Seizure type classification algorithm based on multi-dimensional brain network feature selection

Different seizure types exhibit varying levels of redundant brain network features. In order to select features that contribute significantly to model performance, an algorithm for seizure type classification based on multi-dimensional brain network feature selection (MBNFS) is proposed. Firstly, wavelet packet transform (WPT) is applied to all channels of the electroencephalogram (EEG) signals and all the sub-band signals obtained are used as network nodes. Secondly, mutual information (MI) between different nodes is calculated to get MI matrix as edge weights. Next, we design a new brain network set construction method based on the leave-one-out method and MI matrix to formulate a set of brain networks. After that, network features are extracted through nodes and edges and further evaluated for their contributions using random forest (RF). Then, based on the leave-one-out method and a matrix where the main diagonal is 0 and the remaining elements are 1, the contributions of channels, sub-bands and network features are deduced. Leveraging these three-dimensional contributions, a multi-dimensional iterative selection is conducted, focusing on selecting the dimensional individuals whose contributions exceed the manually set threshold. Finally, the RF model is trained with the refined feature set to classify seizure types, yielding test set detection accuracy, specificity, sensitivity, F1 score and kappa values of 99.86 %, 99.86 %, 99.86 %, 0.9986 and 0.9984, respectively.