Title:Predicting Extubation Failure in Intensive Care: The Development of a Novel, End-to-End Actionable and Interpretable Prediction System

Abstract:Predicting extubation failure in intensive care is challenging due to complex data and the severe consequences of inaccurate predictions. Machine learning shows promise in improving clinical decision-making but often fails to account for temporal patient trajectories and model interpretability, highlighting the need for innovative solutions.
This study aimed to develop an actionable, interpretable prediction system for extubation failure using temporal modelling approaches such as Long Short-Term Memory (LSTM) and Temporal Convolutional Networks (TCN). A retrospective cohort study of 4,701 mechanically ventilated patients from the MIMIC-IV database was conducted. Data from the 6 hours before extubation, including static and dynamic features, were processed through novel techniques addressing data inconsistency and synthetic data challenges. Feature selection was guided by clinical relevance and literature benchmarks.
Iterative experimentation involved training LSTM, TCN, and LightGBM models. Initial results showed a strong bias toward predicting extubation success, despite advanced hyperparameter tuning and static data inclusion. Data was stratified by sampling frequency to reduce synthetic data impacts, leading to a fused decision system with improved performance. However, all architectures yielded modest predictive power (AUC-ROC ~0.6; F1 <0.5) with no clear advantage in incorporating static data or additional features. Ablation analysis indicated minimal impact of individual features on model performance.
This thesis highlights the challenges of synthetic data in extubation failure prediction and introduces strategies to mitigate bias, including clinician-informed preprocessing and novel feature subsetting. While performance was limited, the study provides a foundation for future work, emphasising the need for reliable, interpretable models to optimise ICU outcomes.