Supplementary file.zip

While large language models (LLMs) have shown promise in clinical decision support, existing prompting strategies often rely on single-pass reasoning and lack mechanisms for verifying or correcting diagnostic conclusions, resulting in suboptimal performance in diagnostic tasks. To address this, we introduce a novel“Initial Diagnosis–Verification–Final Diagnosis”prompting framework that guides LLMs through a structured two-stage reasoning process.<br>We designed four verification prompting strategies and evaluated their diagnostic performance across 589 MedQA-USMLE cases and 300 NEJM Case Records, using both GPT-4o and DeepSeek-V3. Each case underwent five diagnostic repetitions, with both diagnoses and reasoning chains recorded. Our results demonstrate that the verification prompting strategy can substantially optimize LLM diagnostic performance across multiple dimensions, including improving accuracy, enhancing consistency, reducing uncertainty, correcting initial diagnostic errors, and minimizing reasoning illusions and errors, thus demonstrating promising applicability in medical diagnostics.<br>We believe this framework offers a practical and scalable approach to improving the reliability of LLM-driven medical diagnostics.