本文针对迷走神经刺激治疗药物难治性癫痫疗效个体差异极大的问题，基于患者术前的心电信号分析心率变异性指标并作为特征向量，后采用机器学习方法实现了癫痫患者术前的迷走神经刺激疗效预测，进一步通过特征选择在评估不同指标对疗效重要性的同时提高了模型预测性能，并探索了不同年龄和昼夜环境下患者术前心率变异性的差异以及对预测性能的影响。首先，介绍了迷走神经刺激疗效的影响因素以及心电信号、心率变异性和自主神经系统三者之间的关联，为了探讨临床上心率变异性对癫痫患者心脏自主神经功能的评估作用，从定义、计算方法以及生理意义等方面阐述了心率变异性的时域、频域和非线性指标。其次，分析了131例药物难治性癫痫患者的人口统计学信息、影像检查结果以及癫痫发作特征等临床资料，并基于术前心电预处理结果对比了患者分别在昼夜4小时的心率变异性及其与迷走神经刺激疗效的相关性，发现睡眠状态下与疗效显著相关的心率变异性指标远高于清醒状态，表明睡眠期间的心电记录能够提供更有效的疗效预测因子，揭示了心率变异性的昼夜节律与迷走神经刺激疗效之间的潜在关联。最后，基于单变量筛选和全变量子集迭代两种特征选择方法，分别在不同年龄的患者群体中评估了睡眠和清醒状态下心率变异性指标与迷走神经刺激疗效的相关性，根据相关性大小的排序结果在随机森林模型上实现了迷走神经刺激疗效预测，并进一步探究了不同心率变异性特征组合、年龄和昼夜状态对于疗效预测性能的影响。本文首次提出基于机器学习方法和术前心率变异性指标的迷走神经刺激疗效预测模型，有助于为术前筛选手术适应症患者提供客观依据，深入对迷走神经刺激治疗癫痫机制的理解。
Aiming at the great difference in efficacy of vagus nerve stimulation (VNS) in the treatment of drug-resistent epilepsy (DRE) , we proposed a statistical model to achieve VNS outcome prediction based on machine learning method and heart rate variability (HRV) indices that analyzed from preoperative electrocardiogram as feature vectors. Besides, feature selection was applied to evaluate the importance of different HRV indices to efficacy prediction while improving the prediction performance of the model. In the whole research process, the variance of preoperative HRV indices under different ages and circadian environments were investigated and their influence on prediction performance were further explored.Firstly, influencing factors of VNS outcome and the relationship between ECG, HRV and the autonomic nervous system were introduced. In order to state the clinical evaluation effect of the cardiac autonomic nervous function of HRV for patients with DRE, the time domain, frequency domain and non-linear indices of HRV were explained on the basis of definition, calculation method and physiological significance.Secondly, the demographic information, imaging examination results, and epileptic seizure characteristics of 131 patients with DRE were analyzed. Based on the preprocessing results of ECG, 4-h HRV and its relevance to the VNS outcome in sleep and awake state of patients were compared. It was found that in sleep state the amount of HRV indices which were significantly related to VNS outcome were much greater than that in awake state, indicating that ECG recorded during sleep could provide a more effective predictor of efficacy, and revealing the potential relationship between the circadian rhythm of HRV and the VNS outcome.Finally, through univariate filter and recursive feature elimination methods, the relevance between HRV indices and VNS outcome were evaluated in sleep and awake state of different age groups respectively, and the prediction of VNS outcome was achieved based on random forest classification model and the importance ranking results of HRV indices. The effect of combination of different top-ranked HRV indices, age and circadian status on the prediction performance were further experimented. To the best of out knowledge, it is the first study that combined machine learning method and HRV indices to achieve the VNS outcome prediction, and investigate the influence of sleep state on the prediction performance. Our statistical model would help to provide an objective basis for preoperative evaluations of patients with surgical indications, and was conducive to an in-depth understanding of the mechanism of VNS in the treatment of epilepsy.
