A multicentre, prospective, cluster-randomised, parallel-controlled real-world effectiveness study evaluating whether a rare-disease diagnostic large language model can improve diagnostic quality, efficiency, and health-economic outcomes for physicians managing patients with suspected rare or diagnostically unresolved disease.

Rare disease patients commonly experience prolonged diagnostic odysseys rooted in limited rare disease recognition, phenotypic heterogeneity, and dispersed diagnostic clues. Diagnostic decision-support large language models may improve first-visit consultations by integrating prior records, generating structured analyses, and proposing candidate diagnoses, thereby shortening diagnostic pathways and improving appropriate genetic testing referral.

Here, rare diseases clinics are established at 12 clinical institutions using a two-level randomisation structure: (1) cluster randomisation of physicians stratified by seniority tier and centre, with physician arm fixed throughout to eliminate contamination; (2) eligible consented patients randomly assigned to consultation rooms, blinding patients to arm assignment and providing allocation concealment.

In the intervention arm, patients interact with the AI before their consultation; physicians will review the AI report and interact with AI when seeing patients. In the control arm, patients are seen under standard hospital workflow without any generative AI tools. Outcomes adjudicated by an independent Expert Committee blinded to arm assignment; adjudicators access no AI-generated materials.

Patient Inclusion Criteria:

• Any age. Legal guardian co-signs consent for minors or individuals lacking legal capacity.
• Diagnostically unresolved or suspected rare disease, with at least one prior complete clinical evaluation at a secondary-level or higher institution yielding no confirmed explanatory diagnosis.
• First presentation to the enrolling institution for the current condition, with no prior records in the institutional HIS or outpatient system.
• No prior genetic testing related to the current condition; no results or reports available.
• Written informed consent provided voluntarily by patient or legal guardian, with commitment and ability to complete structured follow-up.

Patient Exclusion Criteria:

• Confirmed diagnosis (clinical, pathological, or molecular) explaining the primary symptoms.
• Emergency presentation, critical illness, or any condition incompatible with trial participation.
• Neither patient nor legally authorised proxy able to complete follow-up.
• Concurrent enrollment in another interventional study with diagnostic accuracy or genetic testing yield as a primary endpoint.
• Prior use of another AI system has already yielded a confirmed diagnosis for the current condition.

Physician Inclusion Criteria

• Licensed physician in internal medicine, neurology, pediatrics, general medicine, rare disease, or a related specialty.
• ≥2 years of clinical practice; competent to manage rare disease patients; stratified into junior or senior tier.
• Voluntary participation with written informed consent.

Physician Exclusion Criteria

• No longer in clinical practice, or unable to fulfill required outpatient duties during the study period.
• Unwilling to provide informed consent or to permit protocol-required collection of consultation and questionnaire data.
• Currently enrolled in another AI-assisted clinical workflow, or expected to be unable to comply with the procedures.

Patient level: open label (intervention arm patients must interact with AI before consultation, making concealment impossible).

Treating physicians: open label (as they cannot be blinded). Outcome adjudicators (independent expert committee) and statisticians: blinded. Adjudicators do not know arm assignment and are not shown any AI-generated or AI-attributed material; statistical analysis coded A/B until unblinding.