Variable weight algorithm for convolutional neural networks and its applications to classification of seizure phases and types

Deep learning techniques have recently achieved impressive results and raised expectations in the domains of medical diagnosis physiological signal processing. The widely adopted methods include convolutional neural networks (CNNs) recurrent (RNNs). However, existing models possess static connection weights between layers, which might limit generalization capability classification performance as different layers are fixed after training. Furthermore, to deal with a large amount data, network sufficiently size is required. This paper proposes variable weight (VWCNNs), type structure employing dynamic instead their fully-connected layers. VWCNNs able adapt characteristics input data can be viewed an infinite number traditional, fixed-weight CNNs. We will show that proposed VWCNN outperforms conventional CNN terms accuracy, capability, robustness when inputs contaminated by noise. In this paper, applied three seizure phases (seizure-free, pre-seizure seizure) based on measured electroencephalography (EEG) data. achieve 100% test accuracy strong phases, thus potential useful tool for diagnosis. seven types seizures investigated using world’s largest open source database recordings, TUH EEG corpus. Comparisons CNNs, RNN, MobileNet, ResNet, DenseNet traditional machine including random forest, decision tree, support vector machine, K-nearest neighbours, standard networks, Naïve Bayes being conducted realistic sets. demonstrate advantages over other classifiers robustness.