Physiological Signal Classification And Application

Physiological signal classification in smart medical fields can help the health assessment of human body,assist doctors in disease diagnosis and even effective treatment without tedious clinical examination and treatment to a certain extent.In the process of promoting the development of smart medical care,how to achieve accurate physiological state perception is an important problem to be solved,which places a higher demand on the accuracy of intelligent algorithms at the same time.Although the application of various deep learning models has made great progress in recent years,the existing models still face some important challenges in practical application scenarios.In order to build high-accuracy classification models and effectively encode the physiological signals,this paper presents a series of novel signal classification deep learning models for related downstream applications,by conducting an in-depth and effective study on the key challenges such as direct capture of salient waves,effective fusion of the complementary features,and overcoming the subject variability.The main contribution of this paper includes:1)Propose a salient wave detection network for physiological signal classification,aiming at capturing salient waves in the raw physiological signal directly and efficiently without any preprocessing.To accurately describe salient waves,a U~2 structure based on fully convolutional neural networks is proposed.Meanwhile,a multi-scale extraction module is designed to capture multi-scale transition rules among sleep stages with the help of dilated convolutions of different sizes,solving the problem that recurrent neural networks are difficult to optimize.Experimental results on single-channel EEG datasets for sleep stage classification demonstrate the effectiveness of this model.2)Propose a physiological signal classification model for fusing the spatial-spectral-temporal features in a unified framework.To accurately model the characteristics of the physiological signals from different perspectives,a spatial-spectral-temporal 3D data representation of physiological data is constructed firstly.Subsequently,a 3D densely connected convolutional neural network with a two-stream structure for fusing features is designed.Meanwhile,a spatial-spectral-temporal attention mechanism is designed to enhance the performance of model classification to adaptively capture the local patterns.Experiments on multi-channel EEG datasets for emotion recognition verify that the model achieves excellent emotion recognition performance.3)A physiological signal classification model fusing multimodal data is proposed,aiming to fully model the heterogeneity and correlation of multimodal physiological signal.To accurately model the spatial relationship of multimodal data,a heterogeneous graph sequence construction method for multimodal data is designed to organize multimodal physiological signals,and a graph transformation network is proposed to model the heterogeneity among multimodal physiological signals.Simultaneously,a graph convolutional network is adopted to model the spatial correlation between multimodal physiological signals,and a gated recurrent unit is used to model the temporal correlation of multimodal physiological signals.Experiments on multimodal physio-logical signal datasets for emotion recognition demonstrate the effectiveness of this model on the emotion recognition task.4)A physiological temporal classification model based on domain generalization is proposed,aiming to capture rich brain topological properties while extracting subject-invariant features.To reduce the negative effect of subject variability on the generalization ability of the model,an adversarial domain generalization method is introduced to enable the model to learn depersonalized subject-invariant features.Two different brain views are constructed based on the spatial proximity and functional connectivity of the brain to model the topological relationships of the brain from different perspectives.A spatial-temporal graph convolutional network based on the attention mechanism is also designed to extract the key spatial-temporal features to serve the high accuracy classification.Experiments on two multi-channel EEG datasets for sleep stage classification in a cross-subject scenario show that the model is able to achieve high performance on the sleep stage classification task.