Atrial Fibrillation Recognition Based On Deep Learning Feature Selection And Semi-Supervised Learning

In recent years,the number of cases of cardiovascular disease has continued to increase with the development of population aging,and the mortality rate caused by cardiovascular disease has remained high.Paroxysmal atrial fibrillation,as a common arrhythmic disorder,is associated with high mortality and morbidity from many cardiovascular diseases,and its diagnosis requires wearable long-term ECG.In order to reduce the workload of segment-by-segment discrimination and diagnosis of Holter artificially,the machine learning algorithms for automatic diagnosis of atrial fibrillation fragments are of great significance.Faced with the current situation that labeling data is difficult to obtain in clinical,semi-supervised learning has advantages.Based on the high classification accuracy of deep learning networks and the interpretability of traditional machine learning,this thesis proposed an atrial fibrillation recognition method for ECG signals based on the semi-supervised learning classifier Laplacian Support Vector Machine(LapSVM)using deep learning features and manual features.The main work of the thesis is as follows:(1)Data pre-processing on ECG signals.Wavelet decomposition is used to remove high-frequency noise,and mean filtering is used to remove baseline drift.Manual features extraction includes that three R-R interval time domain features and Entropy-based AF were extracted to reflect the change of the R-R interval of AF signal,and the QRS-zero-setting sample entropy after amplitude normalization was extracted to reflect the change of f wave in atrial fibrillation.(2)An atrial fibrillation recognition framework of CNN-LSTM combined with semi-supervised model LapSVM was constructed for ECG signal.The model architecture of Convolutional Neural Network(CNN)is improved,and the Long Short Term Memory Network(LSTM)is combined to extract the deep learning features of short-term ECG samples.LapSVM was used to classify the deep learning features which were extracted from short-time ECG samples and combined with manual features.Through experiments on the AFDB database,the classification effect of the proposed method with other methods was compared.By performing cross-dataset experiments on the self-collected dataset,the generalization performance of the proposed method was tested.By experimenting with less labeled data,the classification recognition ability of the semi-supervised algorithm LapSVM when there are fewer labeled samples is verified.(3)An atrial fibrillation recognition network based on genetic algorithm used to dimensionality reduction was built.The deep learning features extracted by CNN-LSTM were selected to dimensionality reduction using genetic algorithm.The changes of feature collection for the recognition of AF before and after feature dimensionality reduction were compared.Search for the best feature subset corresponding to different original feature dimensions to find the optimal deep learning feature selection scheme.Study the change in demand for deep learning feature dimensions as the amount of labeled data decreases.