Our inspiration is to help health professionals avoid overwhelming hospital infrastructure with non-critically ill patients seeking initial Covid-19 screenings through machine learning. Machine-learning has made significant progress in sound analytics for medical applications. Recent studies have shown that multiple diseases, such as Alzheimer’s and Parkinson’s, are detectible from respiratory audio. We believe distinguishing COVID-19 from other respiratory illnesses would allow large-scale systematic screening & isolation of patients, with no scaling limitations of cost or time. The model could reduce cases globally and prevent a new wave of the disease. We set out to develop an application that will allow the general public to submit their symptoms and receive a high-quality COVID-19 risk assessment, even among asymptomatic individuals, without creating an additional burden for global healthcare infrastructure.
What it does
Cov-Audio is a mobile COVID detection platform delivered through a web application. It uses machine learning to help potential Covid-19 patients screen their symptoms from home. By uploading a 6-second audio sample of their cough to check for respiratory illnesses and answering additional questions about their symptoms and medical history, users will receive a COVID-19 risk assessment. From this assessment, they can predict if they require further testing with a reasonable degree of certainty.
How we built it
We build two applications, one that allows users to answer screening questions and another that will enable users to upload a cough audio sample so that our application can use machine learning to detect respiratory ailments. The combined data of a COVID-19 risk score and respiratory illness detection can be used to determine the likelihood of someone having Covid-19.
Challenges we ran into
Collecting the data to build machine learning prediction models was challenging. There are few data sets containing cough samples with covid-status information. Additionally, we tried obtaining data from the University of Cambridge (Academia) but could not receive it in time. Additionally, researching and Implementing two machine learning models so that users can easily upload a cough audio-sample to check potential respiratory ailments and a questionnaire that would output a COVID-19 risk score. We implemented this system from scratch after researching the following academic papers: https://arxiv.org/pdf/2006.05919.pdf, https://digitalreality.ieee.org/images/files/pdf/COVID-AI-Diagnosis-using-Cough-Recordings.pdf. Building a web system so that users could determine their COVID-19 risk. Organizing all team members in different time zones (Central Standard Time, Philippine Time, Turkey Time, Geneva Time) to do all the needed tasks to train to machine-learning pipelines from big-data and design and implement web applications in just a week.
- If you have Risk factor using demographics model > 3 and a lung disease diagnosed then you have high risk of covid19
- If you have Risk factor using demographics model < 2 and no lung disease diagnosed then you have low risk of covid19
Focusing on the global health community, which is currently experiencing an overload of COVID-19 patients. (meeting with health professionals). We developed a screening system that can accurately determine a user’s COVID-19 risk, even among asymptomatic individuals. Implemented two machine-learning pipelines that determine respiratory ailment probability and COVID-19 risk probability from scratch with high accuracy among asymptomatic individuals. Using react native, we implemented a mobile application that takes in user-supplied information, records a user's cough, and displays the corresponding prediction.
What we learned
As a freshman in college, it was daunting to lead a large-scale project like this. Regardless, our team learned much more about the state of the art machine learning pipelines and sound wave analysis in great detail. Most of our group had no experience in machine learning in audio or full-stack development before working on this project. We learned that it was possible to assess users that have been infected with Covid-19 with a reasonable degree of certainty based on symptoms, without an official test. Using this knowledge and distributing it at scale through a web app does not require expensive testing equipment and takes little time compared to current methods. The scalability of this project could help screen users and save lives. Through this application, we can help combat the pandemic while ensuring better safety of the global health community and optimize the efficiency of COVID-19 testing, which would be especially helpful to a worldwide health infrastructure that is short on testing equipment.
What's next for Cov-Audio
There are several other features that we would like to add. For our models, additional data could help their accuracy. For screening, different health metrics such as body temperature and other pre-existing conditions, such as heart disease and diabetes, could help our model’s accuracy. We also would like to implement this not only on a web application but also in hospital infrastructure. For the future, we could obtain voice recording data from individuals with an official COVID-19 test to train a centralized model that is secured with blockchain. Our goal is to deploy this internationally so hospitals can implement this system to protect our global community.