FPGA Accelerating And Deep Learning Based Arrhythmia Classification Research

With the increasing pace of people’s lives and the steep increase of people’s mental stress,induced cardiac like diseases have become common diseases that pose a significant threat to human health.Based on extrapolation from the Chinese cardiovascular disease report 2018,there are nearly 290 million existing patients.Electrocardiogram(ECG)is the most prevalent examination way,doctors can preliminarily grasp the basic physical conditions of patients according to the results of ECG,diagnose arrhythmia,coronary heart disease,hypertension and other diseases.Traditional ECG automatic classification studies,however,have focused on the classification of beat types.Although highly accurate on the test set,the actual clinical effect is poor and the diagnosis still needs to be made manually by physicians.In this paper,a deep learning based automatic classification method for arrhythmia,combined with an FPGA(field programmed gate array)hardware acceleration scheme,can quickly and effectively address the screening difficulties and low real-time performance caused by the design of human features in traditional classification methods.The main research content of this paper includes:(1)ECG signal basis and preprocessing.The sources of signal noise in ECG and the common filter noise method are analyzed,aiming at the common baseline drift of ECG signal,using wavelet decomposition and reconstruction,to filter out the effects of baseline drift.A preprocessing scheme of the electrocardiographic signal is proposed,as well as a method for beat delineation,normalization,to reduce perturbations during electrocardiographic signal acquisition and improve the accuracy of classification.(2)A network based on bidirectional gating loop structure combined with attention mechanism(Bi GRU-Attention)was designed to classify arrhythmias.A Bi GRUAttention hybrid model was proposed and constructed to autonomously learn the spatial timely sequence characteristics of ECG signals by taking advantage of the memory function of Bi GRU structure and the concentrating characteristics of attention mechanism.A software and hardware synergistic acceleration scheme of an FPGA based arrhythmia classification model was designed,and a hardware module against a Bi GRU-Attention network was designed based on the hardware characteristics of the FPGA to accelerate the network training process.(3)Verification of the Bi GRU-Attention mixed model algorithm validity using the MIT-BIH arrhythmia database implemented the AAMI(Association for the advancement of medical instrumentation)recommended classification of the five arrhythmias as N,V,s,F,Q.Comparing the inter patient scheme with other literature methods,the sensitivities of N,V,s and F in the test set were 98.28%,97.48%,83.65%,66.10%,respectively,for positive positivity,98.20%,96.82%,94.10%,80.41%,respectively,which indicated that the model algorithm proposed in this paper had some advantages.Finally,FPGA is utilized to operate in parallel to verify the rationality of the accelerated optimization design of the FPGA based arrhythmia classification algorithm by analyzing the acceleration ratio and acceleration effect.