Abstract
Effective outcome prediction is crucial in the management of spontaneous intracerebral hemorrhage (SICH). This study developed machine learning models to predict the 180-day Glasgow Outcome Scale (GOS) score using clinical parameters and identify key prognostic factors. All patients with SICH were randomized into training and internal validation cohortsin a 7:3 ratio. After the most relevant variables were selected by the SVM-RFE and LASSO algorithms, the predictive efficiency of the gradient boosting decision tree (GBDT), XGBoost, random forest (RF), and lightgbm models was evaluated through an exhaustive suite of performance indicators. The Shapley additive explanations (SHAP) and local interpretable model-agnostic explanations (LIME) methods were engaged to explain the top-performingmodel. The model was developed for 215 patients. The GBDT model showed stable performance in predictive power in both the internal and external validation cohorts, with the following metrics assessed in the internal test set: AUROC (0.865 ± 0.069); AUPRC (0.834 ± 0.090); accuracy (0.822 ± 0.065), F1 (0.716 ± 0.114); precision (0.787 ± 0.147); recall (0.676 ± 0.138); sensitivity (0.902 ± 0.070) specificity, (0.787 ± 0.147). The GBDT model was assessed in the outer test set with the following metrics: AUROC: 0.785 (0.752, 0.832), AUPRC: 0.693 (0.649, 0.762), accuracy: 0.770 (0.739, 0.807), F1: 0.769 (0.738, 0.806), precision: 0.769 (0.738, 0.806), recall: 0.770 (0.739, 0.807), sensitivity: 0.826 (0.793, 0.863), Specificity: 0.710 (0.662, 0.761). We found that the blood platelet count, serum calcium level, and hemorrhage in the left occipital lobe, right temporal lobe, and right parietal lobe were the 5 most important features for GOS prediction in the GBDT and RF models. The outcomes of SHAP and LIME were consistent with those of previous studies. The GBDT model exhibited the best prediction performance. Moreover, it has the potential to aid clinicians in identifying high-risk patients and guiding clinical decision making.