COVID has caused 400k deaths in the US alone and brought the economy to a standstill. It has also introduced a number of challenges worldwide.
The primary reason why COVID is so intractable is because there is a significant delay between infection and diagnosis. Testing is not scalable, sometimes inaccurate, and introduces more risk for people who go to a public place to get tested. It is also very expensive and most countries have faced difficulties buying more test kits.
Simply put, we need a solution that can work everywhere and requires minimal scaling. It must also be based on sound science and identify at risk individuals effectively.
Enter COVID CoughNet.
COVID CoughNet is a deep network that recognizes the differences between COVID Coughs and non COVID coughs. It takes as input raw audio files, and throws out a diagnosis of whether that cough comes from a COVID infected individual.
We believe this solution is far more scalable than contact testing. it's much easier to install microphones in public places than it is to invite people to get tested. It also reduces the latency between time of infection and diagnosis significantly. People can use the network to get a better sense of their diagnosis from home, without needing to go out. This reduces transmission to other individuals as well.
We achieved state of the art results among open source implementations at about 83% accuracy. This is comparable to a study done by Cambridge in July, achieving 80% accuracy. We used CNN, RNN, LSTM and ensemble models to achieve this result. We also compared our results with the new vision transformer architecture which produced worse but promising results.
Challenges we ran into
- Training the models was time-consuming, but we got it done!
- Collecting data was difficult given the reduced amount of available open-source data
Accomplishments that we're proud of
- Achieving SOTA, on par with Cambridge University researchers
- Developing the data pipeline was initially difficult, but once we had it, it was much easier to develop the CNN and Vision Transformer
What's next for the COVID CoughNet?
We need to build a webapp where people can record their cough and get an instant diagnosis. Of course, this won't be an official medical recommendation -- but it will help people limit their interactions and stop the spread of COVID.
Ideally, we could develop our web app in such a way as to anonymize collected data for future researchers to use, unlike several other existing implementations. Furthermore, we could develop this web app to be open-source, and therefore it could be extended and used by municipalities across the globe.
This is the ultimate goal for our project. This way, we can improve our models more.