PhD Scientific Days 2025

Pharmaceutical Sciences and Health Technologies III.

Investigating the Accuracy of Machine Learning Models in Predicting Extubation Success in Mechanically Ventilated Patients - A Systematic Review and Meta-Analysis

Name of the presenter

Bakos Péter

Institute/workplace of the presenter

Csolnoky Ferenc Kórház, Veszprém

Authors

Péter Bakos¹, Shir Galin¹, Dávid Laczkó¹, Caner Turan¹, Bence Szabó¹, Péter Hegyi¹, András Lovas¹, Zsolt Molnár¹

1: Semmelweis University, Centre for Translational Medicine

Text of the abstract

Introduction

Optimal timing of extubation in mechanically ventilated patients remains a challenge in intensive care. Both failed and delayed extubation contribute to increased morbidity and mortality. Currently no gold standard exists for predicting succesful extubation. Machine learning (ML) models show promise in outcome prediction, but their application to extubation success lacks standardized validation.

Aims

To systematically review and quantitatively evaluate the performance of ML models in predicting extubation success in critically ill, mechanically ventilated patients.

Methods

We searched PubMed, Embase, and the Cochrane Library for studies involving adult ICU patients undergoing planned extubation, where ML models were used to predict extubation success or failure. Clinical scoring systems were also included for comparison. Performance metrics were meta-analyzed and subgroup analysis was conducted. The review followed Cochrane Handbook methodology. Risk of bias was assessed using modified versions of QUADAS-2 and QUADAS-C, adapted for ML studies.

Results

Twenty-six studies were eligible for systematic review, 43 ML models from 13 studies were included in the meta-analysis. The most commonly reported metric was the area under the ROC curve (AUC). The rapid shallow breathing index (RSBI) was the most frequently reported clinical score.
Model performance varied, with the best AUCs ranging from 0.66 to 0.97. Pooled AUCs by model type were 0.88 (95% CI: 0.77–0.94) for traditional ML models and 0.85 (95% CI: 0.66–0.94) for deep learning models. Tree-based models showed the highest performance (AUC: 0.92, 95% CI: 0.78–0.97), while linear models performed the worst (AUC: 0.66, 95% CI: 0.23–0.93). RSBI alone demonstrated poor predictive value (AUC: 0.58, 95% CI: 0.17–0.90). Study heterogeneity was high, driven by differences in predictor selection and model design.

Conclusion

ML models show good accuracy in predicting extubation success and may support clinical decision-making. Further research is needed to develop generalizable tools suitable for real-world implementation.

Funding

EKÖP-KDP 2024-2028