Marking And Predicting The Probability Of ECG Segments Being Abnormal Based On Multiple Instance Learning

Objective:The incidence of cardiac diseases such as arrhythmia,myocardial ischemia and atrioventricular hypertrophy increases year by year,and the mortality also shows an increasing trend.ECG is one of the channels observing the overall condition of the heart in clinical practice,through which the information of cardiac activity can be timely and efficiently obtained and the patient's condition can be determined.The recognition and analysis of ECG by computer aided diagnosis technology is both important and necessary in clinical practice.The purpose of this study is to combine the convolution network under the framework of weakly supervised learning to automatically detect and mark the abnormal points of ECG and give the positive probability value.Methods:In this study,the ECG of 16,488 patients were selected from the ECG database of the***,including 4,001 normal cases and 12,487 abnormal cases.Abnormal ECG includes conduction block,hypertrophy,tachyarrhythmia,bradyarrhythmia and T-wave changes.In the study,the method of multiple instance learning was used.The heart beat was taken as the ‘bag',and the sampling point in it was taken as the ‘instance'.The label of abnormal heart beat was set as 1,and that of normal heart beat was set as 0.Through the classification marker built by the CNN,the coarse label of 'bag' is used to predict the fine label of 'instance' level,which used the thought of weakly supervised learning,so as to mark the probability value of the abnormal point in the heart beat.Results:For the ECG of seven subgroups of the five categories,the heatmap shows that,compared to the ECG's abnormal areas marked by the professionals manually,the marker designed in this study can mark and predict the probability of ECG areas being abnormal,and the F score of the regions is above 0.54,up to 0.83.The AUC value of the seven minor abnormal ECG test results was above 0.89,up to 0.994.Conclusion:This study combined multiple instance learning and CNN to design the model structure,training the model with the method of weakly supervised learning,and selected 7 types of abnormal ECG to test the model.The heatmap was used to visually show the comparison between the model designed in this study and the existing classical methods for labeling abnormal points of ECG.Meanwhile,manual labeling by professionals was used as a reference.The heatmap,AUC value and regional F score results all showed that this model had ECG abnoraml points labeling and probability prediction functions for the above seven ECG.