Multimodal machine learning for predicting postoperative functional outcomes in surgically treated supratentorial deep intracerebral hemorrhage: a prospective multicenter study

TL;DR

This study developed a machine learning model combining clinical and biological data to predict recovery outcomes after brain hemorrhage surgery, showing strong performance and interpretability.

Contribution

A novel multimodal machine learning model using clinical, imaging, physiological, and biomarker data for predicting postoperative outcomes in sICH patients.

Findings

A Random Forest model achieved an AUC of 0.883 in predicting functional outcomes after surgery for intracerebral hemorrhage.

Admission GCS and hematoma volume were identified as the most important predictors by SHAP analysis.

The model included eight key predictors selected via LASSO, including biomarkers like TNF-α and GFAP.

Abstract

Early prediction of functional outcomes after surgery for spontaneous supratentorial deep intracerebral hemorrhage (sICH) remains difficult. This study developed and validated multimodal machine-learning models incorporating clinical, imaging, physiological, and biomarker data, including temperature management strategies, and explored interpretability using SHAP.

This prospective multicenter cohort enrolled 285 surgically treated sICH patients. Outcome was defined as favorable (mRS 0–3) vs. unfavorable (mRS 4–6). Data were split by stratified random sampling into a training set (n = 199) and a test set (n = 86). LASSO with 10-fold cross-validation (1-SE rule) selected key predictors. Five classifiers (Random Forest, neural network, decision tree, k-nearest neighbors, naïve Bayes) were trained with 10-fold cross-validation and evaluated on the test set. Performance was assessed using AUC (95% CI) and standard classification metrics; AUCs were compared by DeLong's test. SHAP was applied to the best model.

LASSO identified eight predictors: admission GCS, hematoma volume, TNF-α, GFAP, IL-1β, admission NIHSS, mean body temperature, and peak ICP. On the test set, Random Forest achieved the highest performance (AUC 0.883, 95% CI 0.829–0.937; accuracy 0.824; F1-score 0.836), with no significant AUC difference versus the neural network (AUC 0.867; P = 0.312). SHAP ranked admission GCS and hematoma volume as the most important features, followed by TNF-α and GFAP.

A multimodal Random Forest model provided good discrimination for predicting postoperative functional outcomes in surgically treated sICH, and SHAP improved interpretability by quantifying feature contributions.

Figures