Machine learning for predicting 365-day mortality in spontaneous intracerebral hemorrhage patients in the intensive care unit

Abstract

Abstract SICH is associated with high rates of disability and mortality. This study aims to develop a predictive model to assess the 365-day mortality rate of patients with SICH. We conducted a retrospective study and identified patients with SICH from the MIMIC-IV database (n = 2069). Baseline variables were screened to identify potential predictors using three methods: univariate analysis, least absolute shrinkage and LASSO regression, and the Boruta algorithm. Among seven well-established models, the optimal model was selected based on the area under the curve (AUC) in the validation cohort. Model calibration and decision curve analysis were further performed to evaluate predictive performance, and Shapley Additive Explanations (SHAP) were used to visualize the contribution of each variable to the model. The Random Forest model demonstrated the most efficient and robust predictive performance. For predicting 365-day mortality in all patients, the Random Forest model showed excellent discriminative ability in the validation cohort, with an AUC of 0.83 (95% CI 0.796–0.867), a specificity of 90.86%, and an F1 score of 63.41%. Calibration and decision curve analyses indicated no significant bias and suggested potential clinical utility. Overall, the Random Forest model demonstrated moderate-to-good predictive performance for estimating 365-day mortality risk in patients with SICH and may provide supplementary support for early risk stratification, highlighting its potential value in the development of early warning systems. A total of 2069 patients with SICH were included in this study, among whom 717 experienced poor outcomes. Among the machine learning models evaluated in the training cohort, theRF model achieved the best overall performance, with a mean AUC of 0.83 (95% CI 0.796–0.867). Regarding feature importance for predicting long-term outcomes in patients with SICH, the GCS score ranked highest, followed by length of hospital stay (LOS_hospital), age, SAPS II, OASIS, and other variables. Ultimately, guided by the variable importance weights of the RF model and insights from the model’s ROC curve, 14 variables were integrated to establish a tailored long-term prognostic prediction platform for patients with SICH. Based on the results of the Random Forest model, we developed a predictive model incorporating 14 clinically accessible predictors, which demonstrated reliable performance in predicting long-term outcomes in patients with SICH.