Adjusted-GCS-Enhanced Machine-Learning Model Predicts 28-Day Mortality in ICU Stroke Patients

Background : Stroke is a leading cause of mortality and long-term disability worldwide. Accurate prediction of outcomes in critically ill stroke patients, particularly in intensive care unit (ICU) settings, is crucial for timely interventions. The Glasgow Coma Scale (GCS) is commonly used to assess neurological function while traditional GCS scoring is inadequate for mechanically ventilated patients, particularly in the MIMIC-IV (Medical Information Mart for Intensive Care IV) database. This study aims to enhance stroke mortality predictions by introducing an adjusted GCS scoring method and integrating machine learning (ML) models with Shapley Additive Explanations (SHAP) for improved interpretability. Methods : We conducted a retrospective analysis using data from the MIMIC-IV database, including 12,625 ICU-admitted stroke patients aged 18 years or older. We developed and evaluated ten ML models incorporating clinical features such as age, comorbidities, and GCS scores (both original and adjusted). The adjusted GCS was calculated by modifying the verbal component for ventilated patients. Model performance was evaluated using area under the receiver operating characteristic curve (AUC), accuracy, sensitivity, specificity, and F1 score. SHAP analysis was applied to interpret the contribution of individual features to the model predictions. Results : The study cohort included 10,502 survivors and 2,123 non-survivors. Models incorporating the adjusted GCS consistently outperformed those using the original GCS, with the Gradient Boosting Machine (GBM) achieving the highest AUC of 0.868 (95% CI: 0.853−0.882). Adjusted GCS also showed superior calibration and a closer alignment with observed mortality outcomes. SHAP analysis identified key predictors of mortality, including the adjusted GCS, Charlson Comorbidity Index, age, Simplified Acute Physiology Score II (SAPS II), race, renal disease, and others. Conclusions : The adjusted GCS method, when integrated into ML models, significantly improves the prediction of 28-day mortality in critically ill stroke patients. This approach enhances model interpretability and clinical utility, offering a more accurate tool for personalized risk assessment.