Talk · IEEE International Conference on Smart Information Systems and Technologies · May 14, 2025

Deep Learning with Multi-Head Attention for Respiratory Sound Analysis

Dina Koishiyeva, Jeong Won Kang, Assel Mukasheva

Abstract

This paper explores the performances of deep learning algorithms architectures for classification of breath sounds in automated diagnosis of breathing diseases. The investigation encompasses three convolutional neural frameworks: a baseline convolutional network, a bidirectional recurrent convolutional architecture, and an augmented version incorporating a combination multi-head self-attention mechanism. The systems were developed and validated using a publicly available dataset containing respiratory sound recordings from individuals with various pulmonary disorders. Gaussian noise amplification and a class-weighted loss function are applied to remove class imbalance. All three architectures configurations have similar convolutional fronts but differ in bidirectional processing and the inclusion of an attention block with residual reinforcement. While all configurations share a common convolutional backbone, they diverge in their temporal modeling capabilities specifically in bidirectional sequence handling and attention integration. Notably, the attention-augmented model demonstrated the highest predictive accuracy (93.8%), surpassing both the recurrent (91.5%) and base convolutional (89.2%) counterparts. Despite this advancement, differentiating between pneumonia and COPD remains problematic due to acoustic similarities.

Phonocardiography and Auscultation TechniquesArtificial Intelligence in HealthcareCOVID-19 diagnosis using AIrespiratory sound analysismulti-head self-attentionconvolutional neural networks

Deep Learning with Multi-Head Attention for Respiratory Sound Analysis

Abstract

More by Assel Mukasheva