Research On Bioelectrical Signal Recognition For Smarter Healthcare

With the aging population and the extension of human life,the risk of suffering from chronic diseases,such as cardiovascular disease,Alzheimer’s disease and stroke etc,has greatly increased in general population.Meanwhile,China’s primary medical resources are seriously deficient.A new type of smarter Healthcare system for chronic diseases,which integrates wearable’s,remote monitoring,intelligent diagnosis and services should be established to meet the development of the population and society.Electrocardiogram(ECG)and electroencephalogram(EEG)are powerful tools for monitoring,diagnosis and treatment of chronic diseases such as cardiovascular diseases,senile dementia and stroke.The ECG and EEG are non-invasive,economical,convenient and flexible,so they have an important application values in the future wearable smarter medical care.Bioelectric signals such as ECG and EEG usually have strong noises,thus containing strong randomness,nonlinearity and chaos,etc.Furthermore,their morphological characteristics shows significant variations for different patients and under different conditions.The signal processing and automatic diagnosis methods are the key techniques for wearables and intelligent medical equipments.The current recognition methods for ECG and EEG signals,especially for clinical data,still have some issues of low accuracy,low diagnostic efficiency and high variability,which seriously hinder the clinical application of the intelligent diagnosis and rehabilitation systems.To solve the above problems,the physiological mechanism of ECG and EEG signals was analyzed and the correlation between signal characteristics and disease phenomena was studied.Advanced signal processing and high efficient adaptive machine learning algorithms along with training mechanisms,novel methods such as deep neural network for real-time auxiliary diagnosis based on medical big data were proposed.Furthermore,these methods were used for the classification and recognition of ECG and EEG signals.The main innovations of the paper are as follows:1.A novel scheme that combined a frequency band selection common spatial pattern algorithm and a particle swarm optimization least squares twin support vector machine classifier for recognition of motor imagery patterns was proposed.Motor imagery EEG signal recognition is the key for the wearable rehabilitation brain-computer interface(BCI)system for patients such as,Alzheimer,stroke etc.To overcome the slow response speed and large variability in the motor imagery BCI system,an adaptive band selection common spatial pattern(CSP)feature and LS-TWIN-SVM algorithm model was proposed for rapid adaptive motor imagery EEG signal recognition based on the analysis of neurophysiological mechanism of motor imagery EEG.This algorithm can overcome the inter-individual variability by adaptively selecting the motor imagery rhythm bands of EEG signals for different patients.The LS-TWIN-SVM classification model replaces complex quadratic programming with the solution of linear equations.It has high training efficiency and its easy to implement in the hardware.Particle swarm optimization,chaotic particle swarm optimization,a genetic algorithm and a quantum genetic algorithm were compared and used to tune the hyper-parameters for the classifiers.Using the BCI Competition IV dataset,the experimental results showed that the proposed method improved the overall recognition accuracy,reduced the variability between individuals compared to the traditional support vector machines,twin support vector machines and their improved algorithms.Furthermore,it achieved a faster CPU running time for training classifiers.The experimental results under the same dataset were better than the recently published artificial intelligence based motor imagery EEG signal recognition results.2.A wavelet adaptive threshold algorithm for ECG signal denoising and Shannon energy combined Hilbert transform algorithm for QRS complex wave detection method were proposed.Electrocardiogram signal filtering and QRS complex wave detection are the preconditions for feature extraction and automatic diagnosis.A wavelet adaptive threshold filtering method was proposed to improve the signal-to-noise ratio of the signal.Meanwhile,it was not distorting the waveform morphology feature.The experimental results based on clinical ECG data showed that this method achieved better results.In order to improve the accuracy of complex wave detection,an algorithm based on the first-order difference,Shannon energy,Hilbert transform adaptive threshold was proposed to detect R-wave of ECG signals.R-wave detection was performed using MIT-BIH public dataset and clinical data respectively.The detection results showed that the proposed method was superior to the traditional ECG complex wave detection algorithm in detection accuracy and efficiency.3.A hybrid feature extraction and GA-based directed acyclic graphs LS-TWIN-SVM method for arrhythmia recognition were proposed.The LS-TWIN-SVM multi-classification strategies,such as decision tree,directed acyclic graph etc.were analyzed and their computational complexity and efficiency were compared.A time intervals and power spectrum feature extraction method was proposed.These heartbeat features were sent to a directed acyclic graph LS-TWIN-SVM model for rapid arrhythmia classification.The number of our training samples was less than 3.2%of all samples.Our proposed method achieved a high classification accuracy of 99.1403%and achieved a more rapid training time of 0.2044 s.Using the same dataset,the evaluation results showed that the proposed method outperformed multi-layer perceptions,extreme learning machines,support vector machines and several twin SVMs in terms of classification accuracy and training efficiency.4.A real-time arrhythmia classification algorithm for clinical ECG big data based on deep residual network was proposed.The ECG signals acquired by clinical electrocardiographs,especially wearable electrocardiographic devices were easily contaminated with external noise and its morphological characteristics show significant variations for different patients and under different conditions.ECG signal detection and recognition for the real-time ECG monitoring and arrhythmia diagnosis are still a challenging task.A wavelet adaptive threshold denoising combined with deep residual network scheme was proposed for arrhythmia diagnosis of clinical ECG data.ECG filtering was implemented using wavelet adaptive threshold technology.A multi-layer convolutional neural network(CNN)containing multiple residual blocks was designed,namely a deep residual network for recognition of arrhythmia signals.Using the residual block local neural network unit to construct a deep residual network,it alleviated the difficulties of deep network convergence,difficulty in tuning and so on,but also overcomed the degradation problem of the traditional convolutional neural networks(CNN)when the network depth increasing.The batch normalization of each convolution layer improved its convergence.Using the MIT-BIH database of 94091 heart beats and clinical 12-lead static ECG data,the experimental results showed that the proposed method was superior to traditional deep learning networks such as Multilayer Perceptron(MLP),LENET CNN and ALEX CNN.Under the same dataset,the classification accuracy is better than that of the currently published deep neural network algorithm.Using the clinical 12-lead ECG data,experimental results showed that the classification accuracy reached 87.3885%,which exceeded the diagnostic level of the clinical experts,so it would have high clinical application value.