Research On Myocardial Infarction Recognition Based On Deep Learning And Multi-source Feature Fusion

Myocardial infarction is one of the most deadly cardiovascular diseases and poses a great threat to people’s health.Therefore,early detection and diagnosis of myocardial infarction is vital for patients.ECG,as an important tool for diagnosing myocardial infarction,has the advantages of wide distribution,low cost,simplicity and convenience.In recent years,with the booming development of computer technology,the field of myocardial infarction recognization based on ECG signals has become a hot issue for research and it has important practical application.Therefore,this paper conducts an in-depth study on the recognition of myocardial infarction ECG signals.The main contents are as follows.(1)Aiming at the recognition of myocardial infarction,we proposed a detection model based on 1D-Dense Net.The peak point of R-wave was located for the preprocessed ECG signal,and 12-lead heartbeat signals of 0.8s duration were extracted as the input sample using the R-wave peak as the reference point.Then we constructed a 1D-Dense Net model.Feature reuse of myocardial infarction ECG signals was enhanced by using the interconnection between the front and back of dense blocks in 1D-Dense Net.At the same time,the data imbalance was improved using weighted cross-entropy as loss function.10-fold cross-validation of the proposed model was performed on the PTB dataset used in this paper,and the model achieved recognition accuracy of 94.49%,sensitivity of 87.38% and specificity of 97.12%.(2)Aiming at the problem of limited expressiveness of a single deep learning model,we proposed a myocardial infarction recognition algorithm based on multi-network stacking model.Firstly,short fragment ECG signals which contain more anterior-posterior correlation information and are more suitable for the multi-network stacking model were used as input samples.Then,three parallel networks,1D-Dense Net,modified 1D-Dense Net,and LSTM,were designed and trained using the stacking ensemble learning algorithm.LSTM was used to extract time-dependent features of ECG signals,while Dense Net was used to extract morphological features.The model combined the strengths of each network and achieved better results on the myocardial infarction dataset,with accuracy,sensitivity,and specificity of 96.05%,92.98%,and 97.59% respectively.(3)Aiming at the problem of lack of comprehensive and rich feature extraction in the model,we proposed a myocardial infarction recognition algorithm based on multi-source feature fusion and multi-network stacking model.Firstly,the mean amplitude spectrum feature map is extracted from the frequency domain transformation result of 12-lead ECG signals.The frequency band in which the T-wave,an important detection index of myocardial infarction,was divided into 12 frequency subbands.The remaining frequency band was divided into 12 frequency subbands.Then,the mean amplitude of each frequency subband was calculated to form a 12×24 feature map.Next,multi-network stacking model was designed for the multi-source feature fusion input,which contains five sub-networks in parallel.1D-Dense Net,the modified1D-Dense Net,and LSTM are used to extract features for the time-domain signal input,and2D-Dense Net and the modified 2D-Dense Net were used to extract features for the mean amplitude spectrum feature.Combining the features of different structures and types of networks,the recognition and classification of myocardial infarction were achieved.The algorithm obtained97.57% accuracy,95.66% sensitivity,and 98.42% specificity in the experimental dataset.