Research On ECG Heartbeat Classification Algorithm And Energy-Efficient Architecture

Cardiac disease is a common chronic disease that threats human health.With the development of artificial intelligence technology,wearable devices for cardiac monitoring can automatically analyze ECG signals and make corresponding diagnoses,but high diagnostic accuracy and energy-efficiency are the major challenges.This research uses heartbeat classification as a specific application scenario,aiming to improve the classification performance and energy efficiency of the heartbeat classification model.From the two aspects of algorithms and circuits,focuse on solving problems of performance deviation of the classification model in the imbalanced learning problem,the limitation of classification algorithms based on manual feature extraction,and excessive energy consumption of classification models.The main contributions of the research include:1.Research of heartbeat classification model based on resampling technology.The distribution of heartbeat types is imbalanced in the heartbeat classification application,which results in classification performance deviation.To mitigate the problem,a data processing algorithm based on SVM resampling technology is proposed.The algorithm extracts necessary samples and eliminates the negative impact of noise by using SVM to undersample the training dataset repeatedly.It can minimize the loss of information and improve the quality of data to the maximum extent.Combined with the characteristics of the data distribution,the weighted oversampling technique is adopted to solve the problem of imbalanced data distribution.Test and comparison are carried out based on MIT-BIH database.Experimental results show that the classification model has advantages of classification performance and the classification accuracy can be 95.8%,78.7%,and 89.7%for heartbeat class N,V’,and S,respectively.2.Research of heartbeat classification model with feature self-learning.Considering the limitation of classification algorithms based on manual feature extraction,RNN with a new structure is proposed for variable-length label sequence identification.The network processes ECG signals by using time window to avoid the defect of R wave detection.By using the time-sequential characteristic of long-sequence ECG signals,LSTM layers are adopted to automatically extract corresponding features.Then combine CTC to output a variable-length label sequence,which enhances the applicability and practicability of the classification model.The proposed network is evaluated on the MIT-BIH arrhythmia database and experimental results show that the RNN classification model can achieve 89.6%heartbeat classification accuracy and 89.5%heartbeat sequence accuracy.3.Research of energy-efficient heartbeat classification architecture.Implement energy-efficient architectures of heartbeat classification models based on artificial feature extraction and feature self-learning and analyse their energy consumption in heartbeat classification applications.By studying the distribution of heartbeats in ECG signals,a cascaded structure with preclassification function is used to optimize the heartbeat classification model based on artificial feature extraction to reduce unnecessary calculation and memory access,which saves average energy consumption by 55.1%.Through a deep compression method,the heartbeat classification model based on feature self-learning realizes the transformation from dense network to sparse network,reducing parameters by 74.6%and relieving the storage burden.The coding scheme is improved by using the weight sharing characteristics after quantization to enhance the reusability of parameters,which can reduce multiplication by 50.3%and memory access by 37.0%.The classification energy consumption can be further saved by 33.9%.The proposed techniques are of research significance and practical application value to the improvement of heartbeat classification performance and energy efficiency of wearable cardiac monitoring devices.