ECG Abnormal Detection Method Based On QRS Morphology Recognition And Label Co-occurrence

Electrocardiogram(ECG),as a noninvasive detection method,is widely used in early clinical precaution and diagnosis of heart disease.However,manual recognition of ECG abnormalities requires professional doctors,which is inefficient and costly,so automatic ECG abnormality diagnosis has received widespread attention.Automatic ECG abnormality detection usually includes four steps: pre-processing ECG signals,identifying waveforms,extracting ECG features,and model prediction.The main technical difficulties lie in:(1)QRS complex is the most significant waveform in ECG,and its accurate identification can help the subsequent detection of P and T waves and the extraction of ECG features,but the current QRS morphology recognition methods either detect too few abnormalities or are parameter-sensitive;(2)ECG abnormality detection is a multi-label classification problem.Due to the complex correlation between different abnormalities,the combination of abnormal labels increases exponentially,resulting in too large search space.The main research efforts to address the above issues are as follows.(1)A Sliding Window Voting strategy based on Hidden Markov Model is proposed to automatically identify QRS morphology.Firstly,each QRS complex is divided into four sections,and a sliding window is set for each section to extract samples;Secondly,the waveforms of each section are regarded as states,and the centers of sample clusters act as observations to construct a state-constrained Hidden Marko Model;Finally,we construct multiple candidate samples for the predicted complex,predict each candidate sample’s waveform through hidden Markov,and vote among all the predicted waveforms to determine the target waveforms of the complex.(2)An ensemble model of ECG abnormal detection based on label co-occurrence is proposed.Firstly,we use chi-square test to find isolated labels that are independent from all other labels and treat them separately;Secondly,a weighted undirected graph is used to describe the correlation of all non-isolated labels,with vertices indicating labels and edges representing the label co-occurrence.Prim algorithm is then used to generate the maximum cost tree;Finally,for each label pair of parent-child nodes in the maximum tree,we collect the positive and negative samples to train the binary classification model,and predict the result according to ensemble technique.The main contributions are as follows: On the one hand,we propose the voting strategy based on Hidden Markov model,which captures the local features of waveforms using sliding-windows and describes the overall characteristics using Hidden Markov.It can accurately identify target QRS morphologies;On the other hand,we propose the ensemble ECG abnormal detection algorithm based on label co-occurrence,which takes advantages of both the simultaneous presence and simultaneous absence of labels to mine the correlations between abnormal labels.The experimental results demonstrate the effectiveness and superiority of the method.