AI for analyzing open-data related to the COVID-19 pandemic

This paper aims to improve the prediction models of COVID-19 evolution in Belgium. To do so, two approaches are explored. On the one hand, Forecastzoo, a collaborative science tool, is developed to provide a framework for any individual wishing to propose a model. On the other hand, specific models are proposed based on an underestimated virus prevalence estimator: the virus concentration in wastewater. The Forecastzoo platform is a website that takes the form of a challenge; the various models are put into a competition to see which ones stand out. Unfortunately, too few users submitted ideas, and no rigorous models emerged. Indeed, they were more of a gut feeling, hand-drawn predictions. Research on SARS-CoV-2 wastewater-based analysis is proposed in collaboration with Sciensano. This paper uses correlation to show a clear link and a 12-days lag between contamination and hospital admission. However, only two days are usable when including sampling times. This relationship is evident but not sufficient to map, without spatial information, the treatment plants to their respective municipalities. In addition to these analyses, these pages use the wastewater figures to prove the vaccine’s effectiveness and the bias of using positive cases as a prevalence indicator. Finally, a large part is dedicated to creating models to predict and understand the evolution of COVID-19 in Belgium. Machine learning (linear, polynomial, random forest, KNN), as well as statistical models (ARIMA, ARIMAX), fail to predict both the trend and the right range of hospitalizations (MSE above 10^4). It is due to the evolution of SARS-CoV-2 (variants) and population immunity (vaccination) through time. The alternative is to propose two new models that consider recent values more important than the old ones. The results are largely improved (MSE of 10^3). However, these models associate a weight to each station, and these weights vary enormously with time. Therefore, the results are abnormally good compared to the interpretations and could be due to overfitting.