Deep learning models for ICU readmission prediction: a systematic review and meta-analysis.

backgroundIntensive Care Unit (ICU) readmissions are associated with increased morbidity, mortality, and healthcare costs. Therefore, determining an appropriate timing of ICU discharge is critical. In this context, deep learning (DL) approaches have attracted significant attention.

methodsWe conducted a systematic review of studies developing or validating DL models for ICU readmission prediction, published up to March 4th, 2025, and indexed in PubMed, Embase, Scopus, and Web of Science. We summarised them along multiple dimensions, including outcome and population definition, DL architecture, reproducibility, generalizability, and explainability, and provided a meta-analytic estimate of model performance.

resultsWe included 24 studies encompassing 49 DL models, predominantly trained on US-based datasets, and rarely subjected to external validation. There was considerable variability across study settings, including the definition and timeframe of the ICU readmission outcome, as well as DL architecture used, alongside a substantial risk of bias. Technical reproducibility and model interpretation were rare. A meta-analysis of AUROC values from 11 studies yielded a mean of 0.78 (95% CI = 0.72–0.84), with very high heterogeneity (I2 = 99.9%). Models targeting disease-specific ICU subpopulations achieved significantly higher performance (mean AUROC = 0.92, 95% CI = 0.89–0.95, p = 0.002), and substantially lower heterogeneity (I2 = 17.1%).

conclusionsDL models showed promising performances in predicting ICU readmissions, but exhibited several shortcomings, including low reproducibility, over-reliance on a few US-based datasets, and limited explainability. Additionally, the high heterogeneity and risk of bias limited our ability to assess their pooled performance through meta-analysis. Taken together, our observations suggest that the quality of the evidence regarding the application of DL approaches to ICU readmission prediction is poor, thus hindering their clinical applicability.