A team led by clinicians and AI researchers at Shanghai Jiao Tong University has unveiled DeepRare, an artificial‑intelligence system that can triage and help diagnose rare diseases using only patients' clinical symptoms. Published in Nature on 19 February 2026, the system is described by its developers as the world's first rare‑disease diagnostic tool whose reasoning steps are traceable — a claim aimed squarely at two of medicine's most persistent problems: the diagnostic odyssey for patients with uncommon conditions and clinicians' reluctance to trust opaque algorithms.
In controlled testing DeepRare reached a first‑time diagnostic accuracy of 57.18% when given clinical features alone, the team reports — an improvement of nearly 24 percentage points over the previous best international model. When genetic data are available, the system's accuracy rises above 70%. That gap matters: many smaller hospitals and clinics, particularly in less resourced regions, lack routine access to high‑throughput genetic testing and specialist interpretation. DeepRare's ability to operate effectively without genotype information could therefore shorten time‑to‑diagnosis where molecular testing is unavailable.
The work is the product of a collaboration between clinicians at Xinhua Hospital (affiliated with Shanghai Jiao Tong University School of Medicine) led by Professors Sun Kun and Yu Yongguo, and computational teams under Professor Zhang Ya and Associate Professor Xie Weidi. Publication in Nature confers scientific prestige and invites wider scrutiny, while the emphasis on a "traceable" inference process speaks to current debates over interpretability in medical AI: clinicians and regulators want to see why a model reached a conclusion before they act on it.
The practical implications are significant. Rare diseases collectively affect an estimated hundreds of millions of people worldwide, and many patients endure years of misdiagnosis or no diagnosis at all. Tools that reliably surface plausible diagnoses from routine clinical data can help primary‑care physicians identify cases that warrant referral, genetic testing, or specialist input, thereby concentrating scarce resources more efficiently and reducing patient suffering.
At the same time, important caveats remain. The article provides headline accuracy figures but does not substitute for extensive, prospective clinical validation across diverse patient populations and health systems. Performance may vary by the case mix, the quality and granularity of clinical notes, and genetic diversity. Furthermore, claims of traceability must be interrogated: explainability methods differ in how faithful they are to a model's internal logic, and superficially plausible rationales can still mislead if the underlying model has learned confounding patterns.
From a policy perspective, DeepRare sits at the intersection of two national objectives for China: to broaden access to higher‑quality health care and to develop competitive AI‑driven biomedicine. If robustly validated and integrated into clinical workflows, the technology could be a practical lever for improving rare‑disease diagnosis in underresourced hospitals domestically and become an exportable capability. Yet regulators, hospital administrators, and clinicians will need to agree standards for validation, liability, data privacy, and how algorithmic recommendations are presented to end users.
In short, DeepRare represents a noteworthy advance in the painful task of diagnosing rare conditions and a test case for the promise — and limits — of explainable AI in medicine. Its immediate value will depend on replication, transparency of methods, and careful implementation that augments rather than replaces clinical judgment.