Diagnosing COVID-19 with a Cough: The Potential of AI-Powered Audio Analysis

The COVID-19 pandemic has highlighted the urgent need for fast, accurate, and accessible methods to diagnose this potentially deadly respiratory virus. While PCR tests remain the gold standard, they require specialized lab equipment and typically take at least 24 hours to deliver results. Rapid antigen tests offer faster turnaround times but are less sensitive than PCR. CT scans can detect COVID-induced pneumonia but involve radiation exposure.

In the search for better solutions, researchers have turned to an unexpected diagnostic tool: coughs. It turns out the sound of a cough contains a wealth of information that, when analyzed by artificial intelligence algorithms, can help determine if someone is likely to have COVID-19. This article will explore how AI-powered cough audio analysis works, its advantages and limitations, the latest research, and the potential applications of this promising technology.

The Sound of Sickness: Features of a COVID Cough

To the human ear, coughs may sound more or less the same. But audio processing software can extract a large number of features from cough recordings that reveal distinct "audio signatures" of different respiratory conditions. Some key features include:

• Mel-frequency cepstral coefficients (MFCCs): Commonly used in speech recognition, MFCCs capture information about the shape of the vocal tract during a cough. COVID coughs tend to have different MFCC patterns than flu or cold coughs.

• Spectral centroid: This measures the "center of mass" of the cough sound spectrum. COVID coughs may have higher or lower spectral centroids compared to other coughs.

• Zero-crossing rate: The rate at which a cough signal crosses the zero line (between positive and negative) can differ between COVID and non-COVID coughs.

• Chroma: Chroma features project cough audio onto a set of 12 bins representing musical notes. The chroma profile of COVID coughs appears to be distinct.

By extracting a large set of audio features from many cough samples, researchers can train machine learning models to recognize the features most indicative of a COVID cough.

AI Listens and Learns: Machine Learning for Cough Classification

With a dataset of cough recordings labeled as COVID-positive or negative, data scientists can leverage the power of machine learning to build cough audio classifiers. Some commonly used algorithms include:

• Deep learning: Neural networks with many interconnected nodes can learn complex patterns in cough audio spectrograms. Convolutional neural networks (CNNs) are well-suited for this kind of audio classification task.

• Transfer learning: Models pre-trained on other audio datasets (e.g. for speech recognition) can be fine-tuned on a smaller number of COVID cough samples, boosting accuracy.

• Ensemble methods: Combining the predictions of multiple models trained on different feature sets can yield better results than any single model.

The best-performing models in recent studies have achieved over 90% accuracy in distinguishing COVID coughs from other coughs or healthy coughs. Some can even predict the severity of COVID cases based on cough sounds.

The Case for Cough Audio Analysis

So why bother with cough audio when we already have COVID tests? There are several potential advantages:

1. Speed: Cough audio can be recorded and analyzed in seconds or minutes using just a smartphone. Results could be nearly instant.

2. Cost: Compared to PCR kits or antigen tests, cough audio requires no special equipment other than a phone or computer. Analysis software can be accessed for free or at low cost.

3. Non-invasiveness: Recording a cough is quick and painless compared to a nasal swab. This could increase willingness to get tested.

4. Asymptomatic detection: People with no obvious symptoms may still have detectable changes in their cough sounds. Cough audio analysis could help catch these stealthy cases.

5. Remote screening: Cough sounds could be analyzed through telemedicine visits or even through recordings submitted via an app. This enables screening without in-person contact.

Challenges, Limitations and Open Questions

While the potential of cough audio analysis for COVID diagnosis is exciting, there are still hurdles to overcome before it can be widely implemented:

• Noise and variability: Cough sounds can be affected by background noise, microphone type, and individual variations in cough strength or duration. Classifiers need to be robust to these variables.

• Demographic differences: Cough acoustics may differ by age, gender, or health status (e.g. smokers). Models need to perform well across all populations to avoid bias.

• Distinguishing from other conditions: Many respiratory illnesses can cause cough. Classifiers need to be specific enough to rule out flu, colds, asthma and other possible confounders.

• Privacy concerns: Users need assurance that their cough recordings will be kept secure and confidential. Transparent data practices are essential.

• Regulatory approval: Like any new diagnostic method, AI cough analysis needs to be rigorously validated in larger-scale trials and cleared by health authorities before clinical use.

The Road Ahead: Research and Applications

Despite the challenges, research on AI-powered COVID cough detection has progressed rapidly since 2020. Some notable projects include:

• The AI Cough app by MIT, which claims 98.5% accuracy in detecting COVID coughs
• The Coswara dataset from India with over 1,500 COVID cough recordings
• A Cambridge University study that achieved 80% accuracy using only crowdsourced cough data
• Partnerships between AI health companies like Hyfe and health systems to collect prospective, real-world cough data

As this technology advances, some exciting potential applications are:

1. Population screening: Imagine a future where you simply cough into your phone each morning for your personal COVID check. Positive results could prompt confirmation by standard tests.

2. Remote monitoring: For those who test positive, cough acoustics could be used to track disease progression and recovery at home, alerting doctors to cases that worsen.

3. Early warning systems: Changes in the aggregate "cough profile" of a city or country could provide early indication of rising COVID cases before they show up in testing data.

4. Expanding to other diseases: The same AI techniques could be adapted to diagnose flu, pneumonia, and other illnesses that affect the respiratory system. Multiplex cough screening could be a powerful public health tool.

Conclusion

The idea of diagnosing a disease as tricky as COVID-19 from something as simple as a cough sound is equal parts fascinating and challenging. The science is still evolving, but early results suggest AI analysis of cough audio could become a valuable addition to our pandemic-fighting toolkit: a low-cost, rapid, and non-invasive screening method to help control the spread of the virus.

However, as with any AI health application, we must be thoughtful about how this technology is developed and deployed. Rigorous clinical validation, protection of user privacy, and transparency about the limitations are essential to build public trust. And even the cleverest AI is no substitute for access to gold-standard diagnostic tests and quality healthcare.

Looking ahead, the turbocharged research into COVID cough acoustics will likely yield insights relevant to other respiratory diseases. Having a readout of lung health as close as your phone could open up many new possibilities in telemedicine and predictive care. At the very least, you may never think about coughs the same way again!