Myocardial Infarction Recognition Method Based On Deep Learning

Cardiovascular disease is one of the main diseases affecting human health worldwide.According to the latest statistics from the World Health Organization,about30 percent of the 58 million deaths in the world are caused by cardiovascular diseases.Myocardial infarction(MI)is one of the most common cardiovascular diseases in daily life.After myocardial infarction occurs,it will cause irreversible and permanent damage to myocardial cells,so early diagnosis is necessary.Electrocardiogram(ECG)is an electrical signal reflecting the activity of the heart and has been widely used in clinical detection because of its advantages(e.g.simple acquisition,low cost,easy used and no trauma).In recent years,the pattern recognition of MI by 12-lead ECG has become a hotspot.However,the existing automatic recognition methods of MI have some limitations,which are mainly reflected in that the object of model recognition is long sequence signals resulting in poor real-time recognition of MI;the insufficient generalization performance of the model leads to the low recognition accuracy of the inter-patient paradigm.The key scientific problems to overcome the limitations of MI recognition are how to fully consider the structural relationship between multiple-leads,improve the feature expression ability of the model,and how to combine the clinician’s diagnostic logic with practical applications to make the designed model interpretable.Facing the above problems,a set of deep learning theory and method simulating doctors’ analysis of ECG receptive field is proposed,which improves the feature learning efficiency and interpretability of ECG multiple-leads data structure,realizes the beat-by-beat MI recognition of the inter-patient paradigm,and breaks through the bottleneck of MI recognition.The main research contents are as follows:(1)A neural network model based on multi-branch weighted features is proposed under intra-patient and beat-by-beat recognition paradigm.Firstly,the 12-lead heartbeat samples were extracted simultaneously and separately,aiming to acquire the local features of each lead and some correlation information obtained by weighted fusion.Secondly,12-lead features were concatenated and features were continue extracted for acquiring the global features of the 12-lead and further digging out the inherent related information to accurately recognize MI.Through 5-fold cross-validation on the PTB dataset,the model achieved an average accuracy rate of99.95%,an average sensitivity of 99.97%,and an average specificity of 99.86%.(2)A neural network model based on multi-channel fusion is proposed under inter-patient and beat-by-beat sampling paradigm.By introducing the doctor’s receptive field as the guiding principle of the model framework design,a low-level neural network for extracting the shallow features of a single-lead and a high-level neural network for extracting the deep features of multiple-leads were built.Based on the general features,the temporal features are further extracted to excavate the implied causal relationship,so as to achieve the purpose of accurate recognition of MI.Through 5-fold cross-validation on the PTB dataset,the model achieved an average accuracy rate of 91.90%,an average sensitivity of 92.15%,and an average specificity of 90.63%.It can be applied to computer-aided diagnosis equipments to assist clinicians diagnosing MI in view of the excellent performance of this model.