知镜 Cognispect · 医学文献与研究工作台

Artificial Intelligence in Neurology: A Critical Narrative Review of Clinical Evidence, Translation, and Responsible Implementation

Zahran A, Fuqha R, Milhem A, et al.
Brain Behav · 2026-11-01 · PMID 42992946

文献摘要

Purpose Artificial intelligence (AI) is increasingly applied to neuroimaging, electrophysiology, clinical text, and digitally measured behavior, but reported performance often exceeds the maturity of clinical evidence. Method We conducted a structured narrative review, informed by SANRA principles, of English-language human studies, reviews, and regulatory sources identified in PubMed/MEDLINE and IEEE Xplore from January 2022 through July 2026; older landmark studies were retained when necessary for historical or methodological context. We applied a common translational-evidence framework across six prespecified neurological domains to distinguish technical performance from external validation, prospective clinical evaluation, implementation, and demonstrated patient benefit. Evidence was appraised according to cohort representativeness, patient-level data separation, external and prospective validation, calibration, clinical comparators, regulatory status, workflow integration, and patient-level benefit. Finding Across stroke, dementia, Parkinson disease, epilepsy, multiple sclerosis, and neuro-oncology, the strongest evidence concerns narrowly defined imaging and signal-processing tasks. A cluster-randomized stroke study demonstrated shorter thrombectomy workflow time, and several externally validated models show improved portability. In contrast, many high-accuracy classification and segmentation studies remain retrospective, single-dataset, internally validated, and uncalibrated. Multimodal learning, foundation models, federated learning, and large language models broaden capability but introduce additional risks related to dataset shift, bias, opacity, privacy, and automation. Conclusion Neuro-AI should be judged by its intended clinical task and evidence maturity rather than headline accuracy. Multicenter external validation, prospective workflow evaluation, calibration, subgroup reporting, human-factors testing, and post-deployment surveillance are prerequisites for safe and equitable clinical use.

在知镜精读原文来源DOI 原文

知镜提供医学文献检索、阅读笔记、人体图谱、统计绘图与研究复现工具。