This study evaluates the performance of machine learning (ML) algorithms in predicting hospital outcomes: readmission, length of stay (LOS), and mortality.

A systematic review was registered on the PROSPERO database (CRD42024505704) and conducted according to PRISMA guidelines. A detailed search was conducted across six databases to identify studies applying ML algorithms to forecast readmission, LOS, or mortality in general, heterogeneous hospital populations. Data collection extracted model types, input variables, and reported performance metrics. Risk of bias was assessed using the PROBAST tool.

Fifty-two studies met the eligibility criteria. Random Forest, XGBoost, and Support Vector Machines demonstrated consistent discriminative capacity across independent cohorts. Reported area under the receiver operating characteristic curve (AUC) values reached up to 0.90 for mortality (Random Forest), >0.75 for LOS (XGBoost), and 0.79 for readmission (Random Forest). AUC was the most frequently reported metric (60–85%), whereas sensitivity, specificity, and calibration measures (e.g., Brier score) appeared infrequently. Age, comorbidity burden, and prior admissions were the predominant predictors. Socioeconomic variables were underrepresented.

Despite high discriminative performance, systemic methodological deficits prevent clinical adoption. The limited reporting of calibration measures restricts the assessment of the reliability of predicted risks, and the underrepresentation of socioeconomic determinants may compromise algorithmic equity. Future clinical implementation mandates standardized metric reporting, prospective validation, and the strict integration of social determinants of health.