Research On Neural Network Based ECG Classification Algorithm And Energy-efficient Architecture

Cardiovascular arrhythmia is a common disease that often occurs suddenly and becomes life-threating.Electrocardiogram based wearable arrhythmia monitoring devices record and analyze the physiological signals in real-time,which greatly improve the life quality and survival rate of patients.However,the achieving of high-accuracy algorithm and energy-efficient architecture design is a great challenge to be solved.This thesis takes automatic heartbeat classification as a specific application scenario,aiming at improving the diagnostic accuracy and energy efficiency of portable electrocardiogram monitoring devices.This research presents a valuable exploration of neural network based long-term arrhythmia monitoring design from both algorithms and circuits.The main contributions of this thesis include:1.Research of heartbeat classification model by integrating BLSTM and CNN.This thesis presents a lightweight automated ECG classification algorithm with high recognition accuracy,which integrates both BLSTM and CNN.BLSTM and CNN can extract the temporal features from sequential heartbeats and the detailed morphological characteristics from current heartbeat,respectively,making our classification algorithm achieve more significant feature extraction.To further improve the classification performance,adaptive segmentation and re-sampling are applied to align the heartbeats of different patients with various heart rates.In addition,a novel clustering method is proposed to recognize the most representative patterns among the common training data.All the designs make the network obtain higher heartbeat classification accuracy with more lightweight model scale,which is promisingly applicable to embedded devices for automatic and accurate arrhythmia monitoring.2.Research of energy-efficient heartbeat classification architecture based on network compression.Neural network algorithms perform better than traditional machine learning methods on ECG classification.However,they are severely restricted by the complex computations and intensive memory consumption on embedded devices.In this thesis,we propose an energy-efficient ECG processor by using an innovative quantified sparse matrix encoding after network compression.The proposed ECG classification architecture with novel encoding format takes full advantage of the data locality of our neural network model,which removes the majority of weight fetching and matrix-vector multiplications.The peak energy efficiency of the ECG processor reaches 3.23 GOPS/mW,which is state-of-the-art among existing neural network accelerators in the field of edge computing.3.Research of energy-efficient heartbeat classification architecture based on ECG similarity.As the ECG waveforms are nearly periodic signals in most situations,the continuous heartbeats show significant similarity between each other,thus results in a large amount of same intermediate results during network processing.To reduce energy consumption of the ECG classification algorithm,the similarity between consecutive heartbeats is exploited to achieve a high degree of computation reuse in hardware architecture.Based on the simulation,this processor achieves an average energy efficiency of 3.52 GOPS/mW under 1.1 V supply at 100 MHz frequency.Compared to the design without computation reuse,the proposed processor provides a speedup ratio of 2.58 and an energy dissipation reduction of 61.27% per classification.The proposed techniques are valuable on both scientific perspectives and practical applications,which are quite useful for the improvement of classification performance and energy efficiency on portable ECG monitoring devices.