Prompt-Induced Diagnostic Bias in Large Language Model Classification of Echocardiography Reports

This study evaluated whether embedding diagnostic expectations within prompts biases large language model (LLM) classification of left ventricular ejection fraction (LVEF) from echocardiography reports, and whether a prompt-level mitigation strategy can counteract this effect. GPT-5 was evaluated under 1 baseline, 3 bias-injected, and 1 explicit instruction-based mitigation prompt condition across 1,500 structured reports from a single institution dataset. Bias prompts significantly altered LVEF classifications across all 3 classes, with shifts directionally consistent with the referenced category. The instruction-based mitigation strategy restored overall accuracy near baseline in all 3 conditions. In conclusion, prompt-induced bias poses a meaningful risk to diagnostic classification accuracy; however, prompt-level safeguards may support reliable LLM deployment in clinical settings.