Machine learning models for mortality prediction in patients with spontaneous subarachnoid hemorrhage following ICU treatment

Background Spontaneous subarachnoid hemorrhage (SAH) is a severe and potentially life-threatening acute cerebrovascular disease. Early identification of the risk of death in patients with spontaneous SAH is of vital importance for improving prognosis, reducing mortality, and guiding clinical treatment. Methods A retrospective cohort study was conducted using the public database, Medical Information Mart for Intensive Care IV (MIMIC)-IV. The primary outcome was in-hospital mortality following intensive care unit (ICU) treatment. All features were extracted from first-day ICU admission data. Data analysis was performed by using R and Python, with feature selection conducted via least absolute shrinkage and selection operator (LASSO) regression. We constructed 8 models based on the 12 selected features in the training set and evaluated them in the test set by various metrics, including area under the curve (AUC), accuracy, precision (positive prediction value), recall (sensitivity), Brier score, Jordan index, and calibration slope. The most effective model was rendered explainable through the SHapley Additive exPlanations (SHAP) approach. Results The study included 1,121 records, with 870 surviving and 251 deceased patients. We selected 43 features for the preliminary baseline analysis. Based on LASSO regression analysis and clinical practical significance, 12 features were finally included in the construction of the machine learning models. We constructed eight machine learning models, among which the logistic regression (LR) model performed the best. Conclusions In our study, the LR model exhibited superior discrimination in predicting risk of mortality among patients with spontaneous SAH compared to other models. This research contributes to facilitating the early identification of mortality risk in patients with spontaneous SAH. External validation and further prospective studies are warranted to confirm and refine these predictive insights for clinical utilization.