Zhu Meitong, Xu Meng, Gao Meng, Yu Rui, Bin Guangyu
Department of Biomedical Engineering, College of Chemistry and Life Science, Beijing University of Technology, Beijing 100124, China.
College of Computer Science, Beijing University of Technology, Beijing 100124, China.
Sensors (Basel). 2025 Apr 7;25(7):2332. doi: 10.3390/s25072332.
Clinically, patients in a coma after cardiac arrest are given the prognosis of "neurological recovery" to minimize discrepancies in opinions and reduce judgment errors. This study aimed to analyze the background patterns of electroencephalogram (EEG) signals from such patients to identify the key indicators for assessing the prognosis after coma.
Standard machine learning models were applied sequentially as feature selectors and filters. CatBoost demonstrated superior performance as a classification method compared to other approaches. In addition, Shapley additive explanation (SHAP) values were utilized to rank and analyze the importance of the features.
Our results indicated that the three different EEG features helped achieve a fivefold cross-validation receiver-operating characteristic (ROC) of 0.87. Our evaluation revealed that functional connectivity features contribute the most to classification at 70%. Among these, low-frequency long-distance functional connectivity (45%) was associated with a poor prognosis, whereas high-frequency short-distance functional connectivity (25%) was linked with a good prognosis. Burst suppression ratio is 20%, concentrated in the left frontal-temporal and right occipital-temporal regions at high thresholds (10/15 mV), demonstrating its strong discriminative power.
Our research identifies key electroencephalographic (EEG) biomarkers, including low-frequency connectivity and burst suppression thresholds, to improve early and objective prognosis assessments. By integrating machine learning (ML) algorithms, such as Gradient Boosting Models and Support Vector Machines, with SHAP-based feature visualization, robust screening methods were applied to ensure the reliability of predictions. These findings provide a clinically actionable framework for advancing neurological prognosis and optimizing patient care.
临床上，心脏骤停后昏迷的患者被给予“神经功能恢复”的预后判断，以尽量减少意见分歧并减少判断错误。本研究旨在分析此类患者脑电图（EEG）信号的背景模式，以确定昏迷后评估预后的关键指标。
依次应用标准机器学习模型作为特征选择器和过滤器。与其他方法相比，CatBoost作为一种分类方法表现出卓越的性能。此外，利用Shapley加法解释（SHAP）值对特征的重要性进行排序和分析。
我们的结果表明，三种不同的脑电图特征有助于实现五重交叉验证接收器操作特征（ROC）为0.87。我们的评估显示，功能连接特征对分类的贡献最大，为70%。其中，低频长距离功能连接（45%）与预后不良相关，而高频短距离功能连接（25%）与预后良好相关。爆发抑制率为20%，在高阈值（10/15 mV）下集中在左额颞叶和右枕颞叶区域，显示出其强大的判别能力。
我们的研究确定了关键的脑电图（EEG）生物标志物，包括低频连接和爆发抑制阈值，以改善早期和客观的预后评估。通过将梯度提升模型和支持向量机等机器学习（ML）算法与基于SHAP的特征可视化相结合，应用了稳健的筛选方法以确保预测的可靠性。这些发现为推进神经功能预后和优化患者护理提供了一个临床可操作的框架。
