AI Models for Surgical Decision Support in Spontaneous Intracerebral Hemorrhage: A Systematic Review in Relation to Trials and Guidelines.
Authors
Affiliations (2)
Affiliations (2)
- Division of Neurosurgery, University of São Paulo Medical School, Dr. Enéas de Carvalho Aguiar Ave., 255, Room 5083, São Paulo, 05403-000, Brazil. [email protected].
- Division of Neurosurgery, University of São Paulo Medical School, Dr. Enéas de Carvalho Aguiar Ave., 255, Room 5083, São Paulo, 05403-000, Brazil.
Abstract
Artificial intelligence (AI) applications for spontaneous intracerebral hemorrhage (ICH) are rapidly expanding, particularly in perioperative imaging analysis and surgical decision support. Because most predictive AI models are developed using independent clinical datasets, they are not expected to explicitly reproduce randomized trial protocols or guideline decision rules. We therefore conducted a descriptive systematic review and evidence-mapping study to crosswalk AI model inputs, predicted targets, and outputs to major surgical trials and clinical guidelines (ENRICH, MIND, MISTIE III, SWITCH, STICH II, CLEAR, and AHA/ASA), identifying areas of overlap where AI operationalizes trial-relevant constructs and areas where AI-derived predictors may be hypothesis-generating for future trial design. Of 37 records identified (31 from database searches and 6 from hand searches), 21 studies met eligibility criteria. Publications increased after 2021, peaking in 2024 (n = 5) and remaining in 2025 (n = 3). Most cohorts were single-center (13/21), mainly from China, the USA, and Germany. Inputs were predominantly non-contrast computed tomography (CT); one study used magnetic resonance imaging (MRI) for trajectory planning. Deep learning was the most common analysis method (14/21), followed by classical machine learning (4/21) and radiomics-based methods (2/21). AI applications focused on perioperative imaging tasks, including eligibility assessment, postoperative quality assurance, trajectory planning, workflow optimization, and treatment-effect modeling. Three studies demonstrated direct alignment with surgical trial thresholds (e.g., MISTIE/CLEAR), ten were indirectly aligned, and eight had no clear linkage. AI models add value to imaging-based perioperative assessment in ICH and frequently target constructs central to trial- and guideline-based decision-making, even when trial criteria are not explicitly encoded. This crosswalk clarifies where AI outputs overlap with trial- and guideline-relevant constructs and where AI-derived decision boundaries diverge, reinforcing measurement targets that can be benchmarked against existing evidence and identifying candidates for hypothesis-generating evaluation in future studies.