Machine Learning for Outcome Prediction and Rehabilitation in Hemorrhagic Stroke: Systematic Review
Simranjeet Nagoke1, Aditi Agarwal2, Binay Panjiyar3, Govind Mann4, SAI VENKATA MANOJ KOTHARU5
1Neurology, Government Medical College Jammu, 2Department of Medicine, Department of Medicine, Bharati Vidyapeeth University Medical College, Pune, India., 3Texas Tech University, 4Neurology, Sant Parmanand Hospital, 5Department of Medicine, Osmania medical college, Hyderabad
Objective:
To systematically review 2023–2025 studies on artificial intelligence for prognosis and rehabilitation after intracerebral and aneurysmal subarachnoid hemorrhage, compare performance with conventional scores, and appraise readiness for bedside use.
Background:
Intracerebral hemorrhage and aneurysmal subarachnoid hemorrhage cause high mortality and long-term disability. Early risk stratification and rehabilitation planning drive outcomes and resource use. Artificial intelligence applied to CT, ICU, and EHR data often outperforms conventional scores, yet prospective validation, calibration reporting, equity assessment, and workflow integration remain limited. We reviewed 2023-2025 studies to quantify performance and interpretability and to assess bedside readiness for prognosis and rehabilitation.
Design/Methods:
Systematic review of adult ICH and aSAH studies from 2023–2025. Inclusion: AI for prognosis or rehabilitation with reported inputs, outcomes, and validation. Dual screening and extraction. Outcomes: mortality, mRS, discharge, FIM. We compared with conventional scores and summarized AUROC, calibration, and external validation. Qualitative bias appraisal.
Results:
Ten studies met criteria: six prediction models, one randomized trial, and three reviews. Three models used CT for ICH outcomes, two used ICU or EHR data for aSAH outcomes, and one predicted rehab discharge status. Five of six models beat traditional scores with accuracy (AUROC) between 0.817 and 0.946. Most offered some explanation of how predictions were made. Few checked performance on outside hospitals, and many did not report how well predicted risks matched reality or the net clinical benefit. The randomized trial showed better functional gains with an AI-supported rehab device. Many studies were small and single-center with mixed endpoints.
Conclusions:
AI already predicts short-term outcomes after hemorrhagic stroke better than common bedside scores and can help plan discharge and rehab. Before everyday use, we need larger multi-center studies, shared outcome definitions, outside-hospital testing, checks that predicted risks track real results, fairness audits, and proof that these tools fit workflows and are cost-effective.
Disclaimer: Abstracts were not reviewed by Neurology® and do not reflect the views of Neurology® editors or staff.