Prediction of Seizure via Residual Networks Based on Decision Fusion

Two seizure prediction models are built based on a decision fusion strategy and residual network by using spatial coupling features and introducing an attention mechanism. First, eight frequency bands are filtered, and the correlation matrices are computed for each frequency of eighteen channels. Second, the eight 18x18 matrices are input to the residual module for classification, and the results are concatenated to form a vector. A fully connected layer is used for decision fusion. Third, to emphasize the coupling relationship among the different frequency bands, a cubic matrix formed by the eight 18x18 matrices is inputted to an attention network, resulting in the enhanced features. A seizure prediction model is thus proposed by combining the nine decisions. The performance of the model is compared with those from state-of-the-art methods, and the sensitivity of the proposed model is improved by 4.45%.