Abstract
Accurate epidemiological classification guidelines are essential to ensure implementation of adequate public health and social measures. Here, we investigate two frameworks, published in March 2020 and November 2020 by the World Health Organization (WHO) to categorise transmission risks of COVID-19 infection, and assess how well the countries' self-reported classification tracked their underlying epidemiological situation. We used three modelling approaches: an ordinal longitudinal model, a proportional odds model and a machine learning One-Rule classification algorithm. We applied these models to 202 countries' daily transmission classification and epidemiological data, and study classification accuracy over time for the period April 2020 to June 2021, when WHO stopped publishing country classifications. Overall, the first published WHO classification, purely qualitative, lacked accuracy. The incidence rate within the previous 14 days was the best predictor with an average accuracy throughout the period of study of 61.5%. However, when each week was assessed independently, the models returned predictive accuracies above 50% only in the first weeks of April 2020. In contrast, the second classification, quantitative in nature, increased significantly the accuracy of transmission labels, with values as high as 94%.