This paper proposes a novel feature representation approach for heartbeat classification using single-lead electrocardiogram (ECG) signals based on adaptive Fourier decomposition (AFD). AFD is a recently developed signal processing tool that provides useful morphological features, which are referred as AFD-derived instantaneous frequency (IF) features and differ from those provided by traditional tools. The AFD-derived IF features, together with ECG landmark features and RR interval features, are trained by a support vector machine to perform the classification. The proposed method improves the average accuracy of the feature extraction-based methods, reaching a level comparable to deep learning but with less training data, and at the same time being interpretable for the learned features. It also greatly reduces the dimension of the feature set, which is a disadvantage of the feature extraction-based methods, especially for ECG signals. To evaluate the performance, the Association for the Advancement of Medical Instrumentation standard is applied to publicly available benchmark databases, including the MIT-BIH arrhythmia and MIT-BIH supraventricular arrhythmia databases, to classify heartbeats from the single-lead ECG. The overall performance is compared to selected state-of-the-art automatic heartbeat classification algorithms, including one-lead and even several two-lead-based methods. The proposed approach achieves superior balanced performance and real-time implementation.