmoid-SVM. Conclusions: Machine learning models with AUC scores ≥ 0.800 and with high sensitivity could facilitate the identification of patients at the highest risk of readmission and subsequent mortality within six months, allowing prevention teams of family physicians, relevant consultants and visiting nurses to augment care in collaboration with patients and caregivers using interventions tested in RCTs with a low risk of bias to prevent rehospitalisation. However, even patients with lower risk of readmission have risk factors of concern. Next steps in research are: (1) Search the literature to identify studies of readmission prevention at a low risk of bias to enable readmission prevention teams of family physicians, involved consultants and a visiting home nurse to design care plans and increase care that supports patients at home and thus avoids readmissions. (2) Increase the database to 437,000 admissions between 2013 and 2024 to provide more patients on which to test the models.