LungDiag Improves Respiratory Case Answers in Randomised Simulated-Case Study
Published in Artificial Intelligence, Journals, Tools.
A Nature Health study published on October 5, 2026 compared LungDiag, a respiratory-focused GPT-4o chatbot inside WeChat, with conventional mobile web search. Researchers randomised 2,400 adults without medical training across 24 Chinese healthcare systems.
The primary compliant-case analysis included 2,176 questionnaires. Model-based overall accuracy was 70.0% with LungDiag versus 55.4% with web search, a difference of 14.6 percentage points. Dedicated AI and generative-AI-assisted search were disabled in the control group.
The largest improvement concerned diagnosis questions: 81.2% versus 54.5%. Triage accuracy improved more modestly, from 47.1% to 53.4%. Participants using LungDiag took longer, with a median completion time of 527 seconds versus 461 seconds.
This was a non-interventional comparison using simulated cases. It does not demonstrate better real-world care-seeking, shorter diagnostic delays or improved clinical outcomes. The results support further evaluation of a constrained information tool, rather than autonomous diagnosis or triage.
Source: Liang and colleagues, Nature Health, October 5, 2026.