Epidemiological model can forecast COVID-19 outbreaks from wastewater-based surveillance in rural communities

Abstract

Wastewater has emerged as a crucial tool for infectious disease surveillance, offering a valuable means to bridge the equity gap between underserved communities and larger urban municipalities. However, using wastewater surveillance in a predictive manner remains a challenge. In this study, we tested if detecting SARS-CoV-2 in wastewater can forecast outbreaks in rural communities. Under the CDC National Wastewater Surveillance program, we monitored the SARS-CoV-2 in the wastewater of five rural communities and a small city in Idaho (USA). We then used a particle filter method coupled with a stochastic susceptible-exposed-infectious-recovered (SEIR) model to infer active case numbers using quantities of SARS-CoV-2 in wastewater. Our findings revealed that while high daily variations in wastewater viral load made real-time interpretation difficult, the SEIR model successfully factored out this noise, enabling accurate forecasts of the Omicron outbreak in five of the six towns shortly after initial increases in SARS-CoV-2 concentrations were detected in wastewater. The model predicted outbreaks with a lead time of 0 to 11 days (average of 6 days +/- 4) before the surge in reported clinical cases. This study not only underscores the viability of wastewater-based epidemiology (WBE) in rural communities-a demographic often overlooked in WBE research-but also demonstrates the potential of advanced epidemiological modeling to enhance the predictive power of wastewater data. Our work paves the way for more reliable and timely public health guidance, addressing a critical gap in the surveillance of infectious diseases in rural populations.

Keywords: COVID-19; Rural communities; SARS-CoV-2; SEIR; Wastewater-based epidemiology (WBE).

Figure 1:. Diagram showing steps of the particle filter method we use to determine the number of active cases from the wastewater titers of SARS-CoV-2.

The particle filter is initialized using the first measurement of virus concentration in the wastewater. We generate a distribution of the possible number of infections in the community and sample many (50000) values from this distribution. These values are used as the possible number of exposed individuals ($E$) on day 1 (top right graph). Each of these values also gets a potential number of Susceptible ($S$), Infected ($I$), and Recovered ($R$) individuals. Each set of values ($S$,$E$,$I$,$R$) is called a particle. The darker dots in the diagram signify a higher number of particles with that value of $E$. We apply one step of the stochastic SEIR model to each particle to predict the number of infections on the next day (bottom left graph). The measurement of the virus in the wastewater on the next measurement is used to determine which particles are more likely than others. Less likely particles are filtered out using a systematic resampling procedure and replaced with more likely particles (bottom right graph).

Figure 2:. Trend in SARS-CoV-2 in wastewater mirrors the dynamic of the COVID-19 outbreak in rural areas.

Each panel represents a city. In each panel, the bar graph shows the time series of the COVID-19 clinically confirmed cases at the specimen collection dates and the second graph shows the measured concentration of SARS-CoV-2 (green dots) with the 7-day moving average (red line). Vertical dash lines represent the estimated start of the outbreak using either the cumulative sum of the copies per day of the N1 target or the cumulative sum of COVID-19 clinically confirmed cases, determined by the Piecewise regression model. Delta shows the difference of days between predicted dates from wastewater-based detection of SARS-CoV-2 and clinically confirmed COVID-19 cases. Cities are ordered by population size (largest on the top left and smallest on the bottom right).

Figure 3:. Daily quantities of SARS-CoV-2 tend to be more spread as the city population get smaller.

Panel A) shows the distribution of the copies per day of the SARS-CoV-2 on the log scale over the sampling period at each site ordered by city size, detailed in panel B. Note: bin width = 1/30. Dots on the x axis show the samples where N1 was under the detection limit (SC: n = 0, R1: n = 8, R2: n = 8, R3: n = 4, R4: n = 6, R5: n = 7). Panels B) shows the population size of the cities sampled, and C) shows the log scale sample variance measured for each city calculated using the log10 of copies per day of the SARS-CoV-2. This essentially shows that the magnitude of the estimate is less consistent as population size gets smaller (i.e., more stochasticity).

Figure 4:. Susceptible-exposed-infectious-recovered model can forecast cases in the early stage of a COVID-19 outbreak.

A) SEIR model framework depicting a population in green with infected people in red. The SARS-CoV-2 shed by a fraction of the exposed population is measured in the wastewater collected at the WWTF. This titer is integrated into a Susceptible (S), Exposed (E), Infected (I), and Recovered (R) model to estimate the number of exposed individuals E. B) Left white side contains known data at the time of the forecast where the blue lines show the fitted predicted active cases from wastewater up to the beginning of the outbreak, and the blue shade shows the data not yet observed at the time of forecast whereas the red lines are active cases forecasted. Vertical dashed lines represent the estimated start of the outbreak based on clinically confirmed COVID-19 cases. Using the wastewater data, the model forecasted the start of the outbreak between 0 to 11 days earlier than the onset of the increase in clinical confirmed cases. 95% confidence intervals are shown by the gray bars. Dots show the active cases determined as the 11-day moving sum of the clinically confirmed cases. Since the mean infectious period from fitting data was 10.88 days, we determined the actual active cases as the 11-day moving sum of new clinically confirmed cases. Breakpoints between fitted and forecast values were chosen to be two days after the start of the outbreak, determined by the Piecewise regression model. Cities are ordered by population size (largest on the top left and smallest on the bottom right).

Figure 5:. Receiver operating characteristic (ROC) curves for a synthetic data set.

A plot showing the number of true increases against false increases for predicted case counts 1,3,5,7,9,11,13, and 15 days beyond the current measurements. A true increase is counted when there was an increase in cases and the model predicted a greater than α probability of an increase. A false increase is counted when there was no increase in cases but the model predicted a greater than α probability of increase.

Figure 6:. Evaluating SEIR model predictability for an emerging COVID-19 outbreak.

Box plot showing the distribution of measured area under the curve (AUC) when computing 50 receiver operating characteristic (ROC) curves when true positive rate is plotted as function of the false positive rate for prediction forecasted from one to 15 days. A random classifier, which represents the outcome if the model randomly picks predictions, has an AUC of 0.5. The further away the curve is from the one of the random classifiers, the higher the AUC and the better it illustrates the ability of the model to forecast a trend, with the 1 representing the highest accuracy corresponding to 100% positive rate and 0% false negatives. In general, for a diagnostic test to be able to discriminate patients with and without a disease, the AUC must be above 0.5. Values between 0.7 and 0.8 are considered to be ‘fair’ or acceptable (Mandrekar, 2010; Nahm, 2022).