Machine learning application in the prediction of postoperative delirium among elderly patients: a systematic review...

This meta-analysis systematically evaluates machine learning (ML) applications for predicting postoperative delirium (POD) among elderly patients. This study compares predictive performance across models, assesses generalizability and evaluates bias risk. The goal is to provide clinicians with evidence-based guidance for selecting and implementing relevant prediction tools.

Following the PRISMA guidelines we searched PubMed Web of Science and the CochraneLibrary for studies published between July 17 2020, and July 17, 2025. We included studies that developed ML models to predict POD in elderly surgical patients. Model performance metrics were extracted. A random-effects model was used for meta-analysis. Subgroup analyses and risk of bias assessments were performed using PROBAST.

A total of 14 studies comprising 49 prediction models were included, collectively enrolling 874,485 patients with a pooled POD incidence of 9.86%. The pooled Area Under the Receiver Operating Characteristic Curve (AUC) for ML models predicting POD was 0.756 (95% CI: 0.740-0.772), indicating moderate predictive performance. Substantial heterogeneity was observed across studies (I² = 95.7%). Among individual algorithms, Logistic Regression (LR) (AUC = 0.794), XGBoost (AUC = 0.773), and LightGBM (AUC = 0.772) showed superior discriminative performance, whereas Decision Trees (DT) performed poorly (AUC = 0.641). In validation analyses, models showed better performance in internal validation (AUC = 0.760) compared with external validation (AUC = 0.745). In external validation, Support Vector Machines (SVM) achieved excellent performance (AUC = 0.808), while XGBoost maintained stable performance (AUC = 0.755). A significant negative correlation was observed between sample size and model performance in external validation (ρ = -0.613, p = 0.015).

ML models demonstrate moderate predictive performance for POD among elderly patients. LR and XGBoost emerged as the top-performing algorithms, whereas DT exhibited relatively poor discriminative ability. Considerable heterogeneity was observed across studies. Notably, external validation revealed a significant negative correlation between sample size and model performance. Future research should enhance methodological quality, conduct rigorous external validation to confirm model generalizability, and prioritize clinical interpretability to promote clinical use.