基于稀疏表示的两阶段脑电癫痫波检测算法研究

Automatic Detection of Epileptiform Transients in EEG by a Two-stage Algorithm Based on Sparse Representation

脑电癫痫特征波的自动检测具有重要的临床应用价值,本研究提出一种基于自适应预测滤波与稀疏表示的两阶段癫痫特征波检测算法。第一阶段,使用自适应预测滤波器粗检出有嫌疑的癫痫波时段,在保证检测正确率的同时,减少数据量,提高后续处理效率;第二阶段,先以高斯函数及其一、二阶导数为原子的生成函数构建一个冗余多成分字典,再应用匹配追踪算法仅获取可疑波段在此字典下的稀疏表示(自适应参数化表示),原子的结构参数能够准确度量瞬时波形的多种形态结构特征如宽度、幅度、锐度等,进而提出基于形态结构匹配的检测算法,对预检输出的可疑时段进行鉴别分类。检测结果表明该算法针对临床癫痫EEG的检测率为93.3%,正确率为88.5%,相应的漏检率为6.7%,误检率为11.5%。

Automatic detection of epileptiform transients has important application in clinic diagnosis. A two-stage procedure was proposed to automatically detect EEG spikes, based on sparse representation of EEG signals and adaptive prediction filter. In the first stage, an adaptive autoregressive prediction filter was used as a pre-detector to detect all the possible epileptiform transients. This pre-detection not only reduced the computation time but also increased the overall detection performance of the procedure. In the second stage, Gaussian function and its first and second derivations were used as generating functions to construct a redundant mnlti-component dictionary. Subsequently the adaptive time-frequency parametrization of EEG signals were obtained using matching pursuit method in our dictionary, providing description of signal＇s morphological structures. Furthermore, a detection algorithm based on morphological structure match was proposed as a post-detector to classify the possible epileptiform transients. The experimental results indicated that the proposed detection technique yielded sensitivity of 93.3 % and selectivity of 88.5% based on clinical EEGs, thus maintaining lost detection rate of 6.7% and false detection rate of 11.5%.