Research On Detection Method Of Atrial Fibrillation Based On Volume Pulse Wave

Atrial Fibrillation(AF)is a serious and common arrhythmia disease,which may cause stroke and impair heart function when the patient develops.Electrocardiogram(ECG)is the gold standard for detecting AF.However,ECG has the shortcomings of short monitoring period and troublesome acquisition,and it is difficult to detect paroxysmal AF through ECG.In contrast,Photoplethysmography(PPG)is easy to obtain and suitable for long-term monitoring.Therefore,it is of great significance to study related methods for the automatic detection of atrial fibrillation based on PPG signals and using deep learning technology.Based on the different preprocessing of the collected PPG signals,this paper builds two deep learning network models with different inputs.Both models are composed of convolutional neural networks(CNN)and long short-term memory(LSTM)hybrid construction.The main research and work contents of this paper are as follows:(1)Collect the PPG needed for research through three public databases,slice them into signal segments with a fixed length of 10 seconds,and label the PPG signal segments according to the synchronized ECG.(2)Expand the PPG signal segment by using a data enhancement method to solve the problem of imbalance of different labeled data.For one-dimensional PPG,discrete wavelet transform(DWT)is used to decompose,reconstruct,filter and denoise the signal segment.For the two-dimensional PPG data,the one-dimensional PPG signal segment is converted into a two-dimensional time-frequency chromatogram and used as the input data through Continuous Wavelet Transform(CWT).This conversion can eliminate the need for noise filtering steps at the same time.The information in the original signal is retained to a greater extent.(3)Propose a 1D-CNN-LSTM model with one-dimensional data as input,and use preprocessed one-dimensional PPG as model input for AF/non AF classification based on three public databases.The experimental results show that the classification of the model is accuracy(ACC),sensitivity(Sen),specificity(Spe),and F1 Score are 96.91%,96.95%,96.85%,96.81%,respectively.The area under curve(AUC)is 0.9928.(4)A 2D-CNN-LSTM model with two-dimensional data input is proposed.The model uses time-frequency chromatograms as input for AF/non AF classification.The experimental verification is also carried out on the basis of three public databases.The results show The model has high classification accuracy,sensitivity,specificity,and F1 Score,which are98.21%,98.00%,98.07%,98.13%,and AUC is 0.9959.(5)In view of the proposed network model,three well-known deep learning models,VGGNet,Google Net and Res Net-50,were compared to compare the advantages of the model in performance performance,and to verify the impact of network architecture superparametrics on the training process and performance.In summary,the two models proposed in this paper,as well as the pretreatment of PPG,can better realize the automatic detection of AF,not only can help doctors diagnose AF,but also provides a way to detect AF through portable wearable devices.