当机器出错时:机器学习模型在急救电话中识别院外心脏骤停的真假预测特征
该研究在哥本哈根急救医疗服务(EMS)2018年9月至2019年12月的临床实践中,测试了一个用于识别院外心脏骤停(OHCA)急救电话的机器学习模型。模型共处理169,236通电话,其中5,811通(3.4%)被怀疑为OHCA,敏感度84.5%、特异度97.1%。未识别出的OHCA中,31%的来电者描述了与OHCA完全不同的疾病,28%报告患者呼吸正常,23%存在语言障碍;假阳性中28%报告患者无意识,13%为已死亡患者被误判为OHCA。作者认为需要持续优化语言模型以提高OHCA预测的敏感度和特异度。
要点
- 模型处理169,236通电话,怀疑OHCA 5,811通(3.4%),敏感度84.5%、特异度97.1%。
- 未识别OHCA中31%来电者描述为完全不同的疾病,28%报告呼吸正常,23%存在语言障碍。
- 假阳性OHCA中28%报告患者无意识,13%为已死亡患者被误判为OHCA。
- 作者认为需持续优化语言模型以改善OHCA预测的敏感度和特异度。
适用人群与场景
- 哥本哈根EMS在2018年9月至2019年12月期间接到的所有急救电话。
- 经丹麦心脏骤停登记处核实的OHCA患者。
研究设计与关键结果
- 研究类型:在临床实践中测试机器学习模型的描述性分析;电话与调度数据库链接,经丹麦心脏骤停登记处核实OHCA,由受训审核员评估假阴性和假阳性预测的电话。
- 模型共处理169,236通电话,怀疑5,811通(3.4%)为OHCA,敏感度84.5%、特异度97.1%。
- 未识别OHCA中,31%的来电者描述了与OHCA完全不同的疾病,28%报告患者呼吸正常,23%存在语言障碍。
- 假阳性OHCA中,28%报告患者无意识,13%的假阳性为机器学习模型将已死亡且存在不可逆死亡征象患者的电话解读为OHCA。
证据类型、等级与局限
- 证据类型:在临床实践中测试机器学习模型的描述性分析。
- 局限:原文未报告局限性。
对床旁实践的改变
- 作者认为需要持续优化语言模型,以提高机器学习模型在急救电话中识别OHCA的敏感度和特异度。
提炼说明
- 本条依据论文摘要整理;摘要未涉及「诊断标准与关键阈值」「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: A machine-learning model trained to recognize emergency calls regarding Out-of-Hospital Cardiac Arrest (OHCA) was tested in clinical practice at Copenhagen Emergency Medical Services (EMS) from September 2018 to December 2019. We aimed to investigate emergency call characteristics where the machine-learning model failed to recognize OHCA or misinterpreted a call as being OHCA.
METHODS: All emergency calls were linked to the dispatch database and verified OHCAs were identified by linkage to the Danish Cardiac Arrest Registry. Calls with either false negative or false positive predictions of OHCA were evaluated by trained auditors. Descriptive analyses were performed with absolute numbers and percentages reported.
RESULTS: The machine-learning model processed 169,236 calls to Copenhagen EMS and suspected 5,811 (3.4%) of the calls as OHCA, resulting in 84.5% sensitivity and 97.1% specificity. Among OHCAs not recognised by machine-learning model, a condition completely different from OHCA was presented by caller in 31% of the cases. In 28% of unrecognised calls, patient was reported breathing normally, and language barriers were identified in 23% of the cases. Among falsely suspected OHCA, the patient was reported unconscious in 28% of the cases, and in 13% of the false positive cases the machine-learning model interpreted calls regarding dead patients with irreversible signs of death as OHCA.
CONCLUSION: Continuous optimization of the language model is needed to improve the prediction of OHCA and thereby improve sensitivity and specificity of the machine-learning model on recognising OHCA in emergency telephone calls.
参考来源
- Nikolaj Blomberg S, Jensen TW, Porsborg Andersen M, et al. When the machine is wrong. Characteristics of true and false predictions of Out-of-Hospital Cardiac arrests in emergency calls using a machine-learning model. Resuscitation. 2023. PMID: 36634755. doi:10.1016/j.resuscitation.2023.109689
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。