Towards Continually Evolved Heart Sound Classifiers with Class-Incremental Learning

Heart sound analysis provides a non-invasive and cost-effective method for identifying and studying various heart diseases. However, when new categories emerge in dynamic environments, traditional heart sound classification models generally require retraining and suffer from catastrophic forgetting. To address this, this study proposes a heart sound classification model via class-incremental learning (HSCIL) that integrates knowledge distillation, data replay, and dynamic network expansion techniques to reuse an existing classifier. Specifically, residual temporal convolutional networks are used as the backbone and dynamically expanded to include new classes. Additionally, MFCC features rather than raw heart sound samples are used for data replay. Knowledge distillation is also adopted to maintain consistency between outputs of the new model and old model. Finally, HSCIL is evaluated on a dataset with eight categories and compared with other four classification models. Experimental results show that HSCIL achieves 97.03% accuracy and 1.11% forgetting rate.