Another breakthrough in AI! The “cough model” was born. A cough can tell you about your health

Written by | Ma Xuewei

Preface

According to data released by the World Health Organization (WHO) in 2021, the world's top 10 causes of death caused a total of 39 million deaths, accounting for 57% of the global total number of deaths (68 million). They are mainly related to two major types of diseases, namely cardiovascular diseases (such as ischemic heart disease, stroke, etc.) and respiratory diseases (such as tuberculosis, chronic obstructive pulmonary disease, lower respiratory tract infections, etc.).

Among them, coughs or breathing caused by respiratory diseases contain a lot of information about our health status. For example, clinicians use the sound of a cough similar to "wheezing" to diagnose whooping cough, and use dying breathing to detect acute cardiovascular events.

So, in the era of artificial intelligence (AI), can we use this technology to extract health information from these sound data and better monitor our physical condition?

A research team from Google and the Tuberculosis Department of the Zambian Center for Infectious Disease Research has taken an important step in this direction. They have jointly launched a bioacoustic basic model HeAR (Health Acoustic Representations) to help them monitor human voices and mark early signs of disease. The relevant research paper, titled "HeAR - Health Acoustic Representations", has been published on the preprint website arXiv.

According to reports, they trained HeAR on 300 million audio data carefully selected from a diverse and de-identified dataset, and specifically used about 100 million cough sounds to train this "cough model."

HeAR is able to discern patterns in health-related sounds, outranks other models on average on a wide range of tasks, and generalizes across microphones. Models trained with HeAR also achieve high performance with less training data, a key factor in the often data-starved medical research field. HeAR is now available to researchers, helping to accelerate the development of customized bioacoustic models with less data, setup, and compute requirements.

“ Solutions like HeAR will make AI-driven acoustic analysis highly useful in TB screening and detection, providing a potentially low-impact, easily accessible tool to those who need it most, ” said Zhi Zhen Qin, digital health expert at StopTB Partnership.

In the future, the research team hopes to use this research to advance the development of diagnostic tools and monitoring solutions in the fields of tuberculosis, chest, lung and other diseases, and help improve health outcomes for communities around the world.

Today, Indian respiratory health company Salcit Technologies has developed a product called Swaasa that uses AI to analyze cough sounds and assess lung health. The company is exploring how HeAR can help expand the capabilities of its bioacoustic AI model.

A cough can detect diseases

The HeAR system consists of three main parts. Through self-supervised learning, the HeAR system leverages a large amount of unlabeled audio data to learn general audio representations that can be transferred to various health acoustic tasks.

Figure｜HeAR system overview

During the data collection step, the research team used a health acoustic event detector. This is a multi-label classification convolutional neural network (CNN) that is used to identify the presence of six non-speech health acoustic events in 2-second audio clips: cough, baby cough, breathing, throat clearing, laughter, and speaking. The detector was trained using the FSD50K and FluSense datasets and annotated with labels in the audio clips (such as "cough", "sneeze", and "breath").

The paper used two datasets, one of which was 2-second audio clips extracted from 3 billion public non-copyright YouTube videos, totaling 313.3 million clips or about 174,000 hours of audio. These clips were screened using the Health Acoustic Event Detector. The other was collected by the Zambian Center for Infectious Disease Research and contains cough audio recordings and chest X-rays from 599 suspected tuberculosis patients.

The research team used a masked autoencoder trained on a large dataset of 313 million two-second audio clips. Through linear probing, HeAR achieved state-of-the-art performance among all healthy audio embedding models on a benchmark of 33 healthy acoustic tasks across 6 datasets.

Figure | HeAR achieves the highest mean ranking (MRR = 0.708) in 33 health audio tasks, surpassing all other baseline models.

HeAR outperforms other models on FSD50K and FluSense datasets, especially ranking second among the models trained with FSD50K.

Figure｜Performance comparison of the health acoustic event detection task on the FSD50K and FluSense datasets.

HeAR outperforms the baseline model in 10/14 cough inference tasks, including demographics, lifestyle, and performs comparably to the best model in TB and CXR tasks.

Figure｜Performance comparison of cough inference tasks.

HeAR outperforms the baseline model on 4/5 lung function test tasks and gender classification task on the SpiroSmart dataset.

Figure ｜Performance comparison of lung function test tasks.

HeAR's performance on the CIDRZ dataset is not affected by different recording devices and is robust to different devices. In addition, HeAR can achieve good performance even with less training data, which is more advantageous in medical research where labeled data is scarce.

However, HeAR also has certain limitations. For example, linear probes cannot fully realize the performance potential of the model, some datasets are small and have class imbalance problems, and models such as HeAR are large and difficult to run on devices such as mobile phones.

The research team said that in the future they could consider fine-tuning the model or adding more features to improve performance, as well as collecting more data and improving data preprocessing methods. They could also consider studying model compression and quantization techniques to enable it to run on local devices.

AI-assisted disease diagnosis has great potential

From assisting doctors to independently diagnosing diseases, AI is increasingly being used in the medical field and has shown great potential.

In June this year, a research team from Imperial College London and the University of Cambridge trained the AI ​​model EMethylNET to identify 13 different types of cancer (including breast cancer, liver cancer, lung cancer, prostate cancer, etc.) from non-cancerous tissues by observing DNA methylation patterns, with an accuracy rate of up to 98.2%.

In July, an AI tool developed by a Boston University research team and its collaborators is expected to help us diagnose 10 different types of dementia (simultaneously), improving the accuracy of neurologists by more than 26%.

Recently, AI has also made a breakthrough in autism, the "invisible killer" of children. A multimodal data analysis AI model developed by a research team at the Karolinska Institutet can not only detect early signs of autism in children around 12 months old, but also has an accuracy rate of 80.5% for children under two years old. More importantly, the entire process only requires relatively limited information.

It is foreseeable that AI will help humans diagnose more diseases in the near future and bring more possibilities to the field of medical health.