Electrocardiogram(ECG) classification is a complex pattern recognition problem. At present, most of the ECG classification methods based on different machine learning model had achieved a high classification accuracy, but the learning efficiency was low. Therefore, a fast ECG learning algorithm was necessary. In this paper, a method of extreme learning machine was presented, which mapped the original feature space into Hilbert space with different kernel functions and made the ECG date in high dimensional space linearly separable. At last, the experimental verification was carried on MIT-BIH standard library. The results show that the proposed method has higher accuracy and faster learning speed than existing methods, which may be a potential tool for detection and analysis of clinical dynamic electrocardiogram and personalized real-time ECG monitoring.