⚡ electrical engineering

XAI-Driven Spectral Analysis of Cough Sounds for Respiratory Disease Characterization

This paper proposes an XAI-driven methodology using occlusion maps to identify disease-specific spectral regions in cough sounds, revealing significant acoustic differences between respiratory disease groups that are undetectable through raw spectrogram analysis.

Original authors: Patricia Amado-Caballero, Luis Miguel San-José-Revuelta, María Dolores Aguilar-García, José Ramón Garmendia-Leiza, Carlos Alberola-López, Pablo Casaseca-de-la-Higuera

Published 2026-02-18

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Patricia Amado-Caballero, Luis Miguel San-José-Revuelta, María Dolores Aguilar-García, José Ramón Garmendia-Leiza, Carlos Alberola-López, Pablo Casaseca-de-la-Higuera

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to identify a specific singer just by listening to their cough. It sounds strange, right? But for doctors, the way a person coughs can actually hold secret clues about whether they have asthma, pneumonia, lung cancer, or a chronic condition like COPD.

This paper is about teaching computers to listen to these coughs and, more importantly, to explain why they think a cough belongs to a specific disease.

Here is the story of their research, broken down into simple concepts:

1. The Problem: The "Black Box" Detective

For a long time, scientists have used Artificial Intelligence (AI) to listen to coughs. They built "Deep Learning" models (think of them as super-smart, but invisible, detectives) that could look at a visual picture of a cough (called a spectrogram) and say, "Yes, that's a cough," or "No, that's just someone talking."

However, these AI detectives had a big flaw: They were "Black Boxes."

The Analogy: Imagine a detective who points at a crime scene and says, "The criminal was here!" but refuses to tell you why. They just give you the answer without showing their work.
The Issue: If the AI says, "This cough looks like COPD," but you can't see which part of the cough made it decide that, doctors can't trust it. They need to know what to look for.

2. The Solution: The "Highlighter Pen" (XAI)

The researchers decided to use a technique called XAI (Explainable AI). Think of XAI as a magical highlighter pen for sound.

How it works: They took the AI's "Black Box" and asked it to look at a cough spectrogram (which looks like a colorful map of sound frequencies over time).
The Occlusion Map: The AI was asked to "cover up" (occlude) small parts of the map one by one. If covering a specific spot made the AI say, "Wait, I'm not sure this is a cough anymore," then that spot was important.
The Result: The AI drew a map showing exactly which parts of the cough sound were the most important for making a diagnosis. It's like the AI saying, "I'm not looking at the whole song; I'm focusing specifically on this high-pitched squeak in the middle."

3. The Experiment: Filtering the Noise

Once they had these "highlighted" maps, they did something clever. They didn't just look at the whole cough sound; they weighted the sound based on the AI's highlights.

The Analogy: Imagine you are trying to hear a friend's voice at a loud party.
- Raw Data: You hear the friend, the music, the clinking glasses, and the crowd. It's a mess. You can't tell if the friend is happy or sad.
- XAI-Driven Data: The AI acts like a noise-canceling headphone that mutes everything except the friend's voice. Suddenly, you can hear the tremble in their voice or the sharpness of their tone.

The researchers took the cough sounds, applied this "noise-canceling" filter based on the AI's highlights, and then analyzed the remaining sound.

4. The Discovery: COPD Sounds Different

When they analyzed the "filtered" coughs, they found something amazing that they missed before:

Before (Raw Data): When looking at the whole cough, the sounds of COPD patients looked very similar to other diseases. It was like trying to tell the difference between two types of coffee beans while they are still in the bag.
After (XAI Data): Once they focused on the specific parts the AI highlighted, the differences became huge.
- The COPD Clue: Patients with COPD (a chronic lung disease) had coughs that were more chaotic and variable in the specific frequency ranges the AI cared about.
- The Metaphor: If a healthy cough is like a steady drumbeat, a COPD cough (in the highlighted zones) sounded like a drumbeat that was being played by someone who was slightly out of breath and stumbling. The "energy" of the sound was spread out differently.

5. Why This Matters

The paper proves two main things:

Trust: By showing where the AI is looking, doctors can trust the diagnosis more.
Accuracy: The AI found disease patterns that were invisible when looking at the "whole picture." It's like finding a fingerprint that was hidden under a layer of dust.

Summary

Think of this research as giving a doctor a pair of X-ray glasses for listening to coughs. Instead of just hearing a cough, they can now see the "skeleton" of the sound—the specific frequencies that reveal if a patient has COPD, cancer, or pneumonia. By using AI not just as a judge, but as a guide that points out the evidence, they are making respiratory disease diagnosis faster, more accurate, and much easier to understand.

1. Problem Statement

Respiratory diseases (e.g., COPD, pneumonia, lung cancer, asthma) are a leading cause of mortality, requiring effective monitoring and early diagnosis. While deep learning (DL) models have achieved high accuracy in detecting cough events, they suffer from two critical limitations:

The "Black Box" Nature: DL models (specifically CNNs) lack interpretability, making it difficult to understand which acoustic features drive specific disease classifications.
Diagnostic Gap: There is a scarcity of research linking detected cough events to the underlying physiological mechanisms of specific diseases. Previous attempts to use Explainable AI (XAI) for cough analysis relied on fitting Gaussian Mixture Models (GMMs) to occlusion maps. This approach had limitations: GMMs often provided poor fits for certain spectrograms, and the resulting features were heavily dependent on the precise temporal delineation (start/end times) of the cough event, which is difficult to achieve in real-world scenarios.

The paper addresses the need for a methodology that not only detects coughs but also extracts interpretable, disease-specific spectral signatures without relying on precise event boundary detection or complex statistical fitting models.

2. Methodology

The authors propose a novel pipeline combining Convolutional Neural Networks (CNNs) with XAI techniques to extract and analyze spectral features from cough sounds.

A. Data Collection and Preprocessing

Dataset: Approximately 15,000 single cough events recorded over 24-hour periods from 20 ambulatory patients in Spain (ages 23–87).
Pathologies: The cohort included patients with COPD, lung cancer, pneumonia, acute respiratory disease (ARD), asthma, bronchiectasis, and sarcoidosis.
Comparison Groups: Patients were divided into six groups (G1–G6) to compare chronic vs. non-chronic, COPD vs. other diseases, and specific disease pairings.
Preprocessing: Audio was down-sampled to 8.82 kHz, segmented into 1-second windows, and converted into 45×100 spectrograms (logarithmically normalized) to serve as CNN inputs.

B. Cough Detection (CNN)

A custom CNN architecture was trained to classify spectrogram windows as "cough" or "non-cough."

Architecture: Multiple convolutional layers (32 to 256 filters), ReLU activations, Max-Pooling, Dropout, and Fully Connected layers with Softmax output.
Training: Used the AdaMax optimizer with 5-fold cross-validation to ensure patient independence between training and testing sets.

C. XAI-Driven Feature Extraction

Once the CNN identified high-confidence coughs (>90%), the following XAI steps were applied:

Occlusion Maps: The input spectrograms were systematically occluded (masked) with a sliding window. The drop in classification probability indicated the importance of specific regions. This generated an "occlusion map" highlighting the most relevant spectral regions for the cough classification.
Weighted Spectrograms: The original spectrograms were pixel-wise multiplied by the thresholded occlusion maps. This created enhanced spectrograms that amplified only the disease-relevant spectral regions while suppressing noise and irrelevant areas.
Spectral Feature Extraction: Instead of fitting GMMs, the authors extracted standard spectral features directly from these weighted spectrograms. The features included:
- Relative AC Power
- Spectral Bandwidth (SpBW)
- Spectral Crest Factor (SpCF)
- Spectral Flatness (SpF)
- Spectral Flux (SpFX)
- Spectral Renyi Entropy (SpRE)
- Spectral Roll-Off (SpR)

D. Statistical Analysis

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

3. Key Contributions

XAI-Enhanced Spectral Analysis: The paper introduces a method to use occlusion maps to weight spectrograms, allowing for the extraction of spectral features that are specific to the disease mechanism rather than just the presence of a cough.
Elimination of GMM Dependency: Unlike previous work [61], this approach does not rely on fitting Gaussian Mixture Models to the data, avoiding issues related to model goodness-of-fit and bimodal distribution assumptions.
Robustness to Event Delineation: Because the features are derived from the spectral distribution of the weighted spectrogram rather than temporal event boundaries, the method does not require precise identification of the cough's start and end timestamps.
Discovery of Disease-Specific Signatures: The study demonstrates that significant differences between disease groups (particularly COPD) are only visible when analyzing XAI-weighted spectrograms, whereas raw spectrograms show no significant differences.

4. Key Results

The study analyzed six comparison groups (G1–G6). The most significant findings were observed in COPD patients compared to other groups:

Discriminative Power: When using XAI-weighted spectrograms, multiple features (AC Power, SpBW, SpCF, SpF, SpRE, SpR) showed statistically significant differences ( $p < 0.05$ ) for groups G2, G3, and G4 (COPD vs. other diseases).
Raw vs. Weighted: When analyzing raw (non-weighted) spectrograms, no feature achieved statistical significance for any group. This proves the necessity of the XAI weighting step.
COPD Characteristics:
- Lower AC Power & SpCF: Indicates a lower frequency "center of gravity" and less variability in the dominant frequency.
- Higher Spectral Bandwidth & Roll-Off: Indicates a more dispersed power distribution extending to higher frequencies.
- Higher Spectral Flatness & Renyi Entropy: Suggests a "flatter" spectrum resembling white noise, implying a more random and broadband distribution of energy.
Comparison with Previous Work: The proposed spectral features outperformed the temporal features (mean and standard deviation of GMM parameters) used in the previous study [61] for almost all comparison groups.
Sample Size Limitation: While trends were visible for Lung Cancer (G6), statistical significance was not reached due to the small sample size (2 patients), though boxplots showed clear separation.

5. Significance and Conclusion

This paper establishes that Explainable AI is not just a tool for validation but a critical component for feature discovery in medical audio analysis.

Clinical Relevance: The findings suggest that COPD coughs possess a distinct "acoustic signature" characterized by higher randomness, broader frequency distribution, and flatter spectra in the regions most relevant to the CNN's decision-making. This could lead to more accurate, non-invasive diagnostic tools for COPD and other respiratory conditions.
Methodological Advancement: By shifting from "black box" detection to "white box" spectral analysis, the authors provide a framework that is both interpretable and robust. The ability to detect disease-specific patterns without precise cough segmentation makes this approach highly suitable for continuous, real-world monitoring via mobile devices.
Future Impact: The methodology paves the way for developing interpretable diagnostic systems that can differentiate between complex respiratory diseases (e.g., distinguishing COPD from asthma or pneumonia) based on the intrinsic acoustic properties of the cough, rather than just its presence.