Key result
Seven Chinese LLMs show no performance difference on AF questions but produce critical safety errors.
Why the study?
The extent to which recently launched Chinese large language models can assist atrial fibrillation patients was unknown.
Cross-Sectional
Single-blind
Certain Chinese LLMs can generate accurate answers to atrial fibrillation patient questions, but safety concerns currently preclude their clinical application.
Safety errors in Chinese LLMs for AF queries preclude clinical use; leaves open optimized models for patient education pending validation.
The first seven Chinese Large language models (LLMs) were launched to the public on August 31st,2023.However, the extent to which Chinese LLMs can assist atrial fibrillation༈AF༉patients remains unknown. We sought to assess the Chinese LLMs performance of providing responses to AF patient questions. This cross-sectional study compared seven Chinese LLM chatbots including ABAB, Baichuan, Chatglm, Doubao, Ernie bot, Sensechat and ZidongTaichu. First, cardiologists compiled a list of frequently asked questions by patients with AF. Responses from LLMs were collected. We developed a scoring system known as SCECCE, which consists of 6 aspects including safety, correctness, error, completeness, conciseness and elaboration. Each response was assessed by an expert committee according to SCECCE scoring system. A total of 231 responses were obtained. Overall, the median SCECCE score was 10[IQR, 7–10] with a mean(SD) score of 8.6(2.0). No significant statistical differences were observed in SCECCE scores among seven LLMs(p = 0.08). The maximum total SCECCE score was 330 points. Ernie bot attained the highest total score of 299 points. Doubao provided safe responses to 97% of the questions. In terms of correctness and error, the overall comparison of each group showed no statistically significant difference. Ernie bot exhibited the best performance with an accuracy rate of 87.9%. Our findings from this preliminary study suggested that certain Chinese LLMs could generate accurate and comprehensive answers to specific patient questions on atrial fibrillation. Nevertheless, this performance does not warrant clinical application, with safety being a foremost concern among the considerable challenges that lie ahead.
No takes yet. Share an insight, caveat, or question.
Liu et al. (2026) conducted a cross-sectional in Atrial fibrillation. Chinese Large Language Models (ABAB, Baichuan, Chatglm, Doubao, Ernie bot, Sensechat, ZidongTaichu) vs. Comparison among models was evaluated on Overall SCECCE score (p=0.08). Seven Chinese large language models showed no significant difference in overall performance scores when answering atrial fibrillation questions (p=0.08), though critical safety errors were observed.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: