The Future of Healthcare: Google’s AI Uses Human Coughs to Diagnose Diseases

The future of healthcare: Google’s artificial intelligence using human cough to diagnose diseases
“Google AI is soon going to be able to use human coughs to diagnose diseases”, according to an article published on the website of HI-TECH+ magazine on March 22, 2024, by Alexander Sheremetyev (news editor of the portal).

The material published on HI-Tech+ portal tells us how a team led by Google scientists developed a machine learning tool to help detect and monitor health conditions by analyzing noises such as coughs and breaths. Our team was interested in this topic, so we decided to check the accuracy of this information and find out what its practical implications are.

According to the article, the HeAR AI system, which has been trained on millions of recordings of human noises, could one day help doctors diagnose diseases such as COVID-19 or tuberculosis, as well as assess how well someone’s lungs function.This is not the first time scientists have investigated sound as a biomarker for diseases. This idea became popular during the COVID-19 pandemic when scientists discovered that respiratory diseases can be detected through coughing.

What’s new about Google’s system called Health Acoustic Representations (HeAR) is the large dataset it was trained on and the fact that it can perform multiple tasks.

The experts involved in this project say it’s too early to tell if HeAR will become a commercial product. For now, the plan is to give interested researchers access to the model so they can use it in their own research. “As part of Google Research, our goal is to drive innovation in this emerging field,” says Sujay Kakarmat, a Google product manager based in New York who worked on this project. In the case of HeAR, the Google team adapted it to detect COVID-19, tuberculosis, and other characteristics such as human smoking. Since the model was trained on a wide range of human sounds, in order to fine-tune it, the researchers had to provide it with only very limited datasets labeled with these diseases.
On a scale where 0.5 represents a model that performs no better than random prediction and 1 represents a model that makes an accurate prediction every time, HeAR scored 0.645 for COVID-19 detection and 0.710 depending on which dataset it was tested on—higher performance than existing models trained on speech data. Regarding tuberculosis diagnosis, the indicator was 0.739.
The fact that the initial training data was so diverse—with different sound qualities and human sources—also means that the results are generalizable, according to Kakarmat. Ali Imran, an engineer at the University of Oklahoma in Tulsa, says that the sheer volume of data used by Google makes the study important. “This gives us confidence that it is a reliable tool,” he said.

The original article is available at nature.com. Before proceeding to further analysis of other articles, we decided to find out whether this was even possible and how the technology worked. Our expert, Ilya Valeryevich Pomerantsev, head of artificial intelligence at GLOBUS-IT, helped us to figure out some issues from the point of view of AI technology development.

Are there any studies confirming that Google’s AI can use human coughs to diagnose diseases such as COVID-19 and tuberculosis?

— Hypothetically, there is a possibility, because in Russia, there have been similar stories related to vibration diagnosis of the heart and speech recognition, which takes into account characteristics of a person’s national and linguistic affiliations. That is, artificial intelligence can identify anomalies in data very well. And if there are specific sounds in a covid cough, it will be easy for AI to find them and identify that this is a covid cough. However, it may also be tuberculous, hypothetically. If they declared this, I think they would have achieved some success. From the technological point of view, I do not see contradictions.

If we talk about detecting any serious disease, what prospects does this study have in its development? Will AI be able to detect COVID-19 stages?

— Again, COVID-19 was initially detected and proceeded quite heavily, but now it can be similar to the flu and so on. And here the question is, again, about that sample, about the data that researchers have. Most likely, with a change in the clinical picture, algorithms will need refining and modification. If we talk about the detection stage, a well-trained AI model requires a lot of statistical data, and in some difficult cases, this data is quite small.

A publication on the same topic was published in SNOB: “The Google neural network can detect disease by cough, specifically recognizing tuberculosis and COVID-1. The article says that an artificial intelligence system was trained on audio clips containing human sounds. To do this, more than 300 million short recordings of coughs and breathing from public videos on YouTube were used. Each recording was given a special label indicating one of the known diseases. Researchers then blocked some recordings to help the model learn how to predict missing fragments. Doctors will be able to use this neural network to diagnose diseases such as COVID-19 and tuberculosis, as well as monitor how well someone’s lungs are working.

AI has been trained to detect diseases by listening to coughs, according to Techinsider.

Unlike many other articles, however, this article explains how the machine was trained—by listening to 300 million recordings.

Most of the artificial intelligence tools being developed in this field are trained on audio recordings of coughing or breathing. The recordings are accompanied by medical information about the person making these sounds, such as a label indicating that the person had bronchitis at the time of recording. These tools allow you to associate features of sounds with data labels in the learning process, which is known as “learning with a teacher.”

Instead, Google researchers have used teacher-assisted learning, which relies on untagged data, using more than 300 million short audio fragments from coughing, breathing, and other human sounds collected from publicly available YouTube videos. Each clip was transformed into a visual representation of sound, the so-called spectrogram. The researchers then “blocked” some segments of spectrograms to help the model learn how to predict missing parts. This is similar to how large language models are taught underlying chatbots.

For example, ChatGPT was trained to predict the next word in a sentence using a variety of text examples. Using this method, researchers created what they called a basic model, which they said can be adapted to a wide variety of tasks. According to Yael Bensoussan, the field of medical acoustics or audiomics is promising: Acoustic science has been around for decades. The difference is that now, thanks to artificial intelligence and machine learning, we have means to collect and analyze very large amounts of data. Bensoussan is the co-head of a research consortium focused on the study of voice as a biomarker for health: “There is a huge potential not only for diagnosis but also for screening and monitoring. We can’t repeat the scan or biopsy every week, so voice is becoming an important biomarker for monitoring disease. It’s non-invasive and requires very little resources.”

The question of using this technology in medicine was extremely interesting, so we decided to ask our experts for more information. There are big doubts about such medical practices. Most likely, there will be a niche for household use, such as a fitness tracker to monitor your condition. Like bracelets do today, they measure your pulse, blood pressure, and other parameters. However, this is not the basis for diagnosing and treating yourself. Instead, we go to a medical facility where the doctor performs additional tests and makes a diagnosis, as commented by our expert, Ilya Valeryevich Pomerantsev.

It turns out that if this technology is implemented, it will require other validations. As a member of the Commission on the Implementation of AI Ethics under the AI Alliance umbrella, I believe the responsibility for diagnosis lies with doctors or other medical professionals. Therefore, we cannot say that this will be an autonomous application that can uniquely diagnose and treat.Society should get used to the fact that any new tool is poorly implemented and there are many difficulties with its implementation. If it were clear, perhaps we could say that this solution is cool and everyone would use it after a month. However, that is out of the question for the near future, so for now it’s at the level of experimentation and some household use. Further research is needed.

The director of the Institute of Clinical Medicine at N.I. Lobachevsky National State University, Dr. Natalia Grigorieva, clarified some questions about the use of AI technologies in diagnosing various diseases.

Based on the existing data, do you think it is possible to train a neural network so that it can automatically determine whether a person has COVID-19 or tuberculosis, for example, by coughing?

— In fact, artificial intelligence probably has nothing to do with it because there are analyses. analyses. Moreover, there is an express diagnosis, and there is a more accurate diagnosis of antibodies to coronavirus. And here, we just don’t need artificial intelligence. By the presence and level of antibodies in the blood, it is possible to judge the fact of infection with SARS-CoV-2 coronavirus in the past, even with an asymptomatic course, and to determine whether an immune response has been formed. The presence of immunoglobulin G antibodies will confirm the infection. Diagnostics is a different matter. This is where artificial intelligence comes into play. There are already works by Chinese and Russian scientists who used artificial intelligence in analyzing CT scan results more accurately to orient doctors and find differences.Indeed, AI could find differences between coronavirus pneumonia and the usual pneumonia we had before.

We discussed with an IT specialist that it would take at least 1,000 people with a confirmed diagnosis to program such a system. How accurate is this estimate, and is it accurate? What limitations might arise if such a system were to be implemented and large-scale data were uploaded to AI?

— Absolutely, that’s right. Based on the results of medical history information, at least 1,000 patients were collected. And it’s extremely important not to forget about sampling parameters for these patients. The more homogenous the sample is, the more accurate the result will be. If we look at young people, we limit ourselves to age as one thing, but there will also be the same symptoms in older people. In elderly patients, it’s necessary to take concomitant diseases into account. If they smoke, they always cough. AI won’t help you here at all because it won’t answer all your questions. It’ll only say how he understands—hhe coughs because of smoking or asthma, not illness. Back to selection: it’s best to limit to age. For example, patients under 40 should be included. Doctors always choose patients for research by writing, for example, the same number of men and women under 40 with no significant concomitants. That is, we must eliminate heart defects and severe congenital pathology if we are talking about COVID-19. Other respiratory diseases should also be eliminated.

Difficulties arise even with data analysis. Are there any prospects for introducing such technology in real medical practice? Do you see a personal prospect?

— Probably, you always want to create something new in medicine. If a platform appears where the patient uploads his symptoms and an audio file of a cough, the virtual assistant can give him a response: “The probability of this disease in you is, for example, 10% or 40%. In that case, why not create such a system? But there is no need for rest, I repeat, only in cough. It should be a combination of symptoms. There is certainly a benefit from AI helping doctors. They are our assistants.

Conclusion

Thus, the advancements in artificial intelligence, such as using a person’s cough for disease diagnosis with Google AI, can indeed become a significant tool in medical diagnostics and aid in the timely detection of illnesses. However, it is important to remember that artificial intelligence is not immune to errors, so doctors should always analyze the results and consider all factors related to the patient’s condition.

Experts and research data indicate that it is essential to maintain a balance between using artificial intelligence in medicine and professional evaluation by doctors. A key aspect is a comprehensive approach to diagnosis that takes into account all aspects of the patient’s health, lifestyle, heredity, and other factors.

Therefore, while artificial intelligence can be a useful tool in medicine, it is important to remember its limitations and not make final decisions based solely on algorithms. Professional evaluation by doctors remains a crucial element in providing high-quality medical care.

Authors: Arina Gorshenkova, Eva Mikheeva, Alina Kalashnikova, Alena Manina, Daniil Builov