An exploration of challenges associated with machine learning for time series forecasting of COVID-19 community spread using wastewater-based epidemiological data

Wastewater-based epidemiology (WBE) has gained increasing attention as a complementary tool to conventional surveillance methods with potential for significant resource and labour savings when used for public health monitoring. Using WBE datasets to train machine learning algorithms and develop predictive models may also facilitate early warnings for the spread of outbreaks. The challenges associated with using machine learning for the analysis of WBE datasets and timeseries forecasting of COVID-19 were explored by running Random Forest (RF) algorithms on WBE datasets across 108 sites in five regions: Scotland, Catalonia, Ohio, the Netherlands, and Switzerland. This method uses measurements of SARS-CoV-2 RNA fragment concentration in samples taken at the inlets of wastewater treatment plants, providing insight into the prevalence of infection in upstream wastewater catchment populations. RF's forecasting performance at each site was quantitatively evaluated by determining mean absolute percentage error (MAPE) values, which was used to highlight challenges affecting future implementations of RF for WBE forecasting efforts. Performance was generally poor using WBE datasets from Catalonia, Scotland, and Ohio with ‘reasonable’ or better forecasts constituting 0 %, 5 %, and 0 % of these regions' forecasts, respectively. RF's performance was much stronger with WBE data from the Netherlands and Switzerland, which provided 55 % and 45 % ‘reasonable’ or better forecasts respectively. Sampling frequency and training set size were identified as key factors contributing to accuracy, while inclusion of too many unnecessary variables (or e.g., flow data) was identified as a contributing factor to poor performance. The contribution of catchment population on forecast accuracy was more ambiguous. This study determined that the factors governing RF's forecast performance are complicated and interrelated, which presents challenges for further work in this space. A sufficiently accurate further iteration of the tool discussed within this study would provide significant but varying value for public health departments for monitoring future, or ongoing outbreaks, assisting the implementation of on-time health response measures.