From cough, to speech, to even breathing, the sounds our bodies make may talk much about our health.Answers to more subtle clues that lie in these bioacoustics sounds might hold the potential to completely revolutionise ways of routine screening, diagnosing, and monitoring a huge number of conditions: for example, TB or COPD. Researchers at Google have recognized a huge potential for sound as an effective indicator of health and also realize microphones in smartphones are essentially everywhere nowadays. Here the company has been exploring ways that it may utilize AI for the extraction of health insights from acoustic data. The HeAR Acoustic Representations of Health, or HeAR, which is a foundational model for bioacoustics that will assist researchers in developing models that listen to sound—be they of human sounds or of any creature that had something wrong in the body—to flag the first sign of the rumblings of disease.
HeAR was trained using 300 million audio pieces of data curated from a heterogeneous de- identified dataset. For example, one model released by Google Research was trained on about 100 million cough sounds.HeAR teaches the model to extract patterns from health-related sounds, thus contributing to a very strong foundation for the analysis of medical audio. HeAR outperforms other baselines on a very wide range of tasks and in generalization capabilities across microphones, thus proving itsbetter ability to capture meaningful patterns in health-related acoustic data. This model also presented high performance with much less data—which is important for health studies.
HeAR is now available for use by researchers to accelerate the development of custom bioacoustics models with less data, setup, and computation, on the research of models for specific conditions and populations even in data-sparse scenarios or if cost or computation barriers exist.
An Indian respiratory healthcare company, Salcit Technologies, has developed a product called Swaasa through which it uses AI to analyze cough sounds and assess lung health. Getting into the nitty-gritty, the involvement of HeAR into their process this time is going to be how it increases the functionality of their bioacoustic AI models. First, Swaasa is using HeAR to help research and enhance its early detection of TB based on cough sounds.
TB is curable, yet every year millions of cases go unnoticed—many times because people don't have quick and easy access to health services. Improving diagnosis is key to ending TB, and artificial intelligence can play an important role in improving detection in ways that will eventually render care accessible and affordable to all.
There is a history of using machine learning at Swaasa to facilitate the early detection of diseases, closing the accessibility, affordability, and scalability gap with location-independent and equipment-free respiratory health assessment. Using HeAR, they see an opportunity to take this work on a much larger scale: extending screening for TB more widely across India.HeAR is a novel development in the research of acoustic health, furthering future diagnostic tools, monitoring solutions for TB, chest, lung, and other diseases towards better health outcomes in communities across the world.
