Abstract: Non-invasively reconstructing the cardiac TransMembrane Potentials（TMP） from Body Surface Potentials（BSP） can be acted as a regression problem with multi input and multi output. Data driven models based machine learning is an effective to approach the regression problem. Here, deep learning architecture is implemented by using Convolution Neural Network（CNN）, run on Caffe framework. Moreover, based on the realistic heart-torso models, ECGSim software is used to simulate the data on the ventricular activation of Kent’s bundle syndrome for training and testing regression model. The experimental results show that compared with the Extreme Learning Machine（ELM） and the kernelized extreme learning machine, the CNN method can perform better regression ability in terms of the TMP reconstruction accuracy.

Key words: Convolution Neural Network（CNN）, Extreme Learning Machine（ELM）, ECG inverse problem, cardiac transmembrane potentials

摘要： 基于横跨膜电位分布的心电逆问题研究，即从身体表面电位无创重建心脏跨膜电位，可视为一种多输入多输出的回归问题（亦即多个体表电位分布输入重构多个心脏跨膜电位分布输出），而基于数据驱动的机器学习模型是解决回归问题的一种有效手段。通过使用深度卷动神经网络（CNN）构建深度学习模型，使用Caffe框架训练神经网络；此外，基于真实的心脏模型，使用ECGSim软件仿真了肯特束综合症心室激活情况的数据，用于训练和测试回归模型。实验结果表明，与极限学习机（ELM）和核化的极限学习机相比，CNN方法在心脏跨膜电位重构方面有更高的精度和泛化性能。

关键词: 卷积神经网络（CNN）, 极限学习机（ELM）, 心电逆问题, 心肌跨膜电位