Rare disease AI promises progress, but the evidence gap is still the bottleneck
Open Access Government asks whether AI can live up to its promises for rare diseases, where data scarcity and fragmented care have long constrained diagnosis and treatment. The central challenge is not model ambition, but proof in low-volume, high-variability conditions.
Rare diseases are one of the clearest use cases for AI because the clinical problem is obvious: too few cases, too much complexity, and too much time lost to diagnostic odysseys. In theory, machine learning can help surface patterns humans miss, connect scattered records, and accelerate matching to trials or specialists.
But rare disease care also exposes AI’s hardest limitation: most systems learn from abundant examples, while rare diseases are defined by the opposite. That means performance claims can be fragile, especially when models are trained on proxy labels, incomplete phenotypes, or data from a handful of academic centers.
The real promise may lie less in diagnosis alone and more in orchestration. AI could help triage patients toward genetic testing, summarize longitudinal records, identify relevant literature, and reduce the burden on families who often become their own care coordinators. Those are meaningful gains even if models never become perfect diagnosticians.
For policymakers and developers, the article serves as a reminder that rare disease AI needs special evaluation standards. Success should be measured not only by accuracy, but by time-to-diagnosis, referral quality, and patient navigation. Without that broader lens, the field risks celebrating elegant demos while missing the everyday realities of rare disease care.