Generative Elastic Networks (GENs) and Application on Classification of Single-Lead Electrocardiogram

Deep neural networks have achieved significant success in various complex machine learning problems. However, their fundamental nature remains that of a black-box model, presenting substantial challenges in interpretability. This limitation significantly hampers their applicability, particularly in tasks e.g. electrocardiogram (ECG) classification. This paper proposes a white-box model named Generative Elastic Network (GEN) and the algorithms for training, validation, and testing, featuring inherent interpretability in its computational process and network architecture. Based on a strict convex optimization problem, its iterative process guarantees definite and strict convergence. The architecture of GEN is not predefined but generated elastically. The essence of GEN lies in filtering and recording key samples and their mapping processes, showcasing inherent adaptability to small and imbalanced dataset. Leveraging the MIT-BIH Arrhythmia Database, by six classes of single-lead ECG with typical shape characteristics are classified. With only 0.4% of the total data for training and 50% for validation, an average testing accuracy of 99.47% is achieved, matching the performance level of state-of-the-art deep networks.