心脏骤停后患者脑水肿的脑电图自动检测
该研究为回顾性单中心队列,纳入2016—2024年124例成人心脏骤停复苏后同时接受神经影像和EEG监测的患者,52.4%发生弥漫性脑水肿。用心脏骤停后>24 h的4—8 h EEG片段训练transformer和LSTM模型检测脑水肿,最佳检测模型为4 h片段transformer(中位AUC 83.5%、准确率75.0%、敏感度80.0%、特异度70.0%);在19例水肿患者与19例匹配对照中,8 h片段预测模型中位AUC 80.0%。作者认为该法经外部验证后或可在常规EEG监测中提示 evolving 脑水肿供临床复核。
要点
- 最佳检测模型为transformer+4 h EEG片段(心脏骤停后>24 h采集):中位AUC 83.5%、准确率75.0%、敏感度80.0%、特异度70.0%。
- 影像学发现前的预测分析仅19例水肿对19例匹配对照,8 h片段模型中位AUC 80.0%、准确率70.0%、敏感度80.0%、特异度60.0%。
- 匹配因素为年龄、性别、目击骤停和EEG时机;transformer在检测已确立水肿和影像发现前识别相似EEG模式两方面均优于LSTM。
- 作者结论强调需外部验证后,该模型才可能用于常规EEG监测中提示脑水肿、辅助早期干预和家庭沟通。
适用人群与场景
- 成人心脏骤停复苏后患者(2016—2024年单中心),均因常规临床需要接受神经影像和EEG监测。
- 检测模型分析124例;预测分析为19例最终发生弥漫性脑水肿者与19例按年龄、性别、目击骤停、EEG时机匹配的无水肿对照。
诊断标准与关键阈值
- 弥漫性脑水肿由神经影像判定(原文未给出具体影像阈值)。
- EEG分析使用心脏骤停后>24 h采集的4—8 h片段;检测模型最佳为4 h片段,预测模型最佳为8 h片段。
- 原文未提及具体EEG模式、评分界值或启动/停止阈值。
处置流程与药物剂量
- 原文未提及具体处置流程或药物剂量;作者仅提出该模型经外部验证后可能用于常规EEG监测中提示脑水肿供临床复核。
红旗征与不能漏的情况
- 原文未提及具体红旗征或需立刻处理的情况。
研究设计与关键结果
- 研究类型:回顾性、单中心队列研究;样本量124例,无多中心数据。
- 检测模型(124例):最佳为transformer+4 h EEG片段,中位AUC 83.5%、准确率75.0%、敏感度80.0%、特异度70.0%。
- 预测分析(19例水肿 vs 19例匹配对照):最佳为8 h片段模型,中位AUC 80.0%、准确率70.0%、敏感度80.0%、特异度60.0%。
- transformer在检测已确立水肿和影像发现前识别相似EEG模式两方面均优于LSTM。
证据类型、等级与局限
- 证据类型:回顾性、单中心队列研究(机器学习模型开发与内部评估)。
- 局限:单中心、回顾性、样本量小(预测分析仅19对19)、无外部验证;作者明确指出需外部验证。
对床旁实践的改变
- 本研究不改变当前急诊/ICU对心脏骤停后脑水肿的常规评估流程;模型输出仅作为提示信号供临床复核,不能替代神经影像或临床判断。
- 适用人群为心脏骤停复苏后已接受常规EEG监测的成人患者;不适用于未行EEG监测或非心脏骤停人群(原文未提及这些人群)。
- 常见陷阱:将模型AUC/敏感度/特异度误当作决策阈值;预测分析样本量小且为匹配设计,结果不能外推为床旁筛查标准。
指南或共识出处
- 原文未提及具体指南或共识名称;来源期刊为Neurocrit Care。
以上提炼依据论文摘要(全文未开放获取,或许可不允许翻译转载);全文请点「阅读原文」。
英文原文(摘要)· 点击展开
OBJECTIVES: To develop an electroencephalography (EEG)-based machine learning model to identify diffuse cerebral edema in patients post-cardiac arrest and evaluate its ability to predict edema before radiographic detection.
METHODS: We performed a retrospective, single-center cohort study of adult patients resuscitated from cardiac arrest (2016-2024) who underwent neuroimaging and EEG monitoring as part of routine clinical care. Machine learning models using transformer and long short-term memory (LSTM) architectures were trained to detect diffuse cerebral edema from 4- to 8-h EEG segments obtained > 24 h after arrest. The best-performing detection model was then evaluated for its ability to predict diffuse cerebral edema using EEG segments preceding radiographic recognition in patients who ultimately developed edema, compared with matched referents without edema (matched by age, sex, witnessed arrest, and EEG timing). Model performance was assessed using area under the curve (AUC), accuracy, sensitivity, and specificity.
RESULTS: Among 124 patients in the detection model, the median age was 53 years, and 74 (59.7%) were male patients. Sixty-five patients (52.4%) developed diffuse cerebral edema. The best-performing detection model, a transformer using 4-h EEG segments, achieved strong performance (median AUC 83.5%, accuracy 75.0%, sensitivity 80.0%, specificity 70.0%). In a secondary analysis of 19 patients with diffuse cerebral edema and 19 matched referents, the top-performing prediction model used 8-h EEG segments (median AUC 80.0%, accuracy 70.0%, sensitivity 80.0%, specificity 60.0%).
CONCLUSIONS: Machine learning models applied to routine EEG data can classify EEG patterns associated with radiographic diffuse cerebral edema in cardiac arrest survivors. In this study, a transformer-based approach outperformed LSTM both for classifying established edema and for identifying similar EEG patterns before radiographic recognition. With external validation, this approach may help flag evolving cerebral edema for clinician review during standard EEG monitoring, potentially informing earlier intervention and family discussions regarding both reversible and irreversible brain injury.
参考来源
- Stafford RA, Srivastava V, Altaf AZ, et al. Automated Cerebral Edema Detection using Electroencephalography in Post-cardiac Arrest Patients. Neurocrit Care. 2026. PMID: 42806266. doi:10.1007/s12028-026-02668-z
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。