人工智能与人类专家:心脏骤停后基于心电图的闭塞性心肌梗死检测
该单中心研究纳入97例心脏骤停后复苏成功的成人患者,对恢复自主循环后心电图采用四种方法判读:人类专家、已验证的深度神经网络Queen of Hearts(QoH)以及两个基于大语言模型的AI聊天机器人(ChatGPT和EKG Analyst)。以冠状动脉造影确定的急性冠状动脉闭塞(ACO,TIMI 0)和闭塞性心肌梗死(OMI,TIMI 0-2或TIMI 3+峰值肌钙蛋白)为主要结局。QoH诊断ACO的AUROC最高(0.846),人类专家次之(0.735),两个AI聊天机器人最低(ChatGPT 0.456;EKG Analyst 0.474);诊断OMI时QoH同样最高(0.745),人类专家0.635,AI聊天机器人最低(ChatGPT 0.495;EKG Analyst 0.626)。两个AI聊天机器人敏感性极高(ACO 100%;OMI 98.36%)但特异性极低,QoH和人类专家敏感性/特异性分布更均衡。作者认为通用大语言模型因无差别过度诊断不适合心电图判读,而QoH等专用工具可能提供互补价值。
要点
- 在97例心脏骤停后复苏患者中,QoH诊断ACO(TIMI 0)的AUROC为0.846(95%CI 0.752-0.939),高于人类专家的0.735(95%CI 0.622-0.848)。
- QoH诊断OMI(TIMI 0-2或TIMI 3+峰值肌钙蛋白)的AUROC为0.745(95%CI 0.647-0.843),人类专家为0.635(95%CI 0.515-0.755)。
- ChatGPT诊断ACO的AUROC仅0.456(95%CI 0.319-0.592),EKG Analyst为0.474(95%CI 0.346-0.603),均低于QoH和人类专家。
- 两个AI聊天机器人敏感性极高(ACO 100%;OMI 98.36%)但特异性极低,呈现无差别过度诊断。
- 作者结论:通用大语言模型因过度诊断不适合心电图判读,QoH等专用工具可能提供互补价值。
适用人群与场景
- 单中心纳入97例成人心脏骤停后复苏成功(ROSC)患者,对其ROSC后心电图进行判读分析。
诊断标准与关键阈值
- ACO定义为冠状动脉造影显示TIMI 0血流。
- OMI定义为TIMI 0-2或TIMI 3伴峰值肌钙蛋白升高。
研究设计与关键结果
- 研究类型为单中心诊断准确性研究,样本量97例成人心脏骤停后复苏成功患者,比较人类专家、QoH、ChatGPT和EKG Analyst四种方法判读ROSC后心电图。
- 主要终点ACO(TIMI 0)的AUROC:QoH 0.846(95%CI 0.752-0.939),人类专家0.735(95%CI 0.622-0.848),ChatGPT 0.456(95%CI 0.319-0.592),EKG Analyst 0.474(95%CI 0.346-0.603)。
- 主要终点OMI(TIMI 0-2或TIMI 3+峰值肌钙蛋白)的AUROC:QoH 0.745(95%CI 0.647-0.843),人类专家0.635(95%CI 0.515-0.755),ChatGPT 0.495(95%CI 0.376-0.614),EKG Analyst 0.626(95%CI 0.508-0.743)。
- 阈值依赖性性能指标显示两个AI聊天机器人敏感性高(ACO 100%;OMI 98.36%)但特异性极低;QoH和人类专家敏感性/特异性分布更均衡。
证据类型、等级与局限
- 证据类型为单中心诊断准确性研究。
- 局限:原文未报告局限性。
对床旁实践的改变
- 作者认为,通用大语言模型(ChatGPT、EKG Analyst)因无差别过度诊断,不适合用于心脏骤停后心电图判读。
- 作者认为,QoH等专用工具可能提供互补价值。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「处置流程与药物剂量」「红旗征与不能漏的情况」「指南或共识出处」,故未列出这些栏目;全文请点「阅读原文」。
英文原文(摘要)· 点击展开
BACKGROUND/AIM: Accurate electrocardiogram (ECG) interpretation after cardiac arrest is essential for identifying occlusive myocardial infarction (OMI), but post-resuscitation artifacts make this challenging. While artificial intelligence (AI) offers promising support, its diagnostic performance in this critical setting remains uncertain.
METHODS: This single-centre study included 97 adult patients resuscitated from cardiac arrest (CA). Post-return of spontaneous circulation (ROSC) ECG were evaluated by four methods: human experts (HE), a validated deep neural network [Queen of Hearts (QoH)], and two large language model (LLM)-based AI Chatbots (AI-CB) - ChatGPT and EKG Analyst. Primary outcome was AUROC for presence and probability of OMI and acute coronary occlusion (ACO), determined by coronary angiography.
RESULTS: For ACO (TIMI 0), QoH yielded highest AUROC (0.846 [0.752-0.939]), followed by HE (0.735 [0.622-0.848]). Both AI-CB resulted in lowest AUROC (ChatGPT: 0.456 [0.319-0.592]; EKG Analyst: 0.474 [0.346-0.603]). For OMI (TIMI 0-2 or TIMI 3 + peak-troponin), QoH again achieved highest AUROC (0.745 [0.647-0.843]), followed by HE (0.635 [0.515-0.755]), AI-CB were lowest again (ChatGPT: 0.495 [0.376-0.614]; EKG Analyst: 0.626 [0.508-0.743]). Threshold-dependent performance metrics revealed high sensitivity (ACO: 100 %; OMI: 98.36 %) for both AI-CB, at the cost of minimal specificity. QoH and HE showed more even distributions of sensitivity/specificity.
CONCLUSION: QoH, despite operating without awareness of the CA-setting and thus likely at a relative disadvantage, and HE showed robust diagnostic accuracy. Due to undifferentiated overdiagnosis, general LLMs remain unsuitable for ECG interpretation. Domain-specific tools like QoH may offer complementary value.
参考来源
- Silwanis C, Eder J, Fellner A, et al. Artificial intelligence versus human expertise: ECG-based detection of occlusive myocardial infarction after cardiac arrest. Resuscitation. 2025. PMID: 41270992. doi:10.1016/j.resuscitation.2025.110905
供执业医师抢救与诊疗参考,不替代临床判断;剂量请结合体重、肝肾功能和本院药品规格复核,以指南原文为准。