Reimagining Care of People Living with Rare Diseases With Artificial Intelligence

Perspectives are commissioned from an expert and discuss the clinical practice or public health implications of a published study. The original publication must be freely available online.

Affiliations Bioinformatics Institute (BII), Agency for Science, Technology and Research (A*STAR), Singapore, Singapore, Maternal and Child Health Research Institute, KK Women’s and Children’s Hospital, Singapore, Singapore, Genetics Service, KK Women’s and Children’s Hospital, Singapore, Singapore, SingHealth Duke-NUS Institute of Precision Medicine, Singapore, Singapore, Duke-NUS School of Medicine, Singapore, Singapore

Affiliation Rare Care Centre, Perth Children’s Hospital, Perth, Western Australia, Australia

Affiliations Genetics Service, KK Women’s and Children’s Hospital, Singapore, Singapore, SingHealth Duke-NUS Institute of Precision Medicine, Singapore, Singapore, SingHealth Duke-NUS Genomic Medicine Centre, Singapore, Singapore

Artificial intelligence (AI) can transform rare disease care when organized around the patient journey. We outline a patient–clinician–AI triad spanning early detection, diagnosis, clinical trials, and individualized therapies.

Citation: Groza T, Baynam G, Jamuar SS (2026) Reimagining care of people living with rare diseases with artificial intelligence. PLoS Med 23(2): e1004966. https://doi.org/10.1371/journal.pmed.1004966

Copyright: © 2026 Groza et al. This is an open access article distributed under the terms of the Creative Commons Attribution License , which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.

Funding: This work was supported by the National Medical Research Council (NMRC) Singapore (Clinician Scientist Award NMRC/CSAINVJun21-0003 and NMRC/CSAINV24jul-0001 to SSJ), the National Medical Research Council (NMRC) Research, Innovation and Enterprise (RIE2025) Centre Grant seed funding (NMRC/CG1/006/2021-KKH to TG), the Angela Wright Bennett Foundation (to GB), The Stan Perron Charitable Foundation (to GB), the Channel 7 Telethon Trust (to GB) and the Perth Children’s Hospital Foundation (to GB). TG received salary from the National Medical Research Council (NMRC) Singapore NMRC/CSAINVJun21-0003.

Competing interests: The authors have declared that no competing interests exist.

Abbreviations: AADCd, aromatic L-amino acid decarboxylase deficiency; AI, artificial intelligence; EHRs, electronic health records

Rare diseases collectively affect hundreds of millions of people worldwide, yet individual conditions are rare, heterogeneous, and often poorly characterized. Patients and families frequently experience years of misdiagnoses, fragmented care, and social disruption—a journey often described as the “diagnostic odyssey.” Rare diseases therefore function both as a stress test for health systems and as a proving ground for digital health infrastructures and emerging artificial intelligence (AI) technologies [ 1 , 2 ].

Recent work has outlined how AI could support rare conditions across public health surveillance, symptom matching, and digital therapeutics, offering a system-level perspective rather than a narrow diagnostic focus. At the same time, analyses of AI in primary care and surveys of clinicians highlight persistent barriers to implementation, including workflow integration, data quality, trust, and accountability [ 2 , 3 ]. In this Perspective, we argue that progress in AI for rare diseases should be organized along the patient journey—from early suspicion, through diagnosis and treatment development—and grounded in a “triad” between patient-family experts, clinicians, and AI, with each contributing distinct and complementary forms of expertise ( Fig 1 ).

A conceptual view of the patient–clinician–AI triad, showing how AI supports the care pathway from initial presentation, health system screening, and diagnosis support through to clinical trial design (including eligibility, stratification, and N-of-1 trials), as well as drug repurposing/development via in silico modeling, toxicity simulation, and prioritization. Icons made by Freepik from https://www.flaticon.com .

Many people with rare diseases leave extensive digital signals in electronic health records (EHRs) long before a rare condition is suspected. These signals include repeated non-specific presentations, unusual constellations of symptoms, and clusters of abnormal laboratory findings distributed across multiple encounters and care settings. In response, AI-based approaches have been developed to retrospectively and prospectively identify such patterns and flag patients who may warrant further evaluation.

Existing methods span a spectrum from relatively transparent, rule-based phenotypic scores to data-driven models that learn patterns across longitudinal records and clinical narratives. Concrete examples illustrate how these approaches translate routine EHR data into actionable rare disease signals. For example, Cohen and colleagues [ 4 ] used natural language processing to pre-screen for aromatic L-amino acid decarboxylase deficiency (AADCd) by mining unstructured clinical notes for recurring combinations of early, non-specific neurological features (e.g., hypotonia, movement disorders, developmental delay) distributed across encounters and specialties. Patients were ranked by similarity to a prototypical AADCd phenotype, and retrospective evaluation showed that confirmed or likely cases could have been flagged years earlier using documentation already present in the EHR. Similarly, Faviez and colleagues [ 5 ] applied a deep-learning natural language processing pipeline to enrich longitudinal phenotyping for Jeune syndrome, extracting and normalizing rich skeletal and extra-skeletal features and improving discrimination from phenotypically overlapping conditions. Together, these studies show how AI can integrate repeated non-specific encounters and fragmented symptom documentation into disease-specific phenotypic signatures that support earlier recognition in real-world clinical data. However, despite these promising results, the applications remain disease-specific and retrospective, and their generalizability and prospective impact require further validation.

At the health-system level, several deployments now suggest that AI-based rare disease screening is technically feasible at scale, with tools scanning hundreds of thousands of records across care networks and identifying patients for targeted review [ 6 ]. These deployments also reveal important constraints. System-wide screening requires sustained computational infrastructure, robust integration with heterogeneous clinical systems, and careful calibration of alert thresholds so that flagged patients can be reviewed without overwhelming services. Data fragmentation and variable data quality remain major barriers, as rare disease signals are often split across institutions and care episodes. Governance and trust are therefore central, particularly when screening undiagnosed populations.

Once a rare disease is suspected, AI tools can also support diagnostic reasoning by integrating genetic, phenotypic, and imaging data. AI-assisted pipelines can prioritize candidate variants, match symptom profiles to diagnoses, and synthesize multimodal evidence with performance levels that in some settings (when detailed phenotyping is available) approach that of expert clinicians [ 1 , 2 ]. Diagnosis, however, is rarely a single event. It unfolds over time as symptoms evolve and new information accumulates. Consistent with this, evidence suggests that the greatest clinical value arises when AI tools are embedded within broader diagnostic pathways that include specialist review, confirmatory testing, counseling, and follow-up, rather than operating as standalone decision-makers ( Fig 1 ).

Once a diagnosis is made, however, treatment is far from guaranteed. Despite the identification of more than 7,000 rare diseases, effective disease-modifying treatments exist for fewer than 10% of them, leaving the vast majority of patients without evidence-based therapeutic options. Accelerating the generation of high-quality clinical trial evidence is therefore a central unmet need in rare disease care [ 7 ]. The same features that complicate diagnosis—i.e., small patient numbers, clinical heterogeneity, and fragmented data—also challenge conventional clinical trial designs for rare diseases. Around 30%–50% of interventional rare disease trials enroll fewer than 50 participants, and many fail to reach planned sample sizes because patients are geographically dispersed and narrowly phenotyped [ 8 ]. As a result, trials often rely on single-arm designs with external or natural-history comparators, enriched inclusion criteria, and short, clinically meaningful endpoints rather than large randomized trials [ 9 ]. Analytical strategies also differ: Bayesian methods that incorporate prior knowledge and real-world evidence are increasingly recommended where frequentist approaches would be infeasible [ 8 , 10 ]. These adaptations are now established practice but introduce additional methodological and operational complexity.