Assessing performance, calibration, and explainability of machine learning versus traditional models for early outcome prediction after spontaneous intracerebral hemorrhage: a systematic review and meta-analysis protocol.

Abstract

<h4>Background</h4>Early outcome prediction after spontaneous intracerebral hemorrhage (ICH) is critical for patient management and counseling. Although machine learning (ML) models are increasingly applied, their comparative performance and explainability relative to traditional statistical models remain unclear.<h4>Objectives</h4>To systematically compare the predictive performance, calibration, and explainability of ML versus traditional models for early outcomes after ICH.<h4>Methods</h4>Following PRISMA-P guidelines and registered in PROSPERO (CRD420251166996), this systematic review and meta-analysis will include studies developing, validating, or comparing ML and traditional models for predicting early mortality or poor functional outcome (mRS ≥ 3 or GOS ≤ 3) after ICH. Data sources will include PubMed, Embase, Scopus, Web of Science, Cochrane CENTRAL, IEEE Xplore, and major Chinese databases (CNKI, Wanfang, VIP, CBM). Two reviewers will independently screen studies, extract data, and assess risk of bias using the PROBAST + AI tool, which extends and replaces the original PROBAST framework for prediction models incorporating machine learning. Pooled analyses will employ random-effects models; confidence in the body of evidence will be summarized using an adapted approach informed by GRADE principles for prognosis evidence.<h4>Expected results</h4>This review will explore whether ML-based models demonstrate differences in discrimination, calibration, and explainability compared with traditional models.<h4>Conclusions</h4>This review will provide a comprehensive, evidence-based assessment of prognostic modeling for ICH, guiding future model design, validation, and clinical application.<h4>Systematic review registration</h4>PROSPERO CRD420251166996.