Research On Electrocardiogram Signal Detection And Classification Based On Convolutional Neural Network

As people pay more attention to their own health,the demand for personal health monitoring system is increasing.Given the advantages of convolutional neural network in automatic feature extraction and classification accuracy,the thesis focuses on electrocardiogram signal detection and classification by using convolutional neural network.This thesis gives a brief summary on convolutional neural network theory and its application in the fields of electrocardiogram signal detection and classification,researches the mapping algorithm of spiking convolutional neural network,researches the performance comparison between spiking convolutional neural network and regular convolutional neural network,and researches both electrocardiogram signal detection and classification methods by using convolutional neural network.The main contributions of this thesis are presented as follows.1.Cross-layer based mapping algorithm of spiking convolutional neural network.To address the inefficient problem of spiking convolutional neural network mapping onto network-on-chip,a cross-layer based neural mapping algorithm is proposed.The algorithm maps synaptic connected neurons belonging to adjacent layers into the same on-chip network node,which restricts pakcet transfer within the same node and reduces the link load of the network.In order to adapt to various input patterns,the strategy also takes input spike pattern into consideration and remap neurons for improving mapping adaptability.For the application of heartbeat classification,the average packet transfer latency and packet transfer energy cost can be reduced by 16.2%and 14.9%respectively.For this application,experimental results also show that spiking neural network cannot yet compare with convolutional neural network in the aspects of classification accuracy and efficiency.2.Two-level convolutional neural network based QRS complex detection.The existing detection methods largely depend on hand-crafted manual features.To solve the problem,a QRS complex detection using two-level convolutional neural network is proposed.A simple electrocardiogram signal preprocessing technique which only contains difference and averaging operations in temporal domain is adopted.The convolutional neural network consists of object-level and part-level convolutional neural networks for extracting different grained electrocardiogram morphological features automatically.All of the extracted morphological features are used by multi-layer forward network for QRS complex detection.The method not only achieves comparable 99.7%accuracy,but also reduces computation cost.3.Attention-based arrhythmia detection.The detection methods used in the current practice largely depend on hand-crafted manual features.In the basis of QRS complex detection,to mitigate the above problem,an attention-based method for patient-specific electrocardiogram arrhythmia detection is proposed.The preprocessing of this method includes signal resolution modification.The convolutional neural network consists of object-level and part-level convolutional neural networks for extracting different grained electrocardiogram morphological features of both a whole heartbeat and its corresponding waves automatically.In addition,the difference between adjacent RR intervals is computed as a dynamic feature.Both of the extracted morphological features and the interval difference are used by multi-layer forward network for classifying heartbeat.The method achieves 98.5%overall heartbeat classification accuracy.The proposed method acquires comparable accuracy of heartbeat classification though electrocardiogram signals are represented by lower resolution.Techniques proposed in this thesis improve the feasibility and reliability of electrocardiogram detection application using convolutional neural networks,and establish the theoretical and practical foundation for researches of light-weight individual healthcare monitoring system.