Outlier detection in digital PCR wastewater time series data

Abstract

Efforts to control the spread of disease in a population require data collection on its prevalence over time. Analysing this data can give insights into how disease spreads and provide feedback on the effectiveness of intervention methods. Wastewater-based epidemiology (WBE) involves sampling wastewater from a treatment plant to collect data on disease prevalence in the population served by a particular sewage catchment. This data can be used to estimate metrics such as effective reproduction numbers (Rt) which can inform public health decision making. WBE contains several sources of uncertainty and the resulting data are noisy, requiring statistical analysis to gain meaningful measurements. This includes normalisation to account for population or system dynamics, and smoothing of signal to understand trends. When data containing outliers are used to inform models about the spread of disease, these outliers can have an impact on the model analysis. The identification and removal of such data points could improve modelling results and provide a more reliable basis for decision making for disease control. The Wastewater Monitoring Laboratory, part of the Swiss Federal Institute of Aquatic Science and Technology, collects data on the concentration of four different respiratory pathogens in the wastewater from fourteen different treatment plants around Switzerland using digital PCR (dPCR). This data set was used to develop an outlier detection method able to deal with multiple challenges presented by the specific nature of wastewater-based dPCR data, thereby allowing for efficient outlier detection. The outlier detection method was then used to study the impact of outliers on estimating Rt from wastewater viral concentration data. This highlighted the importance of robust model design and provided insight into the impact an outlier can have on estimates depending on when it occurs during an infection wave. Additionally, time series of wastewater data were simulated, providing a resource for further method evaluation.