AI Model Accelerates Rare Disease Diagnosis for Physicians

Rare diseases pose a significant global health challenge, with timely diagnosis often hampered by a lack of specialized clinical expertise. While large language models (LLMs) show promise in this area, their clinical utility has been limited by insufficient deployability, a scarcity of clinically grounded evidence, and a lack of training data. This research introduces RaDaR (Rare Disease navigatoR), an open-source, compact 32-billion-parameter reasoning LLM specifically designed for rare disease diagnosis. RaDaR was trained on a substantial dataset comprising 49,170 publicly available free-text cases and an additional 104,666 synthetically generated cases, enhanced with reasoning-focused training. RaDaR demonstrated superior performance compared to other evaluated open-source models, including the much larger 671B DeepSeek-R1, across various public benchmarks and four external validation centers. In a retrospective study, RaDaR prioritized the correct diagnosis before clinical suspicion was documented in over 61% of cases, potentially reducing diagnostic lead time by nearly two months. Furthermore, a randomized physician-assistance trial showed that RaDaR improved diagnostic accuracy by over 21 percentage points compared to internet search alone. The study also highlighted the effectiveness of phenotype-anchored synthetic narratives in training for long-tail rare diseases, providing a deployable model and a reproducible framework for diagnostic AI in data-scarce medical fields.