How Large Language Models Can Affect Clinical Reasoning: A Randomized Clinical Trial

Abstract

Abstract Importance LLMs have encoded a vast array of medical knowledge and are being integrated into clinical settings as decision-support tools to improve physician performance across various aspects of care. However, evidence of the impact of LLMs on the clinical reasoning of physicians remains limited. Objective To evaluate the impact of LLM on core aspects of physicians’ clinical reasoning: diagnostic reasoning, information gathering, and management reasoning in primary care scenarios. Design, Setting, and Participants We conducted three identical randomized controlled trials (RCTs) in 2024–2025 with 249 physicians in Indonesia, Kenya, and the Netherlands. Participants completed four or five clinical vignettes designed to simulate real-world primary care consultations, with half randomized to have access to ChatGPT-4o. Main Outcomes and Measures Physician quality of care was evaluated using a rubric based on evidence-based clinical guidelines, scored across nine steps of the clinical reasoning process. Primary outcomes were quality scores for diagnostic reasoning, information gathering, and management. Secondary outcomes were quality per answer, number of answers, and less obvious answers. Results Access to LLMs enhanced information gathering and management reasoning across all countries. Physicians who were assigned the LLM achieved significantly better quality-of-care scores in diagnostic steps in Indonesia ( b =7.9%, CI: 4.0% to 11.8%, p <.001) and Kenya ( b = 15.1%, CI: 10.2% to 19.9%, p <.001) but not the Netherlands ( b =1.4%, CI: −1.6% to 4.4%, p =1.00). Physicians with LLM access also performed better in investigative steps, in Indonesia ( b =10.7%, CI: 4.2% to 17.1%, p =.004), Kenya ( b =17.1%, CI: 10.3% to 23.9%, p <.001), and the Netherlands ( b =11.9%, CI: 7.7% to 16.1%, p <.001). We also found LLM access affected physicians’ scores in management steps (Indonesia: b =15.7%, CI: 8.6% to 22.9%, p <.001; Kenya: b =27.3%, CI: 19.9% to 34.7% p <.001; the Netherlands: b =12.3%, CI: 7.1% to 17.5%, p <.001). We found that LLM access was less useful in management reasoning for more cognitively demanding cases compared to standard patient cases in Indonesia ( b =-14.1%, CI: −21.4% to −6.8%, p <.001) and Kenya ( b =-12.1%, CI: −19.6% to −4.6%, p =.006). Conclusions and Relevance In this cross-country randomized control trial, we assessed that access to an LLM had significant positive effects on physicians’ clinical reasoning. The effects we found are promising for the further roll-out of LLMs to supplement physicians in their care tasks. They also suggest that the extent to which LLMs can supplement physicians is context dependent. Key Points Question To what extent do large language models (LLMs) increase physicians’ quality of diagnostic reasoning, information gathering and management reasoning? Findings In a randomized clinical trial including 249 physicians in Indonesia, Kenya, and the Netherlands, access to an LLM significantly enhanced clinical reasoning performance in information gathering and management reasoning across all countries, and diagnostic reasoning in Kenya and Indonesia. Meaning This study shows that the use of an LLM can enhance clinical reasoning of physicians. Further research is needed to effectively understand the augmentation of physician clinical practice.

Ähnliche Arbeiten