Interpretable Machine Learning Model Integrating Electrocardiographic and Acute Physiology Metrics for Mortality Prediction in Critical Ill Patients

Background: Critically ill patients in the intensive care unit (ICU) are characterized by complex comorbidities and a high risk of short-term mortality. Traditional severity scoring systems rely on physiological and laboratory variables but lack direct integration of electrocardiogram (ECG) data. This study aimed to construct an interpretable machine learning (ML) model combining ECG-derived and clinical variables to predict 28-day mortality in ICU patients. Methods: A retrospective cohort analysis was performed with data from the MIMIC-IV v2.2 database. The primary outcome was 28-day mortality. An ECG-based risk score was generated from the first ECG after ICU admission using a deep residual convolutional neural network. Feature selection was guided by XGBoost importance ranking, SHapley Additive exPlanations, and clinical relevance. A three-variable model comprising ECG score, APS-III score, and age (termed the E3A score) was developed and evaluated across four ML algorithms. We evaluated model performance by calculating the AUC of ROC curves, examining calibration, and applying decision curve analysis. Results: A total of 18,256 ICU patients were included, with 2412 deaths within 28 days. The ECG score was significantly higher in non-survivors than in survivors (median [IQR]: 24.4 [15.6–33.4] vs. 13.5 [7.2–22.1], p < 0.001). Logistic regression demonstrated the best discrimination for the E3A score, achieving an AUC of 0.806 (95% CI: 0.784–0.826) for the test set and 0.804 (95% CI: 0.772–0.835) for the validation set. Conclusions: Integrating ECG-derived features with clinical variables improves prognostic accuracy for 28-day mortality prediction in ICU patients, supporting early risk stratification in critical care.