Global performance of machine learning models to predict all-cause mortality: systematic review and meta-analysis

Abstract

We aimed to review the literature on the performance of machine learning models to predict all-cause mortality. The systematic review was protocolled in PROSPERO (CRD42023476567) following PRISMA guidelines. Searches were conducted in PubMed, LILACS, Web of Science, and Scopus databases. Studies predicting all-cause mortality using machine learning were analyzed with random-effects models, with heterogeneity assessed using I2 statistics and quality evaluated using TRIPOD + AI. The meta-analysis included 88 studies. Most of the studies were from the United States (n = 25) and China (n = 20). Overall pooled AUC was 0.831 (95% CI 0.797–0.865), with extreme heterogeneity (I2:100%). The majority of the studies included no social variables in the models (89.8%). Subgroup analysis showed similar performance between general population studies and disease-specific populations. Models from high-income countries were similar to those from low- and middle-income countries. Meta-regression showed covariates that affected the results: algorithm, population type, study quality score, and study CI imputation. Equity-oriented sub-group analysis (< 10%) and external validation in other datasets (8.0%) were scarce. Overall, machine learning models showed high performance to predict all-cause mortality, but also highlighted equity gaps. The limitations reduce the potential of public health’s evaluation and deployment due to the risk of perpetuation of social disparities. Extreme heterogeneity indicates highly context-dependent performance requiring local validation before implementation assessment.