Machine Learning-Based Prediction of Poor Outcomes in Intracerebral Hemorrhage: A Systematic Review and Meta-Analysis

Area of Science:

• Neurology
• Medical Informatics
• Biostatistics

Background:

• Spontaneous intracerebral hemorrhage (ICH) carries high mortality and disability risks.
• Accurate early prediction of ICH outcomes remains a significant clinical challenge.
• Current prediction methods require enhancement for improved patient management.

Purpose of the Study:

• To systematically evaluate machine learning (ML) model performance in predicting adverse outcomes in ICH.
• To assess the predictive accuracy for hematoma expansion (HE), poor functional outcome, and mortality.
• To provide evidence for future research and clinical translation of ML in ICH.

Main Methods:

• Systematic literature search across major databases (PubMed, Embase, Web of Science, Cochrane Library) up to September 2025.
• Inclusion of studies developing and validating ML models for ICH outcome prediction.
• Meta-analysis of pooled concordance index (C-index), sensitivity, and specificity using random-effects or bivariate models.

Main Results:

• Eighty-three studies involving over 136,840 patients were analyzed.
• Integrated clinical-radiomics ML models demonstrated highest discriminative performance: HE (C-index 0.822), poor outcome (C-index 0.850), mortality (C-index 0.860).
• Logistic regression models performed comparably to more complex ML algorithms; external validation data is limited.

Conclusions:

• Machine learning models, especially integrated clinical-radiomics approaches, show robust predictive capabilities for ICH outcomes.
• These models can significantly enhance risk stratification and personalize patient management strategies.
• Further validation in diverse patient cohorts is crucial for widespread clinical adoption.