Predicting Tuberculosis from Real-World Cough Audio Recordings and Metadata

This study demonstrates that a mobile phone-based application utilizing automated cough audio recordings combined with clinical metadata can effectively predict tuberculosis with an AUC of approximately 0.81, offering a cost-effective tool to enhance community case-finding efforts.

The Gist

Imagine you have a very sick friend who keeps coughing. In the old days, to figure out if they have Tuberculosis (TB)—a serious lung infection—you'd have to send them to a clinic, have them spit into a cup, and wait days for a lab to look at the sample under a microscope. It's slow, expensive, and many people in remote areas never get tested.

This paper proposes a new, high-tech way to solve that problem: listening to the cough itself.

Here is the story of how the researchers tried to teach a computer to be a "super-listener" for TB, explained in simple terms.

The Big Idea: The "Cough Detective"

The researchers, working with an app called Hyfe, collected thousands of cough recordings from people in India, Africa, and Asia. They didn't just ask people to cough; they used a smartphone app that guided them, recorded the sound, and automatically picked out the actual coughs from background noise.

They then asked a simple question: "Can a computer tell the difference between a TB cough and a regular cold cough just by listening?"

How They Taught the Computer (The Two Experiments)

The team ran two different "training camps" for their computer algorithms:

1. The "Blindfold" Test (Cough-Only)

In this experiment, the computer was given only the sound of the cough. It was like trying to guess what kind of animal is in a room just by hearing it bark, without seeing it or knowing its size.

• The Tools: They used two types of "ears" for the computer:
  • The Statistician: This looked at the sound's basic math (how loud it is, how fast the pitch changes, the "texture" of the noise).
  • The Deep Learner (CNN): This is a fancy type of AI that looks at the cough sound like a picture. Imagine turning the sound wave into a colorful map (a spectrogram). The AI scans this map like a detective looking for a specific pattern, similar to how your phone recognizes a face in a photo.
• The Result: The computer got about 70% right. It was better than a coin flip, but not perfect. It could hear something different about TB coughs, but it wasn't confident enough to be a doctor on its own.

2. The "Detective with Clues" Test (Cough + Metadata)

In the second experiment, they gave the computer more information. Along with the cough sound, they fed it a "patient profile" (called metadata). This included:

• Age and gender
• Body temperature
• Heart rate
• How long they've been coughing
• Whether they've lost weight or had night sweats
• Whether they've had TB before

Think of this like a detective who can now see the suspect and ask them questions, not just listen to them.

• The Result: The accuracy jumped to 81%. By combining the sound of the cough with the patient's symptoms, the computer became much smarter.

The "Secret Sauce" (How the Computer "Hears")

You might wonder, how does a computer understand a cough? Humans hear a cough as a "hack" or a "wheeze." The computer breaks it down into tiny pieces:

• The "Energy": How hard did they cough?
• The "Pitch": Is it high-pitched or low and rumbly?
• The "Texture": Is it smooth or scratchy?
• The "Map": They turned the sound into a visual map (like a topographic map of a mountain) so the AI could see the shape of the sound.

Why This Matters (The Real-World Impact)

The authors aren't trying to replace doctors. Instead, they want to create a triage tool (a sorting system).

Imagine a community health worker in a village with 50 people coughing. They can't send all 50 to the city for expensive lab tests.

1. The health worker asks everyone to cough into their phone.
2. The app analyzes the sound and the symptoms.
3. The app says: "These 5 people have a high probability of TB. Please prioritize them for the lab test."

This saves money, saves time, and ensures the sickest people get help first.

The Bottom Line

This study shows that AI can listen to a cough and spot TB with surprising accuracy, especially when it also knows the patient's other symptoms.

• Without extra info: The AI is a good guesser (70% accuracy).
• With extra info: The AI is a strong assistant (81% accuracy).

The researchers conclude that if we put this technology into mobile apps in countries where TB is common, we could find many more cases of the disease, stop it from spreading, and save lives, all without needing a massive lab in every village. It's like giving every community health worker a super-powered stethoscope that never gets tired.

Technical Summary

1. Problem Statement

Tuberculosis (TB) remains a leading cause of death globally, with approximately 40% of cases undiagnosed due to barriers in accessing healthcare facilities. Traditional diagnosis relies on clinical exams and sputum testing, which can be costly, invasive, and slow. While cough is a primary symptom of TB, distinguishing TB-related coughs from those caused by other respiratory conditions using human hearing or simple observation is difficult. The paper addresses the challenge of developing a low-cost, non-invasive, digital screening tool using Artificial Intelligence (AI) to analyze cough sounds and metadata to predict TB status, thereby improving case-finding efforts in resource-limited settings.

2. Methodology

Dataset

• Source: Data was derived from the CODA TB DREAM Challenge, involving 7 countries (India, Philippines, South Africa, Uganda, Vietnam, Tanzania, Madagascar).
• Participants: 1,107 individuals (297 TB-positive, 810 TB-negative) aged 18+ with a cough persisting for at least two weeks.
• Data Collection: Coughs were recorded via the Hyfe Research app on mobile phones. The app prompted users to cough, capturing "explosive" sounds within a 5-second window.
• Volume: A total of 9,772 cough sounds were analyzed.
• Ground Truth: TB status was confirmed via molecular and culture testing of sputum samples.
• Metadata: Demographic and clinical variables (age, sex, heart rate, temperature, cough duration, history of TB, hemoptysis, fever, night sweats, weight loss, smoking status) were collected.

Feature Extraction

The authors utilized two main categories of features:

1. Low-Level Descriptors (LLDs): Extracted from audio signals using a sliding window approach (50ms frame size, 25ms hop size, Hamming window).
   • Temporal Features: Energy, Zero-Crossing Rate (ZCR), Intensity.
   • Spectral Features: Spectral Centroid, Spread, Roll-off (90%), Entropy, and Flux.
   • Spectrotemporal Features: Mel-Frequency Cepstral Coefficients (MFCCs) and Log-mel spectrograms.
   • Statistical Summarization: For standard ML models, frame-level features were aggregated using statistics (mean, standard deviation, skewness, kurtosis) to create a fixed-length vector per audio file.
2. Metadata: Tabular data including clinical and demographic factors.

Experimental Design

The study conducted two distinct experiments using stratified grouped cross-validation (ensuring all coughs from a single participant remained in either the training or testing set, never both) to prevent data leakage.

• Experiment 1: Cough-Only

  • Goal: Predict TB using audio features alone.
  • Models:
    • Conventional ML: Logistic Regression (LR), Support Vector Machine (SVM), Multi-Layer Perceptron (MLP), Random Forest (RF), AdaBoost (AB).
    • Deep Learning: 2D Convolutional Neural Networks (2D-CNNs) trained directly on spectrotemporal representations (treating spectrograms as images).
  • Architecture: The best-performing CNN used 3-4 convolutional layers with batch normalization, pooling, and dropout, followed by dense layers.

• Experiment 2: Cough + Metadata

  • Goal: Predict TB using a fusion of audio features and clinical metadata.
  • Approach: Audio features (flattened log-mel spectrograms/MFCCs) were concatenated with tabular metadata.
  • Models: Conventional ML models (LR, SVM, MLP, RF, AdaBoost) capable of handling mixed data types. CNNs were not used here due to the difficulty of jointly training 2D image-like data with 1D tabular data in this specific pipeline.

3. Key Contributions

• Large-Scale Real-World Dataset: This is one of the first studies to evaluate TB prediction on a large, automatically collected dataset (nearly 10k coughs) from diverse geographic regions, rather than small, controlled lab settings.
• Hybrid Approach: The paper demonstrates the efficacy of combining acoustic AI (cough analysis) with clinical metadata to improve diagnostic accuracy.
• Benchmarking: It provides a comprehensive comparison between deep learning (CNNs) and traditional machine learning algorithms for cough-based TB detection.
• Open Science: The work utilizes publicly available training data from the CODA challenge, establishing a baseline for future research in this domain.

4. Results

Performance Metrics (Area Under Curve - AUC)

• Cough-Only Experiment:

  • Cough-Level: The CNN achieved the highest average AUC of 0.70 ± 0.05. Conventional ML models generally performed slightly lower.
  • Participant-Level: When aggregating predictions per participant, Logistic Regression achieved the highest average AUC of 0.69 ± 0.07, slightly outperforming the CNN.
  • Observation: Aggregating probabilities at the participant level generally improved performance over isolated cough classification.

• Cough + Metadata Experiment:

  • Performance Boost: The inclusion of metadata significantly improved results across all models.
  • Top Performer: AdaBoost achieved the highest average AUC of 0.82 ± 0.05 (cough-level) and 0.81 ± 0.04 (participant-level).
  • General Trend: All models (except MLPs) achieved mean/median AUCs well above 0.80. Simple models like Logistic Regression and SVM performed competitively with complex ensemble methods.

Key Findings

• Audio features alone provide moderate predictive power (AUC ~0.70).
• Adding demographic and clinical metadata yields a substantial performance jump (AUC ~0.81–0.82).
• Ensemble methods (AdaBoost) and simple linear models (Logistic Regression) outperformed deep learning (CNNs) and MLPs in the metadata-fusion scenario, suggesting that for tabular data fusion, traditional ML may be more efficient or robust in this context.

5. Significance and Future Work

• Public Health Impact: The results suggest that mobile phone-based applications integrating cough sound analysis and clinical symptoms can serve as effective triage tools. They can help community health workers prioritize which patients require definitive (and more expensive) microbiological testing, reducing costs and improving case-finding rates.
• Scalability: The approach is non-invasive and leverages ubiquitous technology (smartphones), making it suitable for low-resource, high-TB-burden countries.
• Limitations & Future Directions:
  • The study did not exhaustively explore all feature selection schemes or classifier architectures.
  • Future work needs to determine the optimal number of features and the best balance between model complexity and convergence speed.
  • Validation in larger, community-based studies is required to confirm these findings in real-world screening scenarios.
  • Integration into existing digital health systems (e.g., Nikshay in India) is proposed as a next step to enhance epidemiological tracking.

In conclusion, the paper validates that cough audio analysis, especially when combined with clinical metadata, is a viable and promising avenue for the early detection and triage of Tuberculosis.