Machine Learning, Large Language Models, and Multimodal AI for Diagnosing Pediatric Rare Diseases: Scoping Review

Pediatric rare diseases often cause a prolonged diagnostic odyssey. AI, including machine learning, deep learning, large language models (LLMs), and multimodal systems, may support diagnosis, but these applications in children have not been systematically mapped.

The aim of the study is to map diagnostic applications, data modalities, validation strategies, and evidence maturity of AI methods for pediatric rare diseases.

We conducted a scoping review following Joanna Briggs Institute methodology and reported it according to PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews). On June 26, 2026, we searched PubMed, Scopus, Web of Science Core Collection, Embase, China National Knowledge Infrastructure (CNKI), Wanfang Data, and the Cochrane Library for records published from January 1, 2015, through June 1, 2026. We additionally searched medRxiv and arXiv and hand-searched the reference lists of included studies and relevant reviews. Eligibility was defined using the population-concept-context framework: pediatric rare diseases, diagnostic AI, and any clinical or research setting. JZ and JL independently screened titles and abstracts and assessed potentially eligible full-text reports. JZ charted the data, and JL verified every field. Findings were synthesized descriptively according to disease focus, AI technology, input modality, diagnostic task, validation strategy, and evidence maturity.

Database searches identified 2557 records; 2063 remained after deduplication. Of 106 full-text reports assessed, 77 database studies and 4 studies from hand searching and preprint servers were included, yielding 81 studies. Studies were published from 2016 through 2026, with 55 of 81 (67.9%) published from 2024 through 2026. Using a mutually exclusive primary technology classification, classical machine learning accounted for 38 (46.9%) studies, facial AI for 18 (22.2%), deep learning for 15 (18.5%), LLMs for 6 (7.4%), and multimodal AI for 4 (4.9%). Electronic health records, claims, clinical text, or structured clinical vignettes were used in 25 (30.9%) studies, facial images in 17 (21%), and other medical imaging in 13 (16%). Evidence remained mainly retrospective and internally validated: 62 (76.5%) studies included a retrospective component and 76 (93.8%) reported internal validation, whereas 21 (25.9%) included external validation and 12 (14.8%) included a prospective component.

Research on AI-assisted diagnosis of pediatric rare diseases has expanded rapidly, but evidence maturity has not kept pace. Most studies established technical feasibility rather than generalizable clinical benefit, and performance should be interpreted by task, inputs, reference standard, and validation design rather than used to rank technologies. Evidence for LLMs and multimodal AI remains limited. Future research should prioritize multicenter validation, reproducible task-specific benchmarks, prospective evaluation, and assessment of incremental clinical value.