ElectroCardioGram (ECG) heart beat classification plays an important role in clinical diagnosis.However, there is a serious imbalance of the available data among four types of ECG, which restricts the improvement of heart beat classification performance. In order to solve this problem, a class information extracting method based on Convolutional Neural Network (CNN) was proposed. Firstly, an general CNN model based on equivalent data of four ECG types was constructed. And then based on the general CNN model, four CNN models that more effectively express the propensity information of the four heart beat categories were constructed. Finally, the outputs of the four categories of CNN models were combined to discriminate the heart beat type. The experimental results show that the average sensitivity of the proposed method is 99.68%, the average positive detection rate is 98.58%, and the comprehensive index is 99.12%; which outperform the two-stage cluster analysis method.