Lin Wentao, Yang Danni, Chen Chen, Zhou Yuanfeng, Chen Wei, Wang Yalin
School of Information Science and Technology, Fudan University, Shanghai, China.
School of Materials Science and Engineering, Lanzhou University of Technology, Lanzhou, China.
CNS Neurosci Ther. 2025 Jan;31(1):e70196. doi: 10.1111/cns.70196.
Drug-refractory epilepsy (DRE) refers to the failure of controlling seizures with adequate trials of two tolerated and appropriately chosen anti-seizure medications (ASMs). For patients with DRE, surgical intervention becomes the most effective and viable treatment, but its success rate is unsatisfactory at only approximately 50%. Predicting surgical outcomes in advance can provide additional guidance to clinicians. Despite the high accuracy of invasive methods, they inevitably carry the risk of post-operative infection and complications. Herein, to noninvasively predict surgical outcomes of DRE, we propose the "source causal connectivity" framework.
In this framework, sLORETA, an EEG source imaging technique, was first used to inversely reconstruct intracranial neuronal electrical activity. Then, full convergent cross mapping (FCCM), a robust causal measure was introduced to calculate the causal connectivity between remodeled neuronal signals within epileptogenic zones (EZs). After that, statistical tests were performed to find out if there was a significant difference between the successful and failed surgical groups. Finally, a model for surgical outcome prediction was developed by combining causal network features with machine learning.
A total of 39 seizures with 205 ictal EEG segments were included in this prospective study. Experimental results exhibit that source causal connectivity in α-frequency band (8~13 Hz) gains the most significant differences between the surgical success and failure groups, with a p-value of 5.00e-05 and Cohen's d effect size of 0.68. All machine learning models can achieve an average accuracy of higher than 85%. Among them, the SVM classifier with Gaussian kernel function and Bayesian optimization demonstrates the highest accuracy of 90.73%, with a PPV of 87.91%, an NPV of 92.98%, a sensitivity of 90.91%, a specificity of 90.60%, and an F1-score of 89.39%.
Our results demonstrate that the source causal network of EZ is a reliable biomarker for predicting DRE surgical outcomes. The findings promote noninvasive precision medicine for DRE.
药物难治性癫痫（DRE）是指在充分试用两种耐受性良好且选择恰当的抗癫痫药物（ASMs）后仍无法控制癫痫发作。对于DRE患者，手术干预成为最有效且可行的治疗方法，但其成功率仅约为50%，并不理想。提前预测手术结果可为临床医生提供更多指导。尽管侵入性方法准确性高，但不可避免地存在术后感染和并发症风险。在此，为了无创预测DRE的手术结果，我们提出了“源因果连接性”框架。
在此框架中，首先使用脑电（EEG）源成像技术sLORETA反向重建颅内神经元电活动。然后，引入全收敛交叉映射（FCCM）这一稳健的因果度量方法，计算致痫区（EZs）内重塑的神经元信号之间的因果连接性。之后，进行统计检验以找出手术成功组和失败组之间是否存在显著差异。最后，通过将因果网络特征与机器学习相结合，开发出手术结果预测模型。
这项前瞻性研究共纳入39次癫痫发作及205段发作期EEG。实验结果表明，α频段（8~13Hz）的源因果连接性在手术成功组和失败组之间差异最为显著，p值为5.00e - 05，科恩d效应量为0.68。所有机器学习模型的平均准确率均高于85%。其中，具有高斯核函数和贝叶斯优化的支持向量机（SVM）分类器准确率最高，为90.73%，阳性预测值（PPV）为87.91%，阴性预测值（NPV）为92.98%，灵敏度为90.91%，特异度为90.60%，F1分数为89.39%。
我们的结果表明，EZ的源因果网络是预测DRE手术结果的可靠生物标志物。这些发现推动了DRE的无创精准医学发展。
