Researchers present RaDaR, an open-source 32B parameter reasoning large language model designed to accelerate the diagnosis of rare diseases by addressing challenges in clinical deployability and data scarcity. The model was trained on nearly 50,000 public cases and over 100,000 synthetic cases, demonstrating superior performance across benchmarks and external validation centers.
- RaDaR prioritized final diagnoses before documented clinical suspicion in 61.06% of retrospective cases, offering a potential lead time of 1.87 months.
- In a randomized physician-assistance trial, RaDaR improved diagnostic accuracy by 21.44 percentage points compared to internet search alone.
- The model outperformed evaluated open-source models, including the 671B DeepSeek-R1, across public benchmarks and four external validation centers.
- Synthetic-data ablations indicate that phenotype-anchored narratives provide a useful training signal for long-tail rare diseases with a monotonic scaling trend.
RaDaR provides a deployable reasoning model and a reproducible development framework for diagnostic AI, helping to overcome the scarcity of specialized clinical expertise and training data in rare disease diagnosis.