Labeling Ecg Abnormal Segments Based On Multi-instance Multilabel Learning

Objective: The incidence and mortality of cardiovascular diseases are increasing year by year,which has become a major public health problem.These disorders,especially arrhythmias,present morphologic changes on the electrocardiogram,and the evaluation of these low-amplitude and subtle morphologic changes by specialists are time-consuming and challenging.For abnormal ECG segments detection,at present,it mainly gives a detection result according to the fixed length signal,but not indicates the specific abnormal beats in the signal.In order to achieve heartbeats classification,the signal is firstly located,segmented,and then diagnosed.This process requires a mass of beat labels marked by doctors.In this research,for heterogeneously segmented ECG signals of variable lengths,abnormal heartbeats detection algorithms are put forward based on multi-instance multilabel framework.The multi-label annotation training model of signals can be used to achieve multi-label classification for unknown signal segments and locate abnormal heartbeats within the signals corresponding to different diseases,assisting clinicians to make more accurate disease diagnosis.Methods: Based on the multi-instance multi-label learning framework,this study further combines the convolutional neural network design MLNN+ marker and U-net model design MIMLU-net marker.The two markers use the aggregation function to complete the probabilistic combination of the instances,which enables the multi-label classification of unknown signals and the abnormal heartbeats positioning within the signal.Five subsets of A-E containing different numbers of heartbeats were intercepted from the MIT-BIH arrhythmia database according to eight categories.The MLNN+ model uses the four data subsets A-D for training and testing,while the MIMLU-net model uses the five data subsets A-E for training and testing.This paper evaluates the results of the models using different evaluation metrics at both bag and instance levels.Results: At the bag level,both MLNN+ and MIMLU-net models have improved performance compared with the contrast algorithms,and all the results are better than the contrast algorithms in all metrics.MLNN+ model achieves the best results for all indicators on four data subsets containing fixed heartbeats.MIMLU-net model realizes multi-label classification of variable-length signals.At the instance level,the MLNN+ model also achieved the best performance on the datasets containing a fixed heartbeat number,and the model combined with the Max function was more suitable for long-term ECG segments and produced fewer false negative errors than the ISR function.Models proposed in this paper can achieve high signal classification and abnormalities heartbeats identification accuracy except for fusion of ventricular and normal beats and atrial premature beats.Conclusions: For ECG signals with multi-label,this paper proposes new models combined with multi-instance multi-label learning and deep neural networks to realize the classification of unknown signals and the positioning of abnormal heartbeats within the signals.The class activation maps generated by the prediction faithfully reflect the status of each heartbeat and help clinical doctors to make more accurate disease determinations.