An explainable machine-learning model using routinely available early ICU data showed good discrimination for in-hospital mortality risk among patients with heart failure (HF) in a retrospective cohort analysis published in BMC Cardiovascular Disorders.
The study used data from the MIMIC-IV v3.1 database and included adult ICU patients with HF identified using ICD-10 codes (I50.x) between October 2015 and 2022. Predictors included demographics, comorbidities, pre-ICU cardiovascular medications, and vital signs and laboratory parameters collected within the first 0-6 hours of ICU admission. Implausible values were excluded, and missing data were imputed using medians or modes.
An XGBoost classifier was trained using an 80/20 stratified split with class weighting. Model performance was evaluated using area under the curve (AUC), average precision, sensitivity, specificity, F1-score, and Brier score. SHapley Additive exPlanations (SHAP) analysis was used to assess model interpretability.
A total of 12,110 ICU patients with HF were included, of whom 2,041 (16.9%) died during hospitalization. In the test cohort of 2,422 patients, the model achieved an AUC of 0.797 (95% confidence interval [CI], 0.772-0.822) and an average precision of 0.491. Sensitivity and specificity were 0.632 and 0.793, respectively, at the default threshold, while use of the Youden-optimal threshold increased sensitivity to 0.708. Kaplan-Meier analysis showed significant survival separation across predicted risk quartiles (log-rank P<0.001).
SHAP analysis identified FiO₂, age, lactate, Charlson comorbidity index, blood urea nitrogen (BUN), and systolic blood pressure as key mortality predictors. The findings suggest that explainable ML models using early ICU data may support structured mortality risk stratification in patients with HF while assisting, rather than replacing, clinical judgment.
