Predicting long-term outcomes in intracerebral hemorrhage: a comparative study of machine learning models highlights the prognostic value of hematoma clearance

Abstract

Purpose This study aimed to develop and internally validate postoperative prognostic models for long-term functional outcomes in patients with spontaneous intracerebral hemorrhage (sICH) following hematoma evacuation, and to compare the performance of traditional logistic regression (LR) with machine learning (ML) algorithms in a limited sample size setting. Methods A retrospective cohort of 185 ICH patients who underwent stereotactic surgery was analyzed. Four key predictors—age, admission Glasgow Coma Scale (GCS) score, ICH grade, and hematoma clearance ratio—were selected using Least Absolute Shrinkage and Selection Operator (LASSO) regression. Predictive models were constructed using LR and multiple ML algorithms (Gaussian Naïve Bayes (GNB), Linear Discriminant Analysis (LDA), Random Forest (RF), eXtreme Gradient Boosting (XGBoost)). Model performance was evaluated using area under the curve (AUC), accuracy, sensitivity, specificity, and other metrics. Results Hematoma clearance ratio was associated with 6- and 12-month outcomes. The Youden index suggested a threshold of 67.3%, whereas RCS analysis indicated an exploratory cut point near 81%; therefore, the 81% value should be considered hypothesis-generating. LR demonstrated internally validated performance comparable to the evaluated ML models in small-sample settings. Older age, higher ICH grade, and lower GCS were associated with a lower probability of favorable outcomes. Conclusion In this single-center retrospective cohort, the hematoma clearance ratio was associated with long-term functional recovery after ICH surgery. LR provided performance comparable to the evaluated ML models and may serve as an interpretable research model.