Scholarlink

Background <p>Rare neurological disorders, affecting approximately 30–40% of the estimated 400 million individuals with rare diseases worldwide, present fundamental challenges to artificial intelligence (AI) development due to limited patient cohorts, heterogeneous phenotypes, and fragmented data infrastructure. Whether recent advances in foundation models and large language models can overcome these “small-n” constraints remains an open question with profound implications for precision neurology.</p> Results <p>From 1847 records identified, 89 studies met inclusion criteria after screening. The evidence concentrates in five disease clusters: amyotrophic lateral sclerosis (ALS, 28%), Huntington disease (HD, 18%), myasthenia gravis (MG, 14%), muscular dystrophies (12%), and rare epilepsies (9%). Classical machine learning approaches (random forests, SVMs) predominate (51%), with deep learning (convolutional neural networks, recurrent neural networks) comprising 34%, and foundation model–based approaches representing an emerging but rapidly growing 15%. Neuroimaging is the most common data modality (47%), followed by genomic/multi-omic (26%) and clinical/electrophysiological data (27%). Only 24% of studies reported external validation, and among those, approximately one-third demonstrated clinically significant performance degradation. A parallel assessment of 156 international rare disease registries revealed that only 22% utilize internationally recognized Common Data Elements (CDEs), identifying semantic interoperability as the critical bottleneck for federated AI implementation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Artificial intelligence in rare neurological diseases: a scoping review of the evidence landscape and framework for foundation model integration

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月29日

查看全文
  • Shih-Shuan Fang,
  • Shenghan Chen

摘要

Background

Rare neurological disorders, affecting approximately 30–40% of the estimated 400 million individuals with rare diseases worldwide, present fundamental challenges to artificial intelligence (AI) development due to limited patient cohorts, heterogeneous phenotypes, and fragmented data infrastructure. Whether recent advances in foundation models and large language models can overcome these “small-n” constraints remains an open question with profound implications for precision neurology.

Results

From 1847 records identified, 89 studies met inclusion criteria after screening. The evidence concentrates in five disease clusters: amyotrophic lateral sclerosis (ALS, 28%), Huntington disease (HD, 18%), myasthenia gravis (MG, 14%), muscular dystrophies (12%), and rare epilepsies (9%). Classical machine learning approaches (random forests, SVMs) predominate (51%), with deep learning (convolutional neural networks, recurrent neural networks) comprising 34%, and foundation model–based approaches representing an emerging but rapidly growing 15%. Neuroimaging is the most common data modality (47%), followed by genomic/multi-omic (26%) and clinical/electrophysiological data (27%). Only 24% of studies reported external validation, and among those, approximately one-third demonstrated clinically significant performance degradation. A parallel assessment of 156 international rare disease registries revealed that only 22% utilize internationally recognized Common Data Elements (CDEs), identifying semantic interoperability as the critical bottleneck for federated AI implementation.