Predicting time-to-event outcomes in critically ill patients with intracerebral hemorrhage using machine learning

Background Intracerebral hemorrhage is associated with high mortality in intensive care settings. Current prognostic models have limitations, including poor interpretability and insufficient validation in critically ill populations. In this study, we developed an interpretable machine learning model to predict survival in intensive care unit patients with intracerebral hemorrhage. Methods This retrospective study used data from patients with intracerebral hemorrhage in the eICU Collaborative Research Database for model development and the Medical Information Mart for Intensive Care IV database for external validation. Clinical, laboratory, and physiological parameters within 24 h of intensive care unit admission were extracted. Six machine learning survival algorithms, including the Random Survival Forest, were implemented. Model performance was assessed using the time-dependent area under the curve, concordance index, and Brier score. Model interpretability was achieved through the SHapley Additive exPlanations framework. Results The cohort comprised 5797 patients from eICU and 1423 patients from Medical Information Mart for Intensive Care IV. Random Survival Forest demonstrated superior performance, with a time-dependent area under the curve of 0.88 in internal validation and 0.82 on day 1 of external validation. SHapley Additive exPlanations analysis identified Glasgow Coma Scale score, Acute Physiology Score, age, creatinine, temperature, and systolic blood pressure as key predictors. Critical risk thresholds were Glasgow Coma Scale score <9.8, Acute Physiology Score >52.7, and age >62.9 years. Conclusion We developed a machine learning survival prediction model that demonstrated robust performance and clinical utility. The web-based tool may enhance intensive care unit risk stratification and clinical decision-making.