A comparison of two forecasting models for COVID-19 hospitalizations using wastewater concentration data

Wastewater surveillance has become a prominent part of public health efforts to track circulating pathogens and has been incorporated into disease forecasting models. We compared two recent wastewater forecasting models for forecasting COVID-19 hospitalizations using SARS-CoV-2 concentrations in wastewater: a generalized linear mixed model (GLMM) and a Bayesian mechanistic model known as wwinference. We retrospectively produced 1-week ahead forecasts using these two conceptually different models across 10 regions in New York state, from September 2022 to mid-April 2024. We compared the performance of forecasts produced from the two models fit with wastewater and clinical data to a version of each model fit to only clinical data. Models were scored against observations using Continuous Ranked Probability Scores for each forecast (n = 363 forecasts). Of the two models, the GLMM showed improved forecast performance across space and time when evaluated on a natural scale compared to the wwinference model, but no significant difference in forecast performance was found on a log-scale, which evaluates the relative rather than absolute error. Consequently, either model could be used to forecast COVID-19 hospitalizations. Including wastewater concentrations in these models at these spatial and temporal scales did not provide additional forecasting benefit beyond just clinical data. The performance of a wastewater-only model relative to simple null models; however, demonstrates that there is a clear forecasting signal present in wastewater, thus we think the lack of change is likely due to the strong signal from the clinical measures included in the model at this spatial and temporal scale.