A XAI-based Framework for Frequency Subband Characterization of Cough Spectrograms in Chronic Respiratory Disease

1. Problem Statement

Respiratory diseases, particularly Chronic Obstructive Pulmonary Disease (COPD), pose a significant global health burden. While deep learning (DL) has shown promise in automatically detecting and classifying cough sounds, two major limitations persist:

1. Data Scarcity: High-performance models often require large, diverse datasets which are difficult to obtain.
2. The "Black Box" Problem: DL models lack interpretability, making it difficult to understand the underlying pathophysiological mechanisms driving the classification. This hinders clinical trust and the ability to identify specific acoustic markers for different diseases.

Existing Explainable AI (XAI) methods (e.g., Grad-CAM, LIME) often highlight broad or diffuse regions in spectrograms that may not correspond to truly discriminative features, or they rely heavily on precise temporal onset/offset detection, which is sensitive to noise.

2. Methodology

The authors propose a novel framework that combines Convolutional Neural Networks (CNNs) with XAI techniques to extract frequency subband-specific spectral features. The methodology consists of four main stages:

A. Data Acquisition and Preprocessing

• Dataset: 17 adult patients with persistent coughs were recorded over full days in real-world conditions using a smartphone (44.1 kHz sampling rate).
• Groups: Patients were categorized into six comparison groups (G1–G6) to distinguish between:
  • Chronic vs. Non-chronic (G1).
  • COPD vs. Other diseases (G2, G3, G4, G5, G6).
• Signal Processing: Audio was downsampled to 8.82 kHz. Power Spectral Density (PSD) was computed using 10ms Hanning windows (no overlap) and concatenated into 1-second intervals. The resulting spectrograms were log-normalized to create 45 × 100 input matrices for the CNN.

B. CNN Architecture for Cough Detection

A custom CNN was trained to distinguish coughs from non-cough sounds:

• Structure: Input $\rightarrow$ Conv (32 filters, 2x2) $\rightarrow$ MaxPool $\rightarrow$ Dropout $\rightarrow$ Conv (64 filters) $\rightarrow$ MaxPool $\rightarrow$ Dropout $\rightarrow$ Conv (128 filters) $\rightarrow$ Conv (256 filters) $\rightarrow$ MaxPool $\rightarrow$ Fully Connected (512 neurons) $\rightarrow$ Softmax (Binary classification).
• Training: Used AdaMax optimizer, batch size 128, 50 epochs, and 5-fold cross-validation to prevent data leakage between patients.

C. XAI-Driven Feature Extraction (The Core Innovation)

Instead of analyzing the entire spectrogram, the framework isolates diagnostically relevant regions and decomposes them into frequency subbands:

1. Occlusion Maps: For high-confidence cough predictions ($\ge$90%), a sliding mask occludes parts of the spectrogram. The drop in classification probability indicates the importance of the occluded region. These maps are averaged per patient to create a representative "importance map."
2. Weighted Spectrograms: The original spectrograms are multiplied by a binary mask derived from the occlusion map (thresholded at a specific percentile). This amplifies only the spectral regions the CNN deemed critical for detection.
3. Frequency Subband Decomposition: The weighted spectrograms are segmented into five distinct frequency bands:
   • B1: 0 – 0.5 kHz
   • B2: 0.5 – 1.0 kHz
   • B3: 1.0 – 1.5 kHz
   • B4: 1.5 – 2.0 kHz
   • B5: 2.0 – 4.41 kHz
4. Feature Extraction: Seven spectral features were calculated for each subband (and the global band for comparison):
   • Relative Power (RP)
   • Spectral Bandwidth (SpBW)
   • Spectral Crest Factor (SpCF)
   • Spectral Flatness (SpF)
   • Spectral Flux (SpFX)
   • Spectral Rényi Entropy (SpRE)
   • Spectral Roll-Off (SpR)

D. Statistical Analysis

Hypothesis testing (Student's t-test or Mann-Whitney U-test) was performed to identify statistically significant differences ($p < 0.05$) in these features between the defined patient groups.

3. Key Contributions

• Subband-Specific Analysis: Unlike previous works that analyzed the full frequency spectrum, this paper demonstrates that decomposing cough signals into specific frequency subbands reveals discriminatory patterns that are often masked or cancelled out in global analysis.
• XAI-Weighted Spectrograms: The use of occlusion maps to generate "Weighted Spectrograms" ensures that feature extraction focuses only on the diagnostically relevant parts of the cough, reducing noise interference.
• Interpretability: The framework provides physically meaningful spectral markers (e.g., energy distribution, entropy, flatness) linked to specific frequency ranges, offering insights into the pathophysiology of COPD and other chronic conditions.
• Robustness: By focusing on spectral features rather than precise temporal onset/offset, the method is more robust to variations in cough timing and background noise.

4. Key Results

• Superior Discriminatory Power: Subband analysis consistently outperformed global band analysis. For example, in Group G1 (Chronic vs. Non-chronic), features like Relative Power (RP) and Spectral Roll-Off (SpR) were significant in specific bands (1, 2, 4) but showed no significance in the global band due to compensatory trends across frequencies.
• COPD Specific Patterns:
  • Low Frequencies (B1, B2): COPD patients exhibited significantly higher Relative Power and Spectral Entropy compared to other groups, suggesting increased energy and chaotic spectral content in lower frequencies.
  • Mid Frequencies (B3, B4): COPD patients showed lower Relative Power and Spectral Bandwidth compared to other diseases, indicating a "dampening" effect in these specific ranges.
  • High Frequencies (B5): COPD patients showed higher variability and energy content compared to the global average.
• Feature Behavior:
  • Spectral Flux (SpFX): Was uninformative in global analysis but highly discriminative in subbands, revealing that COPD coughs have rigid temporal variability in bands 1, 2, and 4.
  • Spectral Rényi Entropy (SpRE): Consistently showed higher values for COPD patients across multiple bands, indicating a more chaotic spectrum.
• Comparison with State-of-the-Art: When compared to the previous global-band approach (reference [57]), the subband framework identified significant differences in groups (like G1 and G6) where the global method failed.

5. Significance and Conclusion

This study demonstrates that frequency-resolved, XAI-enhanced analysis is a powerful tool for biomedical signal interpretation.

• Clinical Insight: The findings suggest that COPD alters the acoustic signature of coughs in a frequency-dependent manner (e.g., energy shifts from mid-frequencies to low/high extremes), providing a deeper understanding of the disease's pathophysiology.
• Diagnostic Potential: The identified spectral markers can serve as interpretable features for developing more accurate, transparent, and robust diagnostic tools for chronic respiratory diseases.
• Future Work: While the results are promising, the authors note the need for validation on larger, more diverse datasets to confirm the generalizability of these subband-specific markers.

In summary, the paper successfully bridges the gap between "black box" deep learning and clinical interpretability by showing that where in the frequency spectrum a disease manifests is just as important as what the overall signal looks like.