Extreme Learning Machine Based Imbalanced Active Learning For ECG Recognition

The rapid development of social economy has accelerated the pace of people’s life.The premature arrival of the aging society,the unhealthy living styles in daily life and the sudden outbreak of the Covid-19 coronavirus disease have led to a dramatic increase in chronic diseases such as cardiovascular disease.Electrocardiosignals contain important information for clinical diagnosis,which can be detected by surface Electrocardiogram(ECG)technology.Thanks to the massive advantages of current wearable,Internet of Things and artificial intelligence technologies,real-time ECG based intelligent recognition approach will improve the cognition status and provide early warning and prevention of diseases,which is of great significance for the improvement of human health.This dissertation focuses on the recognition of arrhythmia in cardiovascular diseases based on dynamic ECG signals to address the challenges such as the low generalization performance of existing recognition model,the limited learning capability of the machine learning models in the face of class imbalance,the difficulty in identifying valuable data in collected massive ECG data as well as the high costs on manual labeling.In view of the non-iterative fast training and universal approximation capabilities of Extreme Learning Machine(ELM),our study will be conducted using ELM to achieve accurate,efficient and highly generalizable arrhythmia recognition performance.The main research work is summarized as follows:(1)In view of the complicated types and manifestations of arrhythmia,a multiperspective feature set based ELM heartbeat classification approach is proposed.The approach constructs a multi-perspective heartbeat feature set,adopts the mutual-information-based feature selection strategy to select a small number but more valuable and relevant features,so as to obtain a highly distinguishable heartbeat feature set for arrhythmia.Correspondingly,an ELM-based heartbeat multi-classification approach is presented to achieve accurate and efficient heartbeat recognition.(2)To deal with the class imbalance in ECG data,an intra-class coherence based weighted kernel ELM(ICC-WKELM)imbalanced classification approach is proposed.The approach adopts the intra-class coherence(ICC)to describe the class imbalance in a fine-grained manner by modeling the spatial distribution information of samples among classes.Based on the ICC measure,a corresponding weight assignment strategy is designed and ICC-WKELM imbalance classification algorithm is presented to implement accurate imbalanced heartbeat multiclass classification among individuals.Experimental results show that the proposed approach can effectively improve the arrhythmia recognition performance.(3)To improve the personalized recognition performance of the arrhythmia recognition approaches and reduce the costs of sample labeling,an online weighted kernel ELM(OWKELM)based active learning arrhythmia classification approach is proposed.The approach consists of an active sample selection phase and a model online update phase.Active learning strategy is adopted to iteratively select a batch of samples that are representative and valuable for classifier construction from an unlabeled pool.Specifically,a margin probability and sample distribution based dual-criteria optimization objective is designed,and the problem is formulated as a 0-1 integer programming problem and addressed by semi-positive definite programming.Furthermore,for online update of the model after sample labeling,OWKELM is presented to avoid the retraining and reduce the model update time,together with a weight assignment strategy for different classes to improve the ability for imbalanced learning.Experimental results show that the proposed approach can reduce the costs for manual labeling and ensure the recognition accuracy.(4)To overcome the low efficiencies of feature extraction by manual manner and the existing deep learning models,and to avoid the high costs of labeling a large number of invalid samples in unlabeled pool,a multilayer ELM based ensemble active learning heartbeat classification approach is proposed.It obtains the automatic heartbeat deep representation by an ELM Auto Encoder(ELM-AE)based non-iterative unsupervised signal reconstruction and layer-by-layer stacking of multiple ELM-AEs.In order to reduce the unstable phenomenon caused by the random initialization of ELM hidden layer parameters,an ensemble ELM classification model is constructed and the final classification result is obtained by the decision-level fusion with soft ensemble strategy.In addition,active ensemble sample selection strategy is designed to identify valuable samples.The approach can also achieve the accurate and stable recognition performance on multi-lead ECG signals even with different waveforms,which demonstrates the universality of the approach.