Cross-Attention Transformer Networks with Optimized Feature Selection for Explainable Respiratory Disease Classification

Abstract

Respiratory diseases require accurate diagnosis for effective treatment, yet traditional methods rely on subjective assessments and expensive procedures. This paper presents a transformer-based cross-attention framework for acoustic respiratory disease classification with explainable AI integration. The CrossAttentionAcousticNetwork combines CNN spectral feature extraction with transformer temporal modeling, enhanced by cross-attention mechanisms for multi-modal feature fusion. Namib Beetle Optimization selects discriminative features from 100-dimensional handcrafted and deep spectral representations, while LIME provides clinical interpretability. Evaluation on ICBHI 2017 and KAU datasets achieves 99.0% and 95.0% accuracy respectively, representing 21.39% improvement over existing methods. The framework demonstrates superior performance across asthma, COPD, pneumonia, and heart failure classifications while maintaining computational efficiency for real-time deployment. Integrated explainable AI reveals clinically relevant acoustic patterns, with data augmentation improving minority class recognition by 19%. This approach bridges the gap between high-performance deep learning and clinical transparency requirements for automated respiratory disease screening.