Artificial intelligence for pediatric rare disease diagnosis: a multimethod study integrating published evidence and clinician interviews.

backgroundPediatric rare diseases are highly heterogeneous and are frequently associated with missed or delayed diagnosis, creating substantial burden for patients, families, and clinicians. Although artificial intelligence (AI), including large language model-enabled approaches, has shown potential for diagnostic support, translation into real-world pediatric care remains limited. A key gap is the mismatch between metric-centric evidence reporting and clinician-defined implementation needs. To address this, we integrated published evidence with clinician perspectives to derive an implementation-oriented evidence-to-requirements framework for AI-assisted pediatric rare-disease diagnosis.

methodsWe used a convergent multimethod design with two complementary evidence sources. First, we conducted a PRISMA-ScR scoping review of four databases (PubMed, Embase, Web of Science, and Scopus) from inception to December 2025 and included 28 original studies on AI-assisted pediatric rare-disease diagnosis. Second, we conducted semi-structured interviews with 21 pediatric clinicians from 15 departments at a tertiary children's hospital in Chongqing, China, and analyzed transcripts using inductive thematic analysis. We then integrated findings side-by-side to identify convergences, divergences, and translational gaps.

resultsThe scoping review showed rapid movement toward multimodal and LLM-enabled approaches across several diagnostic task types, including screening or cohort identification, phenotyping, differential diagnostic support, and variant or gene prioritization. Translation-oriented evidence remained uneven, with limited prospective evaluation and inconsistent reporting of fairness, safety, and deployment context. Interview analysis identified four recurrent themes: diagnosis as time-pressured puzzle-solving; AI as a cognitive extender rather than replacement; trust dependent on traceable evidence and transparent reasoning; and demand for structured, actionable outputs with low workflow burden. Integrated analysis revealed a persistent implementation gap between metric-centric publication practices and clinician-defined requirements for real-world adoption.

conclusionsThis scoping review and qualitative interview study does not establish clinical effectiveness of AI-assisted diagnosis. Instead, it identifies implementation requirements that may guide future development and evaluation, including representative multicenter data, prospective validation, evidence traceability, actionability, safety, fairness, and workflow fit. Main limitations include restriction to English-language studies, reliance on umbrella rare-disease terminology, possible missed studies among unscreened records after ASReview-assisted screening, and a single-institution clinician interview sample.