Abstract
OBJECTIVE: To evaluate the diagnostic accuracy of artificial intelligence (AI)-generated clinical diagnoses. PATIENTS AND METHODS: A retrospective chart review of 102,059 virtual primary care clinical encounters from October 1, 2022, to January 31, 2023 was conducted. Patients underwent an AI medical interview, after which virtual care providers reviewed the interview summary and AI-provided differential diagnoses, communicated with patients, and finalized diagnoses and treatment plans. Our accuracy measures were agreement between AI diagnoses, virtual care providers, and blind adjudicators. We analyzed AI diagnostic agreement across different diagnoses, presenting symptoms, patient demographic characteristics such as race, and provider levels of experience. We also evaluated model performance improvement with retraining. RESULTS: Providers selected an AI diagnosis in 84.2% (n = 85,976) of cases and the top-ranked AI diagnosis in 60.9% (n = 62,130) of cases. Agreement rates varied by diagnosis, with greater than or equal to 95% provider agreement with an AI diagnosis for 35 diagnoses (47% of cases, n = 47,679) and greater than or equal to 90% agreement for 57 diagnoses (69% of cases, n = 70,697). The average agreement rate for half of all presenting symptoms was greater than or equal to 90%. Adjusting for case mix, diagnostic accuracy exhibited minimal variation across demographic characteristics. The adjudicators' consensus diagnosis, reached in 58.2% (n = 128) of adjudicated cases was always included in the AI differential diagnosis. Provider experience did not affect agreement, and model retraining increased diagnostic accuracy for retrained conditions from 96.6% to 98.0%. CONCLUSION: Our findings show that agreement between AI and provider diagnoses is high in most cases in the setting of this study. The results highlight the potential for AI to enhance primary care disease diagnosis and patient triage, with the capacity to improve over time.