Machine learning models for predicting delayed cerebral ischemia following ruptured intracranial aneurysms: A systematic review and meta-analysis.

BACKGROUND

Delayed cerebral ischemia (DCI) remains a major morbidity and mortality problem following aneurysmal subarachnoid hemorrhage (SAH). Advancements in neurocritical care permit a slow but accurate identification of patients at high risk for DCI. Machine learning models are now emerging as tools for DCI prediction that may provide more individualized risk assessment than conventional approaches.

METHODS

We systematically searched PubMed and Embase for studies developing or validating ML models for predicting DCI after SAH. Qualitative assessment of study quality was performed with QUADAS-2 tool. The performance metrics were synthesized and compared among algorithm families and dataset types (training, test & validation).

RESULTS

Among these 29 studies, 10,000 patients and more than 100 ML models were analyzed. The best discriminative performance was obtained by ensemble methods (Random Forest, XGBoost) with median AUCs of 0.80-0.85, and logistic regression was still the most common and interpretable model. Deep learning models performed variable and had greater overfitting. Sensitivity and specificity varied across models and cohorts, with ensemble models balancing both metrics best. External validation was, however, scarce.

CONCLUSIONS

ML models and particularly ensemble approaches promise to improve DCI prediction after SAH. External validation, model calibration and prospective clinical integration warrant future work.