Comparison of the predictive performance of machine learning and conventional logistic regression models for poor discharge outcomes in patients with Aneurysmal subarachnoid hemorrhage: A retrospective cohort study.

Aneurysmal subarachnoid hemorrhage (aSAH) is a life-threatening disease with high morbidity and mortality. Although numerous models have been developed to predict poor discharge outcomes in patients with aSAH, few studies have systematically compared the performance of machine learning (ML) with traditional logistic regression (LR). This retrospective cohort study included data from 1,414 patients with aSAH who underwent endovascular treatment between April 2021 and April 2023 at five neurointerventional centers in China. For ML model development, candidate predictors were preselected using least absolute shrinkage and selection operator (LASSO), whereas predictors for the LR model were identified using univariable and multivariable analyses. Six ML algorithms were trained, and the best-performing ML model was compared with the LR model to evaluate their predictive performance in forecasting poor discharge outcomes. Among the six machine learning algorithms evaluated, XGBoost showed the best predictive performance and was therefore selected as the representative model for primary comparison with LR. Compared with XGBoost, the LR model demonstrated more consistent performance across datasets, with areas under the receiver operating characteristic curve (AUC) of 0.902 and 0.856 in the training and external validation cohorts, respectively. In addition, the LR model exhibited better calibration (Brier score: 0.101 vs. 0.115, p < 0.05), and superior net reclassification improvement (NRI: 0.240, p < 0.05). Compared with ML models, the LR model remains a practical and reliable risk prediction tool for large-scale sample modeling in clinical practice, enabling timely in-hospital risk stratification and informing individualized management.