Spatio-temporal real-time predictions of COVID-19 test positivity in Uppsala County, Sweden – a comparative approach.
Vera van Zoest , Georgios Varotsis, Uwe Menzel, Anders Wigren, Beatrice Kennedy, Mats Martinell, Tove Fall
Abstract Existing spatio-temporal COVID-19 prediction models have varying accuracies, lack high geographical resolution and mainly focus on number of new cases. Our aim was to predict trends in test positivity, a powerful measure to assess the need for increased local test capacity, at a high resolution. We use a full year of information on direct and indirect indicators of transmission e.g. mobility data, number of calls to the national healthcare advice line and vaccination coverage (June 29, 2020 – July 4, 2021) from Uppsala County, Sweden. Four models were developed for a one-week-window based on Gradient Boosting (GB), Random Forest (RF), Autoregressive Integrated Moving Average (ARIMA) and Integrated Nested Laplace Approximations (INLA). Three models (GB, RF and INLA) outperformed a naïve baseline model after data from a full pandemic wave became available. An ensemble model of these three models slightly improved the average Root Mean Square Error to 0.039 as opposed to 0.040 for GB, RF and INLA, 0.055 for ARIMA and 0.046 for the naïve model. Our findings highlight the importance of collecting a wide variety of data, in order to make early and accurate spatio-temporal predictions on the occurrence of areas with increased transmission and insufficient testing.
Keywords: COVID-19, SARS-CoV-2, test positivity, random forest, gradient boosting, INLA, ARIMA, Sweden