Abstract: Electrocardiosignal can reflect human cardiac health status, and automatic analysis approach of electrocardiogram (ECG) is of great significance for the timely detection and prevention of cardiovascular diseases. An extreme learning machine (ELM) ensemble approach for heartbeat classification that fuses handcrafted features and deep features is proposed. On one hand, the manually extracted features clearly characterize the heartbeat signal: RR interval features reflect the time-domain characteristic while the wavelet coefficient features reflect the time-frequency characteristic. On the other hand, a 1D convolutional neural network (1D CNN) is designed to automatically extract deep features for heartbeat signal. These features are fused and sent to ELM for classification. Due to the unstable nature caused by the random assignment of ELM hidden layer parameters, bagging ensemble strategy is proposed to integrate multiple ELMs and hence achieves stable classification performance and good generalization ability. The proposed approach is validated on the MIT-BIH arrhythmia public dataset, where classification task on 16 heartbeat types is performed. The classification accuracy on single lead reaches 99.05%. The experimental results also show that the performance of the approach with fused features is better than those with deep feature and handcrafted features, respectively.