Research On Dynamic Graph Mining Method For Physiological Signal Sequence

Time series analysis is a central theme in statistical analysis and a powerful method for characterizing biological,medical,and economic data and discovering the origins of its underlying dynamics.Physiological signal mining is an important application field of time series analysis.Physiological signals can reflect various functions of the human body,including mood,fatigue,and health status.Common physiological signals include electrocardiogram(ECG),electroencephalogram(EEG),respiratory rhythm,electromyography(EMG),galvanic skin response(GSR),blood pressure,etc.Obtaining and mining a variety of physiological signals is of great significance for computer-aided diagnosis of human physiological diseases.Disease diagnosis and prediction with physiological signals mainly faces three challenges.First is the problem of signal noise interference and strong randomness.The human body naturally generates physiological signals which are highly random and non-stationary.These signals are inevitably disturbed by the human body,instruments,or surrounding environment during recording.Therefore,finding an effective method to reduce the interference of noise and extracting effective features has always been a research hotspot.Second is the comprehensive mining of multi-channel signals and the correlation between patient signals.Most existing studies only focus on analyzing data from a single channel independently,and do not exploit the implicit associations between patient signals.It is found that the patients with the same disease always have similar physiological signal patterns,although their physiological signals are not the same.So,using these patterns to get better prediction results is also a problem worth studying.Third,the spatial location connection between signal channels is not equivalent to the functional connection.Close spatial relationships cannot guarantee close functional relationships,which are very important for discriminative physiological signal feature extraction.Therefore,it is necessary to construct the functional relationship between the signal channels dynamically.It is unreasonable to predetermine the functional relationship between the channels according to the spatial positions of the channels in the past.Aiming at the above problems,taking ECG and EEG as the starting points,this thesis studies dynamic graph mining methods for physiological signal sequence,and detects human physiological diseases.The main innovations of the thesis are as follows:(1)A single-channel signal sequence mining method based on weighted multi-scale limited penetrable visibility graph is proposed to solve the problem of signal noise interference and strong randomness.The thesis proposes a Weighted Multi-Scale Limited Penetrable Visibility Graph(WMS-LPVG)model,which dynamically constructs WMS-LPVG according to different ECG signal sequences,processes single-channel ECG signals,and tests atrial fibrillations.First,the ECG is weighted and coarse-grained;Second,the coarse-grained sequence is dynamically transformed into a complex network by the WMS-LPVG method,which reduces the interference of noise;Third,we propose a new network feature — local efficiency entropy,and combine it with other complex network features and original ECG signal features to form multimodal features which are more effective feature representations;Finally,we adopt the e Xtreme Gradient Boosting(XGboost)classifier to detect atrial fibrillations.(2)A multi-channel signal sequence mining method based on the maximization of graph mutual information is proposed to solve the problem of comprehensively processing multi-channel signals and the relationship between patient signals.Based on the above results,we study the multi-channel EEG signals.The thesis introduces the correlation between patient signals,combine with graph mutual information maximization,pre-train Graph Convolutional Neural Networks(GCN),and propose the GCNs-MI model to detect depression patients.First,the multi-channel EEG signals of each sample are sequentially spiced into a long signal sequence as a network node,and the Pearson Correlation Coefficient(PCC)between the spliced signals is calculated to form a graph network.Second,each channel is divided into six bands,the average power of each band is calculated,and they are stitched together in order as node features.Again,we pre-train a three-layer GCN model,fix the first two layers of the GCN,and fine-tune the output layer(third layer)of the model using a contrastive learning method of graph mutual information maximization.(3)A method for mining the relationship between signal channels based on a dynamic graph convolutional neural network is proposed to obtain a stronger signal feature representation and further explore the internal relationship between physiological signal channels.Based on the above achievements,The thesis designs a Dynamic Graph Convolutional Neural Network(D-GCN)model,using each channel as a node of the brain network,and learning the adjacency matrix in a dynamic way to describe the relationship between nodes.That is,the data in the adjacency matrix is adaptively updated as the model parameters change during the training process.Compared with the GCN method,the adjacency matrix learned by D-GCN is more effective,which captures the dynamic internal relationship of the signal channels,which can improve the recognition effect of the disease.This thesis mainly studies two physiological signals,ECG and EEG,and conducts experiments with two datasets.One is the 2017 Cardiology Challenge dataset,a short single-lead ECG dataset for atrial fibrillation classification.The other is a full-lead IEEE Healthcom2020/MODMA for depressive disorder analysis.We utilize multiple indicators to evaluate the model performance.The experimental results show that our proposed dynamic graph mining methods has better results than the state-of-the-art results.