Machine learning models predicting extubation success in mechanically ventilated patients: a systematic review and...

Optimal timing of extubation in mechanically ventilated patients remains a major challenge in intensive care. Machine learning (ML) models have been increasingly proposed to support clinical decision-making, yet their predictive performance and readiness for clinical application in extubation outcomes remain uncertain. This study aimed to evaluate the predictive performance and clinical readiness of ML models for predicting extubation success.

A systematic search was performed in PubMed, Embase and CENTRAL up to November 5, 2024. Studies including mechanically ventilated critically ill adult patients undergoing planned extubation were eligible. The index test was any ML model predicting extubation outcome. Models reporting an area under the receiver operating characteristic curve (AUC) were meta-analyzed, and subgroup analyses were conducted. Risk of bias was assessed using a modified version of the Quality Assessment of Diagnostic Accuracy Studies-Comparative (QUADAS-C) tool, adapted for ML-based prediction studies.

Twenty-six studies were included in the systematic review, and 47 ML models from 14 studies (n = 34,322 patients) were eligible for meta-analysis. Reported AUC values across predictive models ranged from 0.59 to 0.98. Pooled AUCs by model type were 0.88 (95% CI: 0.78-0.94) for classical ML models and 0.85 (95% CI: 0.68-0.94) for deep learning models. The best-performing ML models of each study had a pooled AUC of 0.90 (95% CI: 0.82-0.95). Pooled estimates were derived from internally validated models, as only two studies reported externally validated AUCs with confidence intervals. Study heterogeneity was high, driven by substantial differences in predictor selection and model design.

Machine learning models demonstrate acceptable discriminatory performance for predicting extubation success. However, limited external and prospective validation, substantial heterogeneity, and inconsistent reporting currently preclude their routine clinical implementation.