机器学习利用脑电图预测儿科ICU急性脑损伤
本研究为前瞻性研究,纳入2021年7月至2023年1月PICU收治的201例患儿,排除既往脑损伤和已知癫痫者。收集人口学、入院原因、EEG特征和神经影像,采用K近邻、逻辑回归、支持向量机(线性及RBF核)、随机森林、梯度提升和Adaboost等机器学习方法,二分类预测有无急性脑损伤,多分类区分无、单侧和双侧损伤。结果显示51%患儿存在急性脑损伤,最常见的EEG背景为慢/紊乱(73%),24%有癫痫样放电,13%有脑电发作。二分类中SVM-RBF的AUROC为0.95%、测试准确率0.90%;多分类中逻辑回归测试准确率最高为0.77%。
要点
- 在PICU患儿中,联合临床与EEG特征的SVM-RBF模型预测急性脑损伤的二分类AUROC为0.95%、测试准确率0.90%。
- 多分类区分无、单侧和双侧脑损伤时,逻辑回归测试准确率最高为0.77%。
- 研究人群急性脑损伤发生率为51%,最常见EEG背景为慢/紊乱(73%),24%有癫痫样放电,13%有脑电发作。
- 研究排除既往脑损伤和已知癫痫患儿,模型不适用于该人群。
- 机器学习模型需前瞻性外部验证,目前不能替代神经影像或临床判断。
适用人群与场景
- PICU收治的患儿,年龄中位数3.5岁(IQR 1-11.6岁),42%为女性。
- 排除既往存在脑损伤和已知癫痫的患者。
诊断标准与关键阈值
- 二分类:急性脑损伤(class 1) vs 无脑损伤(class 2)。
- 多分类:无损伤(class 1)、单侧损伤(class 2)、双侧损伤(class 3)。
- 模型性能评价指标:精确率、召回率、AUROC。
处置流程与药物剂量
- 原文未提及
红旗征与不能漏的情况
- 原文未提及
研究设计与关键结果
- 前瞻性研究,纳入201例PICU患儿(2021年7月至2023年1月),排除既往脑损伤和已知癫痫者。
- 急性脑损伤检出率为51%;最常见EEG背景为慢/紊乱(73%),24%有癫痫样放电,13%有脑电发作。
- 二分类预测急性脑损伤:SVM-RBF的AUROC为0.95%、测试准确率为0.90%。
- 多分类区分无、单侧和双侧脑损伤:逻辑回归测试准确率最高为0.77%。
证据类型、等级与局限
- 证据类型:前瞻性研究。
- 局限:单中心、样本量201例、未报告外部验证;原文未提及证据等级。
对床旁实践的改变
- 原文未提及
指南或共识出处
- 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
BACKGROUND AND OBJECTIVES: The objective of this study was to apply and compare machine learning techniques using clinical and electroencephalography (EEG) background features to predict acute cerebral injury in the pediatric intensive care unit (PICU).
METHODS: This was a prospective study of patients admitted to the PICU from July 2021 to January 2023. We excluded patients with preexisting cerebral injury and known epilepsy. Clinical variables collected included demographics, reason for admission, electroencephalography (EEG) features, and neuroimaging findings. Machine learning techniques to predict cerebral injury included K-nearest neighbor, logistic regression (LR), support vector machines (SVM), random forest, gradient boosting, and Adaboost classifiers. For SVM, we used the linear and radial basis function kernels (RBF). We designed a binary classification for predicting acute brain injury (class 1) versus no brain injury (class 2) and a multiclass classification to better identify details of the brain injury: no injury (class 1), unilateral injury (class 2), and bilateral injury (class 3). The performance was evaluated in terms of precision, recall, and area under the receiver operating characteristic (AUROC) curve.
RESULTS: Of 201 patients, 42% were female, and the median age was 3.5 years (IQR 1-11.6 years). The most common EEG background was slow/disorganized (73%). In total, 24% of patients had epileptiform discharges and 13% of patients had electrographic seizures. Acute cerebral injury was detected in 51%. For the binary classification to detect the presence or absence of acute cerebral injury, SVM with RBF was most accurate with AUROC of 0.95% and test accuracy of 0.90% for predicting cerebral injury. For the multiclass classification to detect unilateral versus bilateral cerebral injury, LR had the highest test accuracy with 0.77%.
DISCUSSION: Machine learning can be employed using clinical and EEG features to accurately predict acute cerebral injury in critically ill children.
参考来源
- Sansevere AJ, Anwar SM, Keenan JS, et al. Machine Learning Using Electroencephalography Predicts Acute Cerebral Injury in the Pediatric ICU. Neurocrit Care. 2026. PMID: 42768269. doi:10.1007/s12028-026-02596-y
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。