Compressive sensing has been widely used in wearable health monitoring system for arrhythmia classification of electrocardiogram (ECG) signals since it has the ability to process signal with low energy consumption. However, the ECG data to be classified is not only enormous in amount, but also complex in structure. Therefore, solutions are needed to deal with such huge and diverse datasets. In this paper, the authors develop a novel ECG classification scheme using deep learning in wearable health monitoring system. The proposed method first compresses and reconstructs the ECG signals based on compressive sensing, and then develops a deep learning approach to perform heartbeat classification over the recovered ECG signals. Simulation experiments have shown that the proposed scheme can achieve classification accuracy of up to 97.06% with low energy consumption in sensor node as compared to benchmarking methods.