The Study Of ECG Inverse Problem By Using Deep Learning Method

Electrocardiogram is currently one of the main ways for diagnosing heart disease.However,electrocardiogram can not provide intuitive cardiac electrical imaging.Cardiac electrical imaging is an effective way to solve this problem.According to the geometry information and body surface potential distribution,cardiac electrical function imaging can provide the electrical activity information of the heart,which is also acted as inverse ECG problem.Inverse ECG problem can reconstruct cardiac transmembrane potentials(TMP)from body surface potentials(BSP)noninvasively,which can better reflect the details of cardiac electrical activity.Inverse ECG problem can be regarded as a nonlinear regression problem with multi-input and multi-output(i.e.reconstructing the multiple cardiac transmembrane potentials distribution output from the input of body surface potentials distribution through the regression model).And data-driven machine learning algorithm is an effective method to solve the nonlinear regression problem.As a kind of powerful machine learning algorithm,deep learning is the frontier and hot direction in the field of artificial intelligence.Using deep learning method to solve nonlinear regression tasks can also achieve better prediction accuracy and generalization capability.In this paper,convolution neural network(CNN)is used to build a deep learning model.And we use the Caffe framework is to train the model in parallel with GPUs.In this paper,CNN method is used to construct a deep learning model.CNN has powerful abilities of feature learning and expressing.It can achieve a good approximation to the transmembrane potential of the heart.The experimental results show that compared with the ELM and ELM-kernel methods,the proposed CNN-based method can solve the inverse ECG problem with higher prediction accuracy and better generalization capability,which can reconstruct more accurate information of cardiac electrical activities.When it comes to solving the ECG inverse problem,although using conventional CNN model can achieve good accuracy,the convergence of training process will be slow.And it will lead to large memory cost and high complexity of computational time.By using kernel principal component analysis(KPCA)method for a substantial reduction of the sample data’s dimensional pretreatment and the Maxout sparse strategy being considered in CNN network construction to build a KPCA-CNN deep neural network,this paper further proposed a novel way to ease the difficulty in CNN training under the limited hardware capabilities.Compared with the conventional CNN model,the KPCA-CNN model can greatly reduce the computational time complexity and contain more sparsity.It will improve the efficiency of feature extraction,achieve fast convergence and save on computing cost.Experimental results validate that the KPCA-CNN model has better convergence performance and computational efficiency than the conventional CNN model.