Being able to decide in which children a Covid-Test must be performed from predicting from his or her symptoms if he is highly contagious to have Covid or not.
What it does
Create a model (XGBoost) to predict the presence or non-presence of Covid in children. Moreover, we have added the interpretability of this model to allow doctors to understand the important symptoms that affect to a child on having the Covid and be able to make decisions.
How I built it
We have created Jupyter notebooks (Python) and the more complex model is done in a Python script using Pytorch. The code is in the GitHub repository and the demo is in the notebook named demo_plots_model.ipynb
Challenges I ran into
The first problem was in the data, there were relatively few children with Covid and many unknown values. Our first approach was to make a complex model that required Deep Learning but it was not as we expected so we had to look for an alternative with a less complex model, but also using Machine Learning. Luckily this alternative has worked well and we have been able to get results.
Accomplishments that I'm proud of
The final results, although they are not in an interface, are quite clear for pediatricians to interpret them and help them make decisions. In addition, the requirements of the challenge are satisfied since a set of symptoms that characterize Covid and a set of symptoms that characterize Non-Covid are provided.
What I learned
We have learned how the interpretability of machine learning models works since until now we only looked at the prediction but we did not pay attention to what variables influenced this prediction. We have seen that for doctors this is essential and we have been able to develop it successfully.
What's next for SARS-CoV-2 in pediatric age
The next steps are fundamentally to be able to finish improving the complex model that we had planned to use to predict Covid and not Covid and make the interpretation. As it is a more complex model that works with Embeddings, we assume that it could have more accuracy than the XGBoost that we have finally used for the demo. It would also be interesting to create an interface or software so that doctors can interact with the plots and see for a specific patient which are the symptoms that most affect their probability of having Covid.