Study of seizure detection via multiscale adaptive selection convolutional neural networks

To resolve the problem that the multi-scale convolutional neural network (CNN) in traditional electroencephalogram (EEG) signal classification methods for epilepsy simply concatenates features across different scales but ignores the inherent differences between these features, this study proposes an epileptic seizure detection approach based on multiscale adaptive selection convolutional neural networks (MSASCNNs). The MSASCNN model conducts feature screening on raw EEG signals via convolution kernels of diverse sizes to generate initial features, and further extracts multiscale features by constructing multi-scale adaptive selection convolutional blocks. During the feature extraction process, the model integrates an optimized attention mechanism that adaptively adjusts the weight of features from distinct scales, thus strengthening the model’s emphasis on multi-scale feature information. Additionally, the model incorporates residual connections to alleviate the issues of increased training difficulty and information attenuation as network layers become deeper. For practical epileptic seizure detection, this study integrates twin support vector machines for classification, realizing effective monitoring and diagnosis of epilepsy. The results demonstrate that the MSASCNN model delivers outstanding performance on both the dataset provided by the Epilepsy Research Center of Peking Union Medical College Hospital and the CHB-MIT dataset, with an accuracy of 96.78% and a precision of 95.35%. The MSASCNN-based epileptic seizure detection method proposed in this study provides an innovative and efficient solution for EEG signal classification and epilepsy monitoring, and offers solid technical backing for subsequent research in relevant domains.