Running summarizing tools on a new article

Abstract: Importance : Artificial intelligence (AI) holds potential to revolutionize decision-making in healthcare but remains understudied in primary care.Objective: To evaluate diagnostic accuracy of AI-generated clinical diagnoses.Design: Retrospective chart review.Setting: Virtual primary care practice.Participants: Patients presenting to a virtual care provider with acute symptoms.Exposure: Patients underwent an AI medical interview, received a list of possible related conditions, and elected to visit with a virtual care provider. Providers reviewed medical interview summary and AI differential diagnoses before communicating with patients, finalizing the diagnosis, and rendering treatment decisions.Main Outcomes and Measures: Primary outcomes were agreement between AI diagnoses, virtual care providers, and blind adjudicators and AI performance across demographic characteristics such as race. Secondary aims included AI diagnostic accuracy across presenting symptoms and model performance improvement with retraining.Results: A total of 102,059 clinical encounters with 133 providers were reviewed. Virtual care providers selected an AI diagnosis in 84.2% of cases and the top-ranked AI diagnosis in 60.9% of cases. Agreement varied by diagnosis, with providers selecting an AI diagnosis ≥95% of the time for 35 diagnoses (47% of cases) and ≥90% of the time for 57 diagnoses (69% of cases). One-half of all presenting symptoms had an average diagnostic agreement rate ≥90%. Minimal variation was observed in average diagnostic accuracy based on demographics after adjusting for group differences in case mix. For 63.7% of cases with adjudicators' diagnostic consensus, AI included the diagnoses in the differential 100% of the time. AI accuracy was higher for eye and urinary symptoms and lower for skin conditions. Model retraining was associated with increased diagnostic accuracy (96.6% to 98.0%) for retrained conditions.Conclusion and Relevance: Our findings demonstrate that AI models exhibit high diagnostic accuracy in a virtual primary care setting. High provider-to-AI agreement suggests a potential role of AI integration into primary care for aiding in disease diagnosis and patient triage with the ability to improve over time. Further studies are required to investigate the impact of AI on patient outcomes and cost-effectiveness, and the broader implications of AI implementation in primary care settings.