Introduction: Premature or delayed extubation is associated with increased ICU complications, including reintubation, ventilator-associated pneumonia (VAP), and mortality. Current tools are limited by static scoring systems, inconsistent nursing assessment integration, and lack of real-time feedback. This study aimed to develop and validate a nurse-accessible ensemble learning model for dynamic extubation risk prediction.
Methods: We used time-series data from the MIMIC-IV database (n = 17,016) and external validation from a Chinese ICU cohort (n = 158). Patients who underwent planned extubation were included. A Grid Search-based Soft Voting (GSSV) ensemble model was developed from 16 nurse-observable and routinely charted variables (vital signs, GCS components, ventilator settings, etc.) within a 24-hour window prior to extubation. Model performance was compared across RF, XGBoost, CatBoost, LSTM, and GSSV using 5-fold cross-validation and AUC metrics.
Results: GSSV outperformed all five baseline models in both internal and external validation. In MIMIC-IV test set (n=2615), GSSV achieved an AUC of 0.8601 (95% CI: 0.8449–0.8753), outperforming LSTM (0.7783), RF (0.8542), XGBoost (0.8529), CatBoost-1 (0.8556), and CatBoost-2 (0.8587). In the external validation cohort, GSSV achieved an AUC of 0.8808 (95% CI: 0.7986–0.9631), with sensitivity of 65% at 87.5% specificity. Patients extubated against GSSV “not-ready” predictions had significantly lower success rates (39.01% vs. 92.40%) and higher 48-hour mortality (90.31% vs. 23.49%) compared to GSSV-supported extubations. Delayed extubation despite GSSV readiness was associated with increased rates of atelectasis (5.58% vs. 0%) and VAP (12.15% vs. 1.03%). SHAP analysis revealed that GCS motor and eye-opening scores were the most influential features across models.
Conclusions: Our GSSV ensemble model provides accurate real-time extubation risk prediction based on nursing-accessible inputs and offers actionable insights for bedside decision-making. It enhances extubation safety and holds promise for integration into ICU quality improvement workflows.
