AI-assisted recognition of misdiagnosed paroxysmal nocturnal hemoglobinuria

Abstract Background Paroxysmal nocturnal hemoglobinuria (PNH) is often misdiagnosed, delaying effective therapy. In this proof-of-concept study, we systematically reviewed delayed or misdiagnosed PNH cases and evaluated whether large language models could include PNH among the leading differential diagnoses using standardized clinical vignettes. Results A systematic review identified 63 publications describing 68 patients with confirmed PNH and diagnostic delay. Each case was converted into a standardized vignette and tested across 3 large language models under complete and feature-removal conditions. The prespecified primary outcome was PNH diagnostic rank (1 = top diagnosis; 2–5 = within top five; 6 = not listed). Model performance varied, with PNH included among the top five diagnoses in approximately 81%-89.7% of complete vignettes. Descriptive scenario-level analysis suggested poorer PNH ranking when both urinary discoloration or urinary blood descriptors and laboratory hemolysis cues were absent, a finding limited by unaccounted source-case matching. Testing was typically triggered by unexplained hemolysis, 45.6%, or thrombosis, 36.8%, while frequent initial diagnostic attributions included iron-deficiency anemia, aplastic anemia, myelodysplastic syndromes, and urinary tract disease. The case cohort had a lower median reported age than the Registry baseline population, 30 versus 43.7 years, and had a median reported diagnostic delay of 24 months, IQR 12–60. Race and ethnicity were reported in 23.5% of cases. Limitations include selected case literature, curated vignettes, historical heterogeneity, and single-pass evaluation of evolving commercial models. Conclusions In selected published cases converted into standardized vignettes, large language models frequently included PNH among the leading differential diagnoses when hemolysis-related information was available. These findings support further evaluation of AI as a case-finding tool to identify patients in whom PNH diagnosis may be delayed and who may warrant hemolysis assessment and confirmatory flow cytometry, with real-world diagnostic performance requiring prospective validation.