Recently more and more ambulatory electrocardiograms (AECG) were put in clinical usage, casting heavy diagnosis workloads on cardiologists. Aiming to reduce cardiologists’ workloads as much as possible, in this paper, we propose an active learning framework, called Active-Selection-Random-Selection (ASRS), which combines convolutional neural network (CNN) as the classifier, and margin sampling as active selection criterion. ASRS trains a model from scratch without the need of a large database, and thus will be very useful in practice. The experimental results show that with SVEB and VEB’s F-1 scores as the evaluation metrics, using ASRS can save more than 90% of the total workloads compared to the standard practice of using the first five minutes of data for annotation.