Artificial intelligence (AI) can be used to detect COVID-19 infection in people’s voices via a mobile phone app, according to research to be presented Monday at the International Conference of the European Respiratory Society in Barcelona, Spain.
The AI model used in this research is more accurate than lateral flow/fast antigen tests and is cheap, fast and easy to use, which means that it can be used in low-income countries where PCR tests are expensive and/or difficult to distribute.
Ms. Wafa El-Jabawi, a researcher at the Institute of Data Science at Maastricht University in the Netherlands, told Congress that the AI model was accurate 89% of the time, while the accuracy of lateral flow tests varied widely depending on the brand. Also, lateral flow tests were significantly less accurate in detecting COVID infection in people who had no symptoms.
These promising results indicate that simple audio recordings and precise AI algorithms can achieve high accuracy in identifying patients with COVID-19 infection. Such tests can be offered at no cost and can be easily interpreted. Moreover, it enables remote virtual testing and has a response time of less than a minute. They can be used, for example, at entry points for large gatherings, allowing rapid screening of the population.”
Wafaa El-Jabawi, Researcher, Institute of Data Science, Maastricht University
COVID-19 infection usually affects the upper respiratory tract and vocal cords, resulting in changes in a person’s voice. Ms. El-Jabawi and her supervisors, Dr. Sami Simmons, a pulmonologist at Maastricht University Medical Center, and Dr. Visara Orofi, also from the Institute for Data Science, decided to investigate whether artificial intelligence could be used to analyze sounds in order to detect COVID-19.
They used data from the University of Cambridge’s COVID-19 Sounds app containing 893 audio samples from 4,352 healthy and non-healthy participants, 308 of whom tested positive for COVID-19. The app is installed on the user’s mobile phone, participants report some basic information about their demographics, medical history and smoking status, and then they are asked to record some respiratory sounds. This includes coughing three times, breathing deeply through the mouth three to five times, and reading a short sentence on the screen three times.
The researchers used a sound analysis technique called slope spectroscopy, which identifies various sound features such as loudness, strength, and contrast over time.
“In this way, we can deconstruct the many characteristics of the participants’ voices,” Ms. El-Jabawi said. “In order to distinguish the voices of COVID-19 patients from those who did not contract the disease, we built different AI models and evaluated which ones work best in classifying COVID-19 cases.”
They found that one model called long-term memory (LSTM) outperformed the others. LSTM is based on neural networks, which simulate the way the human brain works and learn basic relationships in data. It works in sequences, making it suitable for modeling signals collected over time, such as from sound, due to its ability to store data in its memory.
Its overall accuracy was 89%, its ability to correctly detect positive cases (true positive rate or ‘sensitivity’) was 89%, and its ability to correctly identify negative cases (true negative rate or ‘specificity’) was 83%.
“These results show a significant improvement in the accuracy of diagnosing COVID-19 compared to the latest tests such as the lateral flow test,” said Ms. El-Jabawi. “The lateral flow test has a sensitivity of only 56%, but a higher specificity rate of 99.5%. This is significant because it indicates that the lateral flow test misclassifies infected people as COVID-19 negative than our test. In other words, with The AI LSTM model, we may miss 11 out of 100 cases that will continue to spread the infection, while the lateral flow test will miss 44 out of 100 cases.
“The high accuracy of the lateral flow test means that only 1 in 100 people will be falsely told they have COVID-19 when, in fact, they were not infected, while the LSTM test will misdiagnose 17 out of 100 uninfected people. However, because this test is virtually free, it is possible to invite people to take PCR tests if the LSTM tests show positive.”
The researchers say their findings need to be validated in large numbers. Since the beginning of this project, 53,449 audio samples have now been collected from 36,116 participants and can be used to improve model accuracy and validation. They also perform further analysis to understand the parameters in the voice that affect the AI model.
In a second study, Sir Henry Glyde, a doctoral student at the University of Bristol’s School of Engineering, has shown that artificial intelligence can be harnessed via an app called myCOPD to predict when COPD patients may have an attack. Their disease, sometimes called acute exacerbation. Episodes of COPD can be very serious and are associated with an increased risk of hospitalization. Symptoms include shortness of breath, coughing and increased production of sputum (mucus).
“Acute exacerbations of COPD have poor outcomes,” he said. “We know that early identification and treatment of exacerbations can improve these outcomes and so we wanted to determine the predictive power of a widely used COPD application.”
myCOPD app is an interactive cloud-based application, developed by patients and clinicians and available for use by the UK’s National Health Service. It was established in 2016, and to date, more than 15,000 COPD patients use it to help them manage their disease.
The researchers collected 45,636 records for 183 patients between August 2017 and December 2021. Of these, 45,007 were for stable disease and 629 for exacerbations. Exacerbation predictors were generated one to eight days before the self-reported exacerbation event. Mr. Glide and his colleagues used this data to train AI models on 70% of the data and test them on 30%.
Patients were “more engaged,” who used the app weekly over months or even years to log symptoms and other health information, log medications, set reminders, and access the latest health and lifestyle information. Physicians can evaluate data via the Physician Dashboard, enabling them to provide oversight, shared management, and remote monitoring.
“The latest AI model we have developed has a sensitivity of 32% and a specificity of 95%. This means that the model is very good at telling patients when they are not about to have an exacerbation, which could help them avoid unnecessary treatment,” Mr. Glide said. When they are about to try one. Improving this will be the focus of the next phase of our research.”
Speaking ahead of the conference, Dr James Dodd, Associate Professor of Respiratory Medicine at the University of Bristol and project leader, said: “To our knowledge, this study is the first of its kind to model real-world data from COPD patients, extracted from a widespread therapeutic application. As a result, the predictive models of exacerbations resulting from this study have the potential to spread to thousands of COPD patients after further safety and efficacy tests. It would enable patients to have more independence and control over their health. This is also a great benefit for their physicians. as this system is likely to reduce patient dependence on primary care.In addition, better managed exacerbations can prevent hospitalization and reduce the burden on the health care system.Further study on patient involvement is needed to determine the level of accuracy acceptable and how the system will work. Exacerbation alerting in practice The introduction of sensing techniques may enhance monitoring and improve the predictive performance of models.”
One limitation of the study is the small number of frequent users of the app. The current form requires the patient to enter their COPD assessment test score, fill out their medication diary, and then report having an accurate exacerbation days later. Normally, only patients who interact most with the app, and use it daily or weekly, can provide the amount of data needed for AI modeling. In addition, because there are significantly more days that users are stable than when an exacerbation occurs, there is a significant imbalance between exacerbation and non-exacerbation data available. This results in more difficulty with models correctly predicting events after training on this unbalanced data.
“A recent partnership between patients, clinicians, and caregivers to prioritize research in COPD has found that the highest-rated question is how to identify better ways to prevent disease exacerbations. We have focused on this question, and will work closely with patients to design and implement the system,” concluded Mr. Glade.
The chair of the ERS Science Council, Professor Chris Breitling, is a senior researcher at the National Institute for Health and Care Research (NIHR) at the University of Leicester, UK, and was not involved in the research. He commented, “These two studies demonstrate the potential of AI and applications on mobile phones and other digital devices to make a difference in how diseases are managed. Having more data available to train these AI models, including appropriate control groups, as well as as validation in multiple studies, It will improve its accuracy and reliability. Digital health using AI models presents an exciting opportunity and is likely to impact healthcare in the future.”