Wavelet multiresolution analysis and dyadic scalogram for detection of epileptiform paroxysms in electroencephalographic signals

Abstract Introduction Early detection of epilepsy by the review of large electroencephalographic (EEG) recordings is very stressful, time-consuming, and subjective for neurologists. Several automatic seizure detection systems have been proposed in the literature to solve this problem. Methods This study proposes two complementary wavelet-based approaches for detecting epileptiform paroxysms in EEG signals. First methodology applied the wavelet multiresolution analysis (MRA) to filter non-epileptiform activity in long-term EEG. Second methodology used the wavelet dyadic scalogram to analyze which scales were related to the epileptiform paroxysms. For tests, 65 wavelet functions were selected between daubechies, biorthogonal, symlets, reverse biorthogonal and coiflet wavelet families in order to evaluate their performance. Results For MRA, it was noted a better performance by using the db4 function, by reaching 48.30% of energy with 8 wavelet coefficients, 0.717658 of correlation and 36.799 of root mean square error (RMSE). For wavelet dyadic scalograms, were chosen bior3.9 and rbio1.5 functions, by reaching 77.98% of sensitivity, 94.08% of specificity, 87.87% of efficiency and 0.9613 of area under the curve (AUC value) by using bior3.9. Conclusion The presented approaches are highly complementary for a whole automatic seizure detection system by using the MRA as pre-processing stage to filter non-epileptiform activity, and wavelet dyadic scalogram for extracting desired features from filtered EEG signals.

摘要 引言
通过审阅大量脑电图（electroencephalogram, EEG）记录以实现癫痫早期筛查，对神经内科医师而言工作量繁重、耗时极长且结果具有主观性。现有文献已提出多种自动癫痫发作检测系统以解决该问题。

方法
本研究提出两种互补的基于小波的方法，用于检测脑电图信号中的痫样阵发性放电。第一种方法采用小波多分辨率分析（wavelet multiresolution analysis, MRA）对长时程脑电图中的非痫样活动进行滤波；第二种方法利用小波二进尺度图，分析与痫样阵发性放电相关的尺度分量。为评估各方法的性能，本研究从多贝西、双正交、对称、反双正交及科伊夫莱特小波家族中选取共65个小波函数开展测试。

结果
针对小波多分辨率分析方法，采用db4小波函数可获得更优性能：其8个小波分量的能量占比达48.30%，相关系数为0.717658，均方根误差（root mean square error, RMSE）为36.799。针对小波二进尺度图方法，选取了bior3.9与rbio1.5小波函数；其中采用bior3.9时，模型的灵敏度达77.98%、特异度达94.08%、效能达87.87%，曲线下面积（area under the curve, AUC）为0.9613。

结论
本研究所提出的两种方法具有高度互补性：将小波多分辨率分析作为预处理阶段以滤除非痫样活动，再通过小波二进尺度图从滤波后的脑电图信号中提取所需特征，可构建完整的自动癫痫发作检测系统。