COVID Wave Predictor

Mobile-first design
Works just as well on a desktop

What inspired your project?

I was inspired to create this project after looking into the code behind some of the more complex COVID models, such as the one by Imperial College. Given the enormity of such a model, I knew that I would not be able to replicate it in any way. Instead, I turned to machine learning and simplified the problem into a basic true/false classification as this seemed like it would be easier to attempt whilst still providing a useful outcome.

What problem does your project solve?

My project is designed to solve the data overload that can come with all the many simulations available for COVID. One can find plenty of data, but it requires digesting graphs showing projections and confidence intervals over time.

How does your project solve the problem?

My project provides a simple metric that estimates the probability of a COVID wave within 14 days. It’s a metric familiar to anyone who’s looked at an election forecast yet still provides useful information. Much like the weather forecast allows one to know whether an umbrella is necessary, this probability allows one to plan ahead, potentially shopping ahead so as to avoid COVID at its worst.

What technologies did your project use? Include programming languages, libraries, and any external tools.

The neural network and data gathering for it is written in Python 3.8 and uses Tensorflow v2.3 with CUDA. NumPy is also used for working with the data. The website is written in pure HTML/CSS/JS, although D3.js and Topojson are used for handling the data.

What challenges did you run into?

When I was first training the neural network I failed to take into account bias in the data, allowing the network to be more accurate simply by choosing as close to zero as possible each time. Working around that wasn’t too hard but did require recognizing the problem and then cutting off some data to balance out the categories. Once I did that the neural network then had the same accuracy as guessing, which worried me quite a bit. I wasn’t able to solve that with changes to the model, so I doubled the amount of data fed to the neural network for each prediction, allowing me to get a more useful accuracy.