Research On Seizure Detection By Combining Deep Features And Hand-crafted Features

Since the newborns have just separated from their mother,their brains are extremely vulnerable,and seizure is the most common brain abnormality.In the clinical application,electroencephalography(EEG)is a convenient and low cost diagnostic tool widely used to identify seizures.However,observing EEG records is an extremely dull and time-consuming work,and even if let a professional doctor analyze an EEG signal,he may also make some mistakes.To address this issue,the paper proposes a seizure detection model that combines deep features and hand-crafted features.The main work and contributions of the paper are as follows.Firstly,to take full advantage of human experience in seizure detection,some conventional hand-crafted features are extracted,including the Hjorth,the kurtosis and the skewness of the signals.In addition,in view of the lack of image information,the features of the amplitude-frequency contour map are designed.When extracting these features,the EEG signals are first reconstructed,and then converted into the amplitudefrequency contour map.Finally,the features are extracted from the converted image.The experimental results show that the extracted conventional hand-crafted features can better reflect the characteristics of seizures,and the features extracted from the EEG amplitude-frequency contour map can also dig out some information that cannot be represented by other conventional features.Secondly,to take full advantage of the feature extraction ability of deep learning,a novel neural network structure is proposed to extract features from original EEG signals.The neural network mainly includes a channel dropout module,a multi-scale convolutional layer,an attention layer,a multi-stream bidirectional RNN layer and a fully connected layer.The channel dropout module randomly dropouts some EEG signal channels during the model training to enhance the robustness of the model.The multi-scale convolution layer uses different sizes of convolution kernels to extract multi-scale features.The attention layer captures the importance of different channels.The multi-stream bidirectional RNN layer further extracts the temporal information.The fully connected layer integrates all the information and outputs the prediction results.The experimental results show that the neural network proposed in the paper is better than the normal neural network.Third,in order to simultaneously utilize human intelligence and deep learning approaches,a feature fusion framework combining hand-crafted features and deep features is proposed.The framework includes the module of feature extraction and transformation,and the enhanced DNN.In the module of feature extraction and transformation,it first extracts some hand-crafted features,and then uses GBDT to convert the hand-crafted features into category features.In the enhanced DNN,a normal neural network is used to extract deep features from original EEG signals,and the embedding layers are used to read the category features.Finally,a fully connected layer is applied to combine the embedding features and the deep features.The experimental results show that the proposed feature fusion framework is excellent in seizure detection,and the framework is also applicable to other similar fields.