AI-Based Screening for Rare Diseases

Abstract

Rare diseases collectively affect over 30 million Europeans but individually occur rarely enough that most clinicians encounter few cases in a career, contributing to diagnostic delays that average 4-7 years from symptom onset to correct diagnosis. The resulting diagnostic odyssey imposes substantial costs: progressive disease damage that could be prevented by earlier intervention, repeated invasive testing, psychological harm to patients and families, and missed therapeutic opportunities where disease-modifying treatments exist. AI-based screening approaches -- applying ML methods to EHR data, facial photographs, imaging, and laboratory profiles to flag patients with clinical patterns suggesting rare disease -- offer a scalable approach to reducing the diagnostic odyssey that targeted genetic testing alone cannot achieve. We evaluated five rare disease screening frameworks applied to 1.2 million patients across 18 European clinical centres and primary care networks in Berlin, Zurich, Stockholm, Vienna, and Dublin between 2018 and 2024, covering over 120 monogenic rare diseases with available treatments. Frameworks ranged from clinical pattern recognition alone through multimodal ML systems integrating EHR, imaging, facial analysis, and genomic screening. Performance was benchmarked using our Rare Disease Screening Effectiveness Score (RDSES), integrating sensitivity at clinically acceptable specificity, diagnostic latency reduction, positive predictive value, implementation feasibility, and ethical acceptability. The multimodal AI screening framework achieved the highest RDSES (0.917), identifying 72.4% of undiagnosed rare disease cases in routine clinical encounters while maintaining specificity above 99% -- a sensitivity improvement of 54.2 percentage points over clinical pattern recognition alone. Lysosomal storage disorders, inherited metabolic diseases, and neuromuscular conditions showed the largest screening benefits, with median time-to-diagnosis reduction from 5.4 years to 8.2 months.