Research On Arrhythmia Classification Algorithm Based On Integrated Deep Learning

At present,cardiovascular disease has become the leading cause of death.As a typical cardiovascular disease,the accurate classification and detection of arrhythmia is of great significance.Electrocardiogram has become the most common basis for doctors to diagnose arrhythmia due to the non-invasive and low-cost.The number of electrocardiogram has increased sharply with the improvement of health awareness and the universal of home electrocardiograph.However,the limited medical resources cannot complete the massive diagnosis of ECG.Therefore,the automatic classification for arrhythmia has become a crucial technology.The ECG features cannot be mined completely since the traditional machine learning methods rely on the design of manual features.The large individual differences limit the generalization of the model.The above factors limit the practical application of arrhythmia classification technology.To solve the problems mentioned above,the integrated learning and semi-supervised learning are introduced to the classification of arrhythmia.This paper makes further research on the two different modes of arrhythmia classification.The main work is as follows:(1)An arrhythmia classification algorithm based on ensemble deep learning is proposed to perform the inter-patient arrhythmia classification.The features are extracted from time domain,frequency domain,geometric and State transition of ECG respectively by bi-directional long Short-Term memory network and improved transfer convolutional neural network.The features are fully extracted and the multi features are integrated to perform the joint classification of ECG,which is beneficial to improve the accuracy of the model.A weighted integration strategy is adopted considering that the final classification results vary with different models.The features of ECG are used fully and reasonably by assigning different weights to models according to the performance of during training.The proposed model is tested on the arrhythmia database of MIT-BIH and the overall accuracy has reached96.2%.(2)An arrhythmia classification algorithm based on the combination of ensemble learning and semi-supervised learning is proposed to perform patient-specific arrhythmia classification.The training data consisted of common data and a few special data comes from specific patient.The final specific model was obtained through semi-supervised training,which can learn the ECG features of a specific patient to improve the generalization of the model.The combination of network uncertainty and prediction probability was introduced to reduce the pseudo-label noise in the process of semi-supervised training,which enhanced the accuracy of pseudo-labels.The various models are integrated to further enhance the stability of the model.The proposed method was tested on the arrhythmia database of MIT-BIH and the overall accuracy had reached 98.1%.