Development of an explainable machine learning model for predicting neurological deterioration in spontaneous intracerebral hemorrhage

Abstract

Background

Intracerebral hemorrhage (ICH) is a severe form of stroke associated with high morbidity and mortality. Early prediction of neurological deterioration (ND)—defined as a decline of at least 2 points on the Glasgow Coma Scale (GCS) within 48 h of admission or mortality at discharge—is essential for timely intervention and improved outcomes.

Methods

We developed an explainable machine learning model to predict ND using clinical, laboratory, and radiological data extracted from electronic medical records (EMR) of a retrospective cohort of 491 ICH patients, with ND observed in 52.3 % of cases. Multiple machine learning algorithms—including random forests, extra trees, and CatBoost—were trained, and model performance was evaluated using metrics such as the area under the receiver operating characteristic curve (AUC-ROC) and F1-score. Shapley Additive Explanations (SHAP) were employed to enhance interpretability.

Results

The final model, a blended ensemble, achieved an AUC-ROC of 0.8743, an F1-score of 0.8077, and a sensitivity of 0.8182 on the test set. Key predictors included initial GCS, hematoma volume, age, and the presence of intraventricular hemorrhage. SHAP analysis provided insights into the relative contributions of these predictors, reinforcing the model's clinical relevance.

Conclusions

Our model demonstrates promising predictive performance, suggesting its potential utility for early risk stratification and guiding interventions in ICH management. Further validation in diverse clinical settings is warranted.