Optimization results comparison.

Respiratory disease diagnosis remains challenging in resource-constrained settings, where limited specialist expertise contributes to diagnostic uncertainties affecting over 300 million people worldwide. This study presents E-RespiNet, a novel multi-modal deep learning architecture that integrates ELECTRA’s discriminative pre-training with a triple-stream convolutional neural network framework, enhanced through Harmony Search with Opposition-Based Learning optimization for automated respiratory sound classification. The architecture simultaneously processes mel-frequency cepstral coefficients, discrete wavelet transforms, and mel-spectrograms through parallel CNN streams, with features integrated through hierarchical fusion and ELECTRA-based contextual enhancement. Comprehensive evaluations on two independent clinical datasets—the Asthma Detection Dataset Version 2 (1,211 recordings across five conditions) and King Abdullah University Hospital dataset (940 samples from 81 subjects across four conditions)—demonstrated exceptional performance with 98.9% and 98.8% accuracy respectively, representing 5.0% and 4.3% improvements over baseline configurations. Cross-institutional validation revealed 75.7% average accuracy with a 23.3% generalization gap, substantially better than typical medical AI cross-domain performance. Gradient-weighted class activation mapping provided clinically relevant interpretability, while the Harmony Search optimization framework enhanced both performance and cross-institutional robustness. These results demonstrate significant advances in automated respiratory sound analysis through discriminative language model integration and metaheuristic optimization, establishing important benchmarks for deployable respiratory diagnostic tools in diverse healthcare settings.