An Improved CNN-Transformer Hybrid Architecture for Heart Sound Classification

Abstract

This paper proposes a new framework of hyper-parameter tuning based integration of CNN and Transformer models for classification of heart sounds, which achieves state-of-the-art performance. As heart sound signals have local feature extraction ability, but CNNs are unable to capture long-range dependencies, and Transformers are too computationally extensive to apply to general use we propose an approach that captures the best of both worlds. The integrated system consists of advanced signal processing with machine learning techniques to provide accurate, clinically unqiue and extensible approach for early diagnostics of cardiac anomalies. Extensive experimental evaluations show that our approach provides a substantial performance gain over the state-of-the-art, from which a valuable mechanism for improving cardiac health monitoring and diagnosis emerges.