An AI-Powered Symptom Assessment and Triage Platform for Virtual Primary Care: A Prospective, Real-World Validation Study

Background and Objective: Telemedicine expansion demands accurate AI-driven symptom assessment for remote triage, yet prospective real-world evidence is lacking. We evaluated an AI-powered platform (TeleTriage-AI) in virtual primary care for diagnostic concordance, triage safety, and user satisfaction. Methods: A prospective observational study enrolled ۱,۴۸۲ adult patients with new complaints across three telehealth clinics (September ۲۰۲۳–March ۲۰۲۴). TeleTriage-AI, a fine-tuned clinical large language model, generated differential diagnoses and triage recommendations (emergency to self-care). Primary outcomes were diagnostic concordance (full/partial) with physician-assigned diagnosis and triage safety versus a blinded physician panel. Secondary outcomes included consultation time and net promoter score (NPS). Results: Among ۱,۴۸۲ encounters (۱۸۹ chief complaints), full diagnostic concordance was ۶۸.۴% (۹۵% CI ۶۶.۰–۷۰.۸) and partial ۲۲.۱%, yielding clinically acceptable concordance of ۹۰.۵%. For triage, sensitivity for urgent/emergency cases was ۹۴.۲% (۹۵% CI ۹۰.۹–۹۶.۵), specificity ۸۲.۷% (۹۵% CI ۸۰.۱–۸۵.۱), and critical under-triage rate ۱.۸%. Median consultation time decreased ۲۹.۴% (from ۱۴.۳ to ۱۰.۱ minutes, p<۰.۰۰۱). NPS was +۴۸, indicating high satisfaction. Conclusion: TeleTriage-AI achieved high diagnostic concordance, acceptable triage safety, and substantial time savings in real-world virtual care. AI-driven symptom assessment can enhance telehealth efficiency and access with ongoing safety monitoring.

کلیدواژه ها:

Artificial intelligence ، Telemedicine ، Triage ، Symptom assessment ، Large language model ، Virtual primary care