Optimizing 2D Input Representations and Sub-phase Fusion Strategies for Differential Diagnosis of Asthma and COPD Using CNN- and GRU-Based Networks

This study evaluates machine learning approaches for distinguishing asthma from COPD using pulmonary sound analysis, comparing MFCC matrices, log-mel spectrograms, and VAR models with CNN and GRU networks. MFCC representations with adaptive-length windowing achieved the best performance (F1-score 0.877), while sophisticated fusion strategies and data augmentation unexpectedly degraded results, emphasizing the importance of authentic clinical data.

This research addresses a critical gap in pulmonary disease diagnosis by systematically comparing spectral representation methods for respiratory sound classification. The study's primary contribution lies in demonstrating that simpler, well-established feature extraction methods outperform more complex alternatives when combined with appropriate temporal alignment strategies. The adaptive-length windowing approach solves a fundamental problem in pulmonary sound analysis—varying respiratory cycle durations that create inconsistent temporal dimensions in spectrograms.

The finding that MFCC matrices substantially outperformed log-mel spectrograms and VAR models challenges conventional assumptions in modern deep learning applications. Typically, researchers gravitate toward more sophisticated architectures and augmentation techniques to improve model performance. This study reveals that in medical sound classification, data authenticity and appropriate feature representation matter more than architectural complexity. The negative impact of augmentation techniques, including mixup methods, suggests that synthetic variations may not capture the physiological nuances distinguishing asthma from COPD.

For healthcare AI development, these results have significant implications. They indicate that domain-specific feature engineering remains valuable despite the rise of end-to-end deep learning approaches. The superior performance of direct feature concatenation over gated recurrent units with attention mechanisms suggests that respiratory diagnostic patterns may not require sequential modeling complexity. This finding could accelerate development of interpretable diagnostic tools, as MFCC features are more transparent than learned representations from sophisticated fusion networks.

The research suggests future work should prioritize collecting larger authentic datasets rather than synthetic augmentation and should carefully evaluate whether architectural complexity genuinely improves clinical utility. These insights could influence how medical AI developers approach respiratory disease diagnosis.

• →MFCC matrices with adaptive-length windowing achieved superior F1-scores (0.877 cycle-based, 0.855 subject-based) for asthma-COPD differentiation.
• →Sophisticated fusion strategies using GRU networks and attention mechanisms did not improve diagnostic performance compared to simple concatenation.
• →Data augmentation techniques, including mixup methods, degraded model performance, highlighting the importance of authentic clinical data over synthetic variations.
• →MFCC feature extraction outperformed modern alternatives like log-mel spectrograms and VAR models in this medical sound classification task.
• →Domain-specific feature engineering proved more effective than complex deep learning architectures for pulmonary sound analysis.