Explainable machine learning versus logistic regression for outcome prediction in primary intracerebral hemorrhage: A multicenter radiomics study

Abstract: CONTEXT: Accurate outcome prediction is essential for clinical decisions in intracerebral hemorrhage (ICH) patients. However, whether machine learning (ML) models outperform traditional logistic regression (LR) remains unclear. AIMS: This study aims to compare six ML algorithms with LR in predicting poor 3-month outcomes after primary ICH, using radiomics features from noncontrast computed tomography and clinical data. SETTINGS AND DESIGN: A retrospective study. SUBJECTS AND METHODS: Seven hundred and four primary ICH patients from two centers were allocated into training ( n = 516), internal ( n = 128), and external validation ( n = 60) cohorts. Radiomics features from hematoma regions were extracted to generate a radiomics score (Rad-score). STATISTICAL ANALYSIS USED: The Rad-score and clinical variables were selected for developing one LR and six ML models: random forest (RF), artificial neural network (ANN), AdaBoostM1, Naive Bayes (NB), XGB, and support vector machine (SVM). Model discrimination was assessed by the area under the curve (AUC), and the best-performing ML model was interpreted using Shapley Additive exPlanations (SHAP). RESULTS: In the training cohort, AUCs were 0.849 for LR, 0.897 for RF, 0.885 for XGB, 0.884 for AdaBoostM1, 0.858 for ANN, 0.848 for NB, and 0.839 for SVM. In the internal and external validation cohorts, AUCs ranged from 0.796–0.823 and 0.806–0.858, respectively. The RF model achieved significantly higher AUCs than LR in both training and external validation sets (both P < 0.05). SHAP plots identified Rad-score and National Institutes of Health Stroke Scale as key predictors. CONCLUSIONS: The RF model, integrating radiomic and clinical data, outperformed LR and showed the highest accuracy in predicting poor 3-month outcomes after primary ICH.