JMIR Publications

ABSTRACT

Background:

Pediatric rare diseases often cause a prolonged diagnostic odyssey. Artificial intelligence (AI)—including machine learning (ML), deep learning (DL), large language models (LLMs), and multimodal systems—may support diagnosis, but the landscape of applications in children has not been mapped systematically.

Objective:

To map and synthesize the published evidence on the use of machine learning, deep learning, large language models, and multimodal artificial intelligence for diagnosing pediatric rare diseases, and to identify key gaps in validation, reporting, and real-world clinical implementation.

Methods:

We conducted a scoping review following JBI methodology and PRISMA-ScR. We searched PubMed, Scopus, and Web of Science from 1 January 2015 to 31 December 2025 for studies using ML/DL, LLMs, or multimodal AI to diagnose rare diseases in children (0–18 years). Eligibility was defined by the PCC framework (Population: pediatric rare disease; Concept: diagnostic AI; Context: any setting). Two reviewers screened titles/abstracts and full texts; data were extracted on disease type, model type, input modality, performance metrics, and reported challenges.

Results:

We retrieved 237 records (1 January 2015 to 31 December 2025) and removed 42 duplicates. After screening 195 unique records, we excluded 140 at title/abstract. Of 55 full-text reports assessed for eligibility, 13 were excluded; 42 studies were included. Studies spanned 2018–2025, with a sharp rise from 2019 and a peak of 16 in 2025. By technology, classical ML accounted for 23 studies (55%), deep learning for 8 (19%), facial AI for 4 (10%), multimodal for 4 (10%), and LLM for 3 (7%). The most represented disease groups were Mendelian/rare genetic (9), metabolic/newborn screening (5), and genetic syndromes (4). Input modalities were dominated by EHR/claims/text (15), facial image (7), and genomic data (5). Reported performance varied widely; LLM studies showed modest top-1 accuracy (8–13% on rare pediatric case reports) and emphasized the need for clinical oversight. Common challenges included single-center or retrospective design, small samples, limited external validation, and ethical issues such as privacy, bias, and false positives/negatives.

Conclusions:

The literature on ML, LLMs, and multimodal AI for diagnosing pediatric rare diseases is growing but fragmented. Classical ML and deep learning dominate the evidence base, while LLM and multimodal applications remain limited and LLM performance in rare pediatric settings is modest. Key gaps include prospective and multi-center validation, LLM studies in Chinese pediatric rare disease practice, and integration of agentic AI with real-world clinical workflows. This scoping map can guide future primary studies and systematic reviews. Clinical Trial: The protocol for this scoping review is registered on PROSPERO (CRD420261326146).