E-RespiNet: An LLM-ELECTRA driven triple-stream CNN with feature fusion for asthma classification.

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Computational Linguistics

Background:

• Respiratory disease diagnosis is challenging in resource-limited settings due to limited specialist expertise.
• Diagnostic uncertainties impact over 300 million people globally.
• Automated analysis of respiratory sounds holds potential for improved diagnostic accuracy.

Purpose of the Study:

• To develop and evaluate E-RespiNet, a novel multi-modal deep learning architecture for automated respiratory sound classification.
• To integrate ELECTRA's discriminative pre-training with a triple-stream CNN framework and Harmony Search optimization.
• To enhance diagnostic accuracy and robustness in diverse healthcare settings.

Main Methods:

• E-RespiNet processes mel-frequency cepstral coefficients, discrete wavelet transforms, and mel-spectrograms using parallel CNN streams.
• Features are integrated via hierarchical fusion and ELECTRA-based contextual enhancement.
• Harmony Search with Opposition-Based Learning optimizes the architecture for improved performance and robustness.

Main Results:

• Exceptional performance on two clinical datasets, achieving 98.9% and 98.8% accuracy.
• Demonstrated a 5.0% and 4.3% improvement over baseline configurations.
• Achieved 75.7% average accuracy in cross-institutional validation with a 23.3% generalization gap, outperforming typical medical AI.

Conclusions:

• E-RespiNet represents a significant advance in automated respiratory sound analysis.
• The integration of discriminative language models and metaheuristic optimization establishes new benchmarks for deployable diagnostic tools.
• The model shows promise for improving respiratory diagnostics in resource-constrained and diverse healthcare settings.