LiteOdyssey is a lightweight framework for rare disease diagnosis that achieves efficient diagnostic performance by extending the reasoning chain of a single AI agent. This system utilizes a diagnostic policy developed through human-AI collaboration and augments it with freely available biomedical tools, avoiding the complexities associated with scaling traditional medical AI systems. By employing Policy Iteration with Human Feedback (PIHF), LiteOdyssey guides a reasoning language model through clinical genetics workflows. On two challenging benchmarks, LIRICAL (n = 370) and PhenoPacket Store (n = 873), LiteOdyssey achieved state-of-the-art performance with an overall disease Recall@1 of 59.3%. Both benchmarks feature a high proportion of ultra-rare diseases (prevalence below 1 in 1,000,000, with ultra-rare shares of approximately 45% and 52.8% respectively). On the more difficult PhenoPacket subset, LiteOdyssey achieved a Recall@1 of 60.7%, compared to just 10.7% for the baseline model (GPT-5.4) without tools. This performance was accomplished without fine-tuning, multi-agent ensembles, or a large case-retrieval database. Gains were also observed in previously unseen cases, a private cohort of real-world rare disease patients, and a smaller open-weights model. LiteOdyssey suggests a path toward more accurate, easier to deploy, and more transparent AI systems for rare diseases.
Blogger's Review: LiteOdyssey's innovation lies in its lightweight design and human-AI collaborative reasoning strategy, overcoming the complexities of traditional medical AI systems. This approach not only enhances diagnostic accuracy but also improves the system's auditability and usability, holding significant potential for clinical applications.