AI Model Aids in Disease Detection Through Cough Analysis

From the sound of a cough to the rhythm of our breath, the noises our bodies produce are packed with health-related info. These subtle bioacoustic signals could totally change the game in screening, diagnosing, monitoring, and managing various health issues, like tuberculosis (TB) or chronic obstructive pulmonary disease (COPD). At Google, we see the huge potential in using sound as a health indicator, especially since smartphone mics are so common. That's why we've been diving into how AI can pull health insights from these sounds.

Earlier this year, we rolled out Health Acoustic Representations, or HeAR, a bioacoustic foundation model that helps researchers create models capable of listening to human sounds and spotting early disease signs. The Google Research team trained HeAR on a whopping 300 million audio clips from a diverse, de-identified dataset. For the cough model specifically, we used around 100 million cough sounds.

HeAR is designed to pick up on patterns in health-related sounds, laying a solid groundwork for medical audio analysis. We've found that, on average, HeAR outperforms other models across a bunch of tasks and is great at working with different microphones, showing its knack for capturing meaningful health-related sound patterns. Plus, models built with HeAR can achieve high performance even with less training data, which is a big deal in the data-limited world of healthcare research.

Now, HeAR is available for researchers to speed up the development of custom bioacoustic models, even when data, setup, or computing power is limited. Our aim is to boost research into models for specific conditions and groups, no matter how sparse the data or how high the costs.

Salcit Technologies, a respiratory health company based in India, has developed Swaasa®, an AI-powered tool that analyzes cough sounds to assess lung health. They're now looking into how HeAR can boost their bioacoustic AI models. Swaasa® is starting by using HeAR to improve their early detection of TB through cough sounds.

TB is treatable, but millions of cases go undiagnosed each year, often because people can't easily access healthcare. Better diagnosis is key to wiping out TB, and AI can make a big difference in detection and making care more accessible and affordable worldwide. Swaasa® has been using machine learning to catch diseases early, making health assessments more accessible, affordable, and scalable without needing special equipment or a specific location. With HeAR, they're looking to expand TB screening across India.

"Every missed TB case is a tragedy; every late diagnosis, a heartbreak," says Sujay Kakarmath, a product manager at Google Research working on HeAR. "Acoustic biomarkers could change this story. I'm really thankful for the part HeAR can play in this journey."

We're also getting support from groups like The StopTB Partnership, a UN-hosted organization that connects TB experts and affected communities to end TB by 2030.

"Tools like HeAR can push forward AI-powered acoustic analysis in TB screening and detection, offering a low-impact, accessible solution to those who need it most," said Zhi Zhen Qin, a digital health specialist with the Stop TB Partnership.

HeAR is a big leap forward in acoustic health research. We're hoping to help develop future diagnostic tools and monitoring solutions for TB, chest, lung, and other diseases, and improve health outcomes for communities everywhere through our research. If you're a researcher interested in HeAR, you can find out more and request access to the HeAR API.