Generative artificial intelligence models in clinical infectious disease consultations: a cross-sectional analysis among specialists and resident trainees

ABSTRACT Background The potential of generative artificial intelligence (GenAI) to augment clinical consultation services in clinical microbiology and infectious diseases (ID) is being evaluated. Methods This cross-sectional study evaluated the performance of four GenAI chatbots (GPT-4.0, a Custom Chatbot based on GPT-4.0, Gemini Pro, and Claude 2) by analysing 40 unique clinical scenarios synthesised from real-life clinical notes. Six specialists and resident trainees from clinical microbiology or ID units conducted randomised, blinded evaluations across four key domains: factual consistency, comprehensiveness, coherence, and medical harmfulness. Results Analysis of 960 human evaluation entries by six clinicians, covering 160 AI-generated responses, showed that GPT-4.0 produced longer responses than Gemini Pro (p<0·001) and Claude 2 (p<0·001), averaging 577 ± 81·19 words. GPT-4.0 achieved significantly higher mean composite scores compared to Gemini Pro mean difference (MD)=0·2313, p=0·001 and Claude 2 (MD=0·2021, p=0·006). Specifically, GPT-4.0 outperformed Gemini Pro and Claude 2 in factual consistency (Gemini Pro, p=0·02 Claude 2, p=0·02), comprehensiveness (Gemini Pro, p=0·04; Claude 2, p=0·03), and the absence of medical harm (Gemini Pro, p=0·02; Claude 2, p=0·04). Within-group comparisons showed that specialists consistently awarded higher ratings than resident trainees across all assessed domains (p<0·001) and overall composite scores (p<0·001). Specialists were 9 times more likely to recognise responses with "Fully verified facts" and 5 times more likely to consider responses as "Harmless". However, post-hoc analysis revealed that specialists may inadvertently disregard conflicting or inaccurate information in their assessments, thereby erroneously assigning higher scores. Interpretation Clinical experience and domain expertise of individual clinicians significantly shaped the interpretation of AI-generated responses. In our analysis, we have demonstrated disconcerting human vulnerabilities in safeguarding against potentially harmful outputs. This fallibility seemed to be most apparent among experienced specialists and domain experts, revealing an unsettling paradox in the human evaluation and oversight of advanced AI systems. Stakeholders and developers must strive to control and mitigate user-specific and cognitive biases, thereby maximising the clinical impact and utility of AI technologies in healthcare delivery.