Quality Evaluation of Large Language Model-Assisted Generation of Initial Senior Physician Ward Round Records for Acute Poisoning Patients: A Cross-Sectional Study (Preprint)

BACKGROUND Large language models (LLMs) have demonstrated potential in certain medical text generation tasks. Senior physician ward round records are critical medical documents in the diagnostic and treatment process, with their quality reflecting the accuracy and continuity of clinical decision-making. The quality of senior physician rounds notes generated by large language models (LLMs) in the specific context of acute poisoning remains unclear. OBJECTIVE This study focused on acute poisoning patients and systematically compared the differences in medical record writing quality among the domestic model DeepSeek, the international model ChatGPT, and human physicians across multiple dimensions to clarify their clinical application value. METHODS A retrospective analysis was conducted on 256 poisoning cases from the emergency department ward of Taihe Hospital, Hubei University of Medicine. The study utilized DeepSeek V3.2 Exp and ChatGPT-5.1 to generate senior physician ward round records based on standardized prompts, which were then compared with actual records from the original medical charts. Three senior emergency physicians conducted blinded evaluations, scoring overall quality across five dimensions using a Likert 1-5 scale: case characteristics, current diagnosis, differential diagnosis, treatment plan, and prognosis assessment. Error frequencies in medical records were also documented. Additionally, potential harm was assessed using a modified AHRQ harm scale. RESULTS The DeepSeek model achieved the highest mean total score (24.14 ± 0.90), significantly higher than ChatGPT (23.30±1.42, P< 0.001) and the physician group (23.86±0.86, P=0.015). DeepSeek demonstrated optimal performance in both differential diagnosis (4.98±0.10) and prognosis assessment (4.73±0.42), while matching the physician group's performance (4.96± 0.15) in case characteristics (4.90±0.23) (P>0.05). Subgroup analysis showed that for drug poisoning and pesticide poisoning, DeepSeek's mean total scores (24.23±0.75, 23.92 ±1.14) were significantly higher than ChatGPT (23.34±1.33, 22.78 ± 1.33; both P