ChatGPT-4o能否应对烧伤?AI与临床医生在初始烧伤护理中的单盲对比基准评测
该单盲对比研究以英国烧伤协会(BBA)指南为基准,用20个烧伤情景评估ChatGPT与临床医生在五个领域的准确性。临床医生总体领域准确率高于ChatGPT(88% vs. 78%,p=0.031),疼痛缓解领域ChatGPT显著较低(30% vs. 65%,p=0.023)。ChatGPT可提供总体安全易懂的初始烧伤指导,可作为临床医生主导护理的辅助工具。
要点
- 临床医生总体领域准确率88%高于ChatGPT的78%(p=0.031)。
- 疼痛缓解准确率ChatGPT仅30%,显著低于临床医生的65%(p=0.023)。
- ChatGPT在急救和转诊领域与临床医生表现相当。
核心结论
- ChatGPT提供的初始烧伤指导总体安全且易于理解,在关键领域与临床医生相当。
- 经进一步验证并加入明确免责声明后,ChatGPT可辅助临床医生主导的护理,在无法及时获得专业意见时作为有价值的补充。
研究设计与样本
- 单盲对比研究,使用20个涵盖不同严重程度、类型和患者群体的烧伤情景。
- 由烧伤顾问医生对临床医生和ChatGPT的回答进行盲法随机评估,评估五个领域:急救、敷料、疼痛缓解、转诊和安全警告。
- 以英国烧伤协会(BBA)指南为基准,正确性按BBA推荐评分为1(正确)或0(错误)。
- 总体回答质量采用改良全球质量评分(mGQS;1-5分),≥4分视为临床可接受。
- 使用McNemar检验和Wilcoxon符号秩检验分别比较领域准确率和平均mGQS。
主要发现
- 临床医生总体领域准确率高于ChatGPT(88% vs. 78%,p=0.031)。
- 急救领域两者表现相当(85%)。
- 转诊领域两者表现相当(100%)。
- 敷料领域ChatGPT准确率略低(85% vs. 90%),安全警告领域ChatGPT略低(90% vs. 100%)。
- 疼痛缓解领域ChatGPT准确率显著低于临床医生(30% vs. 65%,p=0.023)。
关键数值与推荐
- 正确性评分:根据BBA推荐,正确为1分,错误为0分。
- mGQS评分范围1-5分,≥4分视为临床可接受。
- 总体领域准确率:临床医生88%,ChatGPT 78%(p=0.031)。
- 疼痛缓解准确率:ChatGPT 30%,临床医生65%(p=0.023)。
- 平均mGQS:临床医生4.33 ± 0.69,ChatGPT 4.15 ± 0.63(差异无统计学意义)。
临床怎么用
- ChatGPT可用于提供初始烧伤指导,尤其在急救和转诊领域与临床医生表现相当。
- 应谨慎对待ChatGPT在疼痛缓解方面的建议,其准确率显著低于临床医生。
- 在无法及时获得专业意见时,ChatGPT可作为临床医生主导护理的辅助工具,但需进一步验证并加入明确免责声明。
以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。
原文
INTRODUCTION: Artificial intelligence (AI) tools such as ChatGPT are increasingly accessed by the public for health-related advice; however, their accuracy in acute burn management remains uncertain. We aimed to assess the domain-specific accuracy and quality of ChatGPT's burn guidance for public use, compared with clinicians, using British Burn Association (BBA) guidelines as the benchmark. METHODS: A single-blinded, comparative study was conducted using 20 burn scenarios encompassing varying severities, types, and patient groups. The responses from clinicians and ChatGPT were blindly and randomly evaluated by burns Consultants across five domains: first aid, dressing, pain relief, referral, and safety warnings. Correctness was scored as 1 (correct) or 0 (incorrect) according to the BBA recommendations. Overall response quality was assessed using a modified Global Quality Score (mGQS; 1-5), with scores ≥4 considered clinically acceptable. The McNemar's and Wilcoxon signed-rank tests compared domain accuracy and mean mGQS respectively. RESULTS: Clinicians demonstrated higher overall domain accuracy compared to ChatGPT (88% vs. 78%, p = 0.031). Performance was comparable in first aid (85%) and referral (100%), with ChatGPT showing marginally lower accuracy in dressing (85% vs. 90%) and safety (90% vs. 100%). Pain relief accuracy was notably lower for ChatGPT (30% vs. 65%, p = 0.023). Mean mGQS scores were higher for clinicians (4.33 ± 0.69 vs. 4.15 ± 0.63); however, this difference was not statistically significant. CONCLUSION: ChatGPT provides generally safe and comprehensible initial burn guidance, comparable to clinicians in key domains. With further validation and inclusion of clear disclaimers, it can complement clinician-led care and serve as a valuable adjunct when timely professional input is unavailable.
引用
- Lim JY, Tan PH, Gutai FKI, et al. Can ChatGPT-4o handle the heat? Benchmarking AI against clinicians in initial burn care: A single-blinded comparative analysis. J Plast Reconstr Aesthet Surg. 2026. PMID: 41692649. doi:10.1016/j.bjps.2026.01.032
供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。