DeepRare surpasses physicians in a study focused on diagnosing rare diseases.

      DeepRare, an agentic AI system comprising 40 specialized tools, surpassed medical professionals in detecting rare conditions in a direct comparison study published in Nature.

      For countless individuals with rare diseases, the journey to diagnosis is complex. Patients often navigate through general practitioners and various specialists for many years, sometimes decades, trying to assemble symptoms that do not align with typical presentations.

      Eighty percent of rare diseases are genetically based, yet many remain undiagnosed until significant biological damage has occurred. The main challenge lies not in a lack of data, but in finding the vital information within the vast medical landscape.

      A recently published study in Nature indicates that artificial intelligence might expedite this search. Researchers from Shanghai Jiao Tong University’s School of Artificial Intelligence and Xinhua Hospital developed DeepRare, an AI system aimed at simulating how human doctors address diagnostic uncertainties.

      In a direct comparison with five seasoned physicians, each with over ten years of experience, the AI system demonstrated superior accuracy overall.

      The results are notable. DeepRare successfully identified the disease on its initial assessment 64.4 percent of the time, while the doctors achieved a success rate of 54.6 percent. When presented with three suggestions rather than one, the AI reached a diagnostic correctness rate of 79 percent compared to 66 percent for the human specialists.

      Importantly, the physicians agreed with the AI’s reasoning 95.4 percent of the time, indicating that the system not only arrives at correct conclusions but does so in ways that resonate with seasoned clinicians as logical and sound.

      What sets DeepRare apart from previous diagnostic AI systems is its framework. Rather than using a black-box classification model, the system incorporates 40 specialized digital tools and adheres to a clearly articulated reasoning process.

      It generates diagnostic hypotheses, tests them against patient data, explores global medical literature databases, assesses genetic variants, and iteratively refines its conclusions before ranking the possibilities.

      This method reflects the cognitive processes a human diagnostician employs, yet benefits from access to comprehensive medical knowledge and computational speed that far exceed human capabilities.

      The system is already operational beyond research settings. Since July 2025, DeepRare has been implemented on an online diagnostic platform, with over 600 medical institutions worldwide signing up to utilize it.

      The research team aims to further validate the system with 20,000 real-world cases and establish a global alliance for rare disease diagnostics. The authors stress that the system is meant to enhance, not replace, clinicians’ roles, recognizing both the technical limitations of AI and the indispensable human aspect of medical practice.

      The potential impact on patients is significant. An estimated 300 million individuals globally suffer from rare diseases, and the average diagnostic journey can extend to five years or more.

      Each year of delayed diagnosis is filled with uncertainty, incorrect treatments, and the risk of worsening organ damage. An AI system that can shorten this timeline by weeks or months, and highlight potential diagnoses that may otherwise be missed, could significantly alter the initial experience of those living with rare conditions.