深度集成网络在癫痫预测中的应用研究

DEEP ENSEMBLE MODEL IN PREDICTION OF EPILEPSY

针对传统方法在脑电信号上捕获癫痫发作时无法有效利用全部信息的问题,提出深度集成网络的方法从脑电信号中学习多维特征以有效检测癫痫发作。对卷积网络的平移不变性和计算效率进行修复提升后纵向集成循环神经网络以同时捕获频域、空域和时域信息。基于深度集成网络进行患者特异性和跨患者模型的训练,两种模型分别达到了平均98.22%和95.65%的灵敏度以及0.09次/h和0.3次/h的误诊率,优于传统模型的结果。实验表明深度集成网络有能力学习癫痫发作的一般不变表示。

Aiming at the problem that traditional methods cannot effectively utilize all the information when capturing seizures on EEG signals, this paper proposes deep ensemble networks to capture multi-dimensional features from EEG signals, so as to effectively detect seizures. After the translation invariance and computational efficiency of convolution network were repaired and improved, the cyclic neural network was vertically integrated to capture the information in frequency domain, spatial domain and time domain at the same time. The patient-specific and cross-patient model was trained based on deep integration network. The two models achieve an average sensitivity of 98.22% and 95.65%, and a misdiagnosis rate of 0.09/h and 0.3/h, which perform better than the benchmark model. It demonstrates that deep ensemble network has the ability to learn the general invariant representation of seizures.