Project Anemoi

Inspiration

The main inspiration of this project was hearing about experiences of COVID testing and the inefficiencies in the process. In most places, you have to head to a centralized location, get swabbed (which is usually a very uncomfortable experience), then wait for your results. That got me thinking, what if this process could be improved?

What it does

My project diagnoses COVID solely through spectrograms of people's breath. Why spectrograms? Spectrograms aren't very complicated and are accessible to anyone with a device and internet. This allows virtually anyone to get tested by COVID solely through their breath online. It's also very effective, with a 99% validation set accuracy rate.

How we built it

One of the most important parts of an ML model is the data because garbage data in means garbage results out. I got the dataset off of Kaggle, which is generally a good source for ML model data. After I got this data I used FastAI (a library for python) to create the model. I used a Resnet18 for transfer learning to increase the efficacy in training.

Challenges we ran into

The largest challenge I ran into was training the model. The actual data collection and data processing wasn't too bad, but when it came to training the model there were definitely some challenges. Some of the challenges I ran into when training the model included: overfitting, running out of memory, low validation accuracy, and sometimes the code just not working altogether. In the end, I managed to overcome these challenges and learned a lot in the process.

Accomplishments that we're proud of

I'm proud that I was able to finish this project and get it to actually work. To my knowledge, not many people have attempted to diagnose COVID with spectrograms before, making me step out of my comfort zone into the unknown, if I had any problems with my code of process, I couldn't simply search it up or look at a tutorial. I am proud that I was able to get it to such high accuracy as well as overcome the challenges I faced.

What we learned

In general, I learnt a lot about coding in the process, especially with training the model as that is where I faced the most challenges.

What's next for Project Anemoi

Currently, this project is solely running locally on a Jupyter Notebook. I wish to continue this project and make it a reality through a web app. This will allow anyone to diagnose themselves for COVID using my app solely by breathing into the mic of their device. To do this I will need to get some money to run my model on a server as well as get a spectrogram API integrated into the model.