Wastewater as an Early Indicator for Short-Term Forecasting COVID-19 Hospitalization in Germany

doi:10.21203/rs.3.rs-5128866/v1

Download PDF

Research Article

Wastewater as an Early Indicator for Short-Term Forecasting COVID-19 Hospitalization in Germany

https://doi.org/10.21203/rs.3.rs-5128866/v1

This work is licensed under a CC BY 4.0 License

You are reading this latest preprint version

Background

The COVID-19 pandemic has profoundly affected daily life and posed significant challenges for politics, the economy, and the education system. To better prepare for such situations and implement effective measures, it is crucial to accurately assess, monitor, and forecast the progression of a pandemic. This study examines the potential of integrating wastewater surveillance data to enhance an autoregressive COVID-19 forecasting model for Germany and its federal states.

Methods

We explore the correlations between viral load measured in wastewater and COVID-19 hospitalization. The study compares the performance of autoregressive models, including Random Forest regressors, XGBoost regressors, ARIMA models, linear regression, and ridge regression models, both with and without the use of wastewater data as predictors. For decision tree-based models, we also analyze the performance of fully cross-modal models that rely solely on viral load measurements to predict COVID-19 hospitalization rates.

Results

Our findings suggest that wastewater data can serve as an early warning indicator of impending trends in hospitalization at a national level, as it shows a strong correlation with hospitalization figures and tends to lead them by six to seven days. Despite this, including wastewater data in the prediction models did not significantly enhance the accuracy of COVID-19 hospitalization forecasts. The ARIMA model emerged as the best-performing model, achieving a Mean Absolute Percentage Error of 4.69%. However, wastewater viral load proved to be a valuable standalone predictor, offering a cost-effective and objective alternative to classical surveillance methods for monitoring pandemic trends.

Conclusion

This study reinforces the potential of wastewater surveillance as an early warning tool for COVID-19 hospitalizations in Germany. While strong correlations were observed, the integration of wastewater data into predictive models did not improve their performance. Nevertheless, wastewater viral load serves as a valuable indicator for monitoring pandemic trends, suggesting its utility in public health surveillance and resource allocation. Future research should explore broader applications of wastewater data for other pathogens and in conjunction with diverse data sources.

The COVID-19 pandemic has significantly disrupted global society, impacting healthcare systems, economies, and daily life on a worldwide scale. This global crisis has revealed weaknesses in healthcare infrastructure and underlined the necessity for reliable predictive tools to effectively manage future pandemics (1). Accurate forecasting of incident cases and hospitalization is essential for optimizing resource allocation or decision making in particular for potential future pandemics. Many approaches in this regard have been published during the COVID-19 pandemic or even before (2).

One approach to pandemic monitoring is wastewater surveillance, which has emerged as a promising method for early detection of viral spread. Wastewater data can provide a non-invasive, community-level indicator of infection prevalence, offering a lead time advantage over pandemic-related metrics (3, 4). Previous studies have demonstrated the utility of wastewater surveillance in tracking COVID-19 trends. It was shown that wastewater data can be utilized for early warning of pandemic outbreaks and new emerging variants of concern (5–8), and it was also used as a predictor for incident cases and hospitalization (9–11). Some of these studies also investigated these aspects in specific countries, including Germany (5, 7, 8, 11). However, these studies either focused on specific regions or failed to explore the potential predictive power of wastewater viral load in the context of short-term pandemic forecasting, particularly concerning hospitalization.

This study aims to address this gap by examining the relationship between viral loads in wastewater and COVID-19 hospitalization across Germany at a regional and national level. We evaluate the potential of integrating wastewater data into various autoregressive models, such as Random Forest, XGBoost, ARIMA, linear regression, and ridge regression models, to improve their predictive accuracy for forecasting hospitalization in Germany. Furthermore, we investigate the efficacy of models that utilize only wastewater data in predicting hospitalization.

2.1 Surveillance Data

We utilized two primary datasets to investigate the relationship between wastewater surveillance data and COVID-19 hospitalization in Germany.

The first dataset contains the daily incident hospital admissions, which were recorded by the Robert Koch Institute (RKI). In this dataset the hospital admissions are represented as the 7-day hospitalization normalized to 100k people (rate) (12). In the following we will refer to this data as hospitalization rates. One key metric frequently utilized during the pandemic was the number of reported incident cases. However, this metric is heavily influenced by the availability of testing, testing strategies, and underreporting (often referred to as the 'dark figure'). Therefore, we considered hospitalization to be a more reliable metric, and excluded incident cases from the subsequent analysis.

The second dataset consists of wastewater viral load measurements collected from 136 sewage treatment plants, distributed across Germany, participating in the AMELAG (Abwassermonitoring für die epidemiologische Lagebewertung) project (13). The name translates to “wastewater monitoring for the epidemiological situation assessment”. The assessment of this wastewater data involves several key steps. Wastewater samples are regularly collected from treatment plants, typically multiple times a week. These samples are then analyzed in laboratories using techniques like quantitative polymerase chain reaction (qPCR) to detect and quantify viral RNA. To ensure accurate concentration estimates, the viral load measurements are normalized based on the flow rate of the treatment plant. Next, data from multiple treatment plants are aggregated to provide a regional overview, using geometric means to smooth out fluctuations. Daily estimates from the weekly data are computed by applying locally estimated scatterplot smoothing (LOESS) (14).

Additionally, we wanted to assess the relationship between wastewater viral load and recorded diagnoses of respiratory diseases. For this we used data which was gathered by docmetric (https://docmetric.de/), a subsidiary company of CompuGroup Medical (CGM). Docmetric collected anonymized health data from approximately 3000 general practitioners. Part of these data are ICD-10 coded diagnoses. We selected diagnoses related to respiratory diseases and used ICD-10 parent categories to aggregate them to four categories: “U0.7.1 COVID-19”, “J00-J06 Acute upper respiratory infections”, “J10-J18 Influenza and pneumonia”, and “J20-J22 Other acute respiratory infections” (15).

2.2 Correlation Analysis

Our first step was to investigate the correlation between viral load in wastewater data and country-level hospitalization rates. For a prior visual inspection we normalized both the hospitalization rates and wastewater data using a min-max normalization. After visually inspecting the time series, we employed correlation analyses to evaluate the relationship between wastewater viral load and hospitalization rates. For this we calculated Spearman (16) correlation coefficients. Spearman coefficients evaluate monotonic relationships, which can be particularly useful when marginal distributions of individual variables do not meet the assumptions of normality. Our hypothesis is that there exists a consistent time lag between the two variables, with trends in wastewater viral load preceding similar trends in hospitalization rates. This is based on the premise that viral load in wastewater increases or decreases several days before individuals exhibit symptoms severe enough to require hospitalization. Consequently, we performed a lagged correlation analysis to explore this temporal relationship. We shifted the wastewater and hospitalization time series relative to each other to identify the highest positive correlation and, consequently, the time lag at which the wastewater data could most effectively predict hospitalization.

2.3 Forecasting Models

In the following part we wanted to assess whether the inclusion of the wastewater viral load into otherwise autoregressive models, will significantly improve the model forecasting performances. Given the expectation that both the hospitalization rates as well as the wastewater viral load exhibit exponential growth and decay behavior, we applied a log-transformation to both time series. The underlying idea is that this transformation yields piecewise linear behavior.

We employed three classical models: linear regression, ridge regression and ARIMA, and two decision tree based ones: XGBoost and Random Forest. These methods were chosen because they proved useful in our previous study (17). In the following we briefly provide more details.

Linear Regression and Ridge Regression

For our linear regression model, we employed the LinearRegression class from scikit-learn (18), specifically version 1.3.2. Since the optimal solution is derived using the Ordinary Least Squares (OLS) method, there was no need or possibility for hyperparameter optimization. In this model, the dependent variable was the hospitalization rate, while the independent variable comprised integers ranging from 1 to 14. Across the 72 windows, we trained the model using data from days 1 to 7 and made predictions for days 8 to 14. In the non-autoregressive scenario, daily wastewater data was included as a second independent variable. To ensure accurate forecasting, all independent variables had to be known, leading us to compensate for the time lag between wastewater and hospitalization data by shifting the wastewater data forward by seven days. This adjustment allowed the fitting window to utilize wastewater data from the previous week and the forecasting window to use the current week's data. To prevent overfitting, particularly when incorporating the wastewater viral load, we applied ridge regression. Also known as L2 regularization, ridge regression introduces an additional term to the loss function in linear regression, guided by a tunable parameter λ (19).We utilized the corresponding class in scikit-learn for this purpose. The parameter was fine-tuned on the left-out training data (see section 2.4), within a range of (0, 10].

ARIMA

Autoregressive Integrated Moving Average (ARIMA) models leverage the statistical properties of stationary data and are widely used for time series forecasting (20). A stationary time series is characterized by the absence of trends and consistent variation around its mean, allowing for the extraction of short-term random time patterns for forecasting purposes. In this study, we utilized a non-seasonal ARIMA model tailored for short-term periods that are not anticipated to exhibit seasonal effects. The ARIMA models in this context rely on three key parameters:

p: the number of autoregressive terms
d: the degree of differencing applied to achieve stationarity
q: the number of lagged forecast errors

The general forecasting equation for ARIMA is defined as follows:

$$\:\widehat{{y}_{t}}=\:\mu\:+{\phi\:}_{1}{y}_{t-1}+...+{\phi\:}_{p}{y}_{t-p}-{\theta\:}_{1}{e}_{t-1}-...-{\theta\:}_{q}{e}_{t-q}\:$$

In this equation, ŷ represents the forecast, calculated as the deviation from the mean µ of a stationary time series, with φ denoting the slope parameters for the p previous values y, and θ representing the q moving average parameters associated with autocorrelation errors e. This model learns to predict future values based on the mean of a stationary time series, adjusted for autocorrelation errors and lagged periods. To ensure stationarity, the differencing technique is applied, which involves calculating the differences between consecutive values in the time series (21). This transformation often leads to stationarity, particularly in first or second order. To identify the optimal parameters (p, d, q), we utilized the auto-ARIMA functionality from the pmdarima library version 1.8.5 (22). While ARIMA models work by nature autoregressive, it is possible to include exogenous features to the model. In so-called ARIMAX models a term βX is added to the ARIMA equation, where X is the exogenous feature, in our case wastewater viral load, and β the corresponding coefficient that is estimated by the model (23).

XGBoost and Random Forest

Both Random Forest and eXtreme Gradient Boosting (XGBoost) are decision tree-based approaches, but differ significantly in their training algorithms. Random Forest creates an unweighted ensemble of decision trees, which is trained in parallel on different subsets of the data using bagging, averaging the predictions (24). In contrast, XGBoost constructs its decision trees sequentially, correcting the residual errors from the previously trained weighted ensemble using gradient descent (25). Both models are widely used for tabular data and have also proven effective in time series forecasting (26–31). Since these models rely on decision trees, they can only extrapolate based on previously observed training data. When predicting values outside the range of the training data, they tend to predict the average or maximum of the observed values. To address this, we applied the differencing technique (21). That means we calculate from each measured datapoint to the next the slope of a local tangent. The models are hence tasked to learn and predict the relative change from one datapoint (here day) to another. Before evaluation the forecast was back transformed by using the cumulative sum. For Random Forest, we used the scikit-learn library version 1.3.2 RandomForestRegressor class (18), and for XGBoost, we employed the XGBoost library version 1.7.3 (25). For hyperparameter tuning we employed a blocked time series cross validation (32) using optuna version 3.2.0 (33) and the same possible hyperparameters (see Table S1) as in our previous study (17).

For both models we further wanted to test two things. First, we wanted to assess whether the inclusion of federal state level data for training the models would improve the models’ performances in forecasting the hospitalization rates on a country level. Since wastewater data was only available at the country level and by community, we aggregated daily wastewater data for each federal state by averaging measurements from sewage treatment plants within the state. This data was normalized based on the number of connected citizens, as specified in the dataset for each participating city. This approach aimed to improve prediction accuracy by increasing the dataset's granularity. However, only five federal states had complete and usable datasets for the period under examination. For these five federal states we used both the regional wastewater viral load and regional hospitalization rates. Secondly, we tested how well these models performed if they were only given the wastewater viral load for forecasting the hospitalization rates, later referred to as cross-modal models.

2.4 Sliding Window Approach for Model Training and Testing

To account for potential changes in conditions during the study periods - such as new safety regulations, vaccines, or virus variants - and to provide more test windows for a more reliable evaluation of the model’s performance, we employed a sliding window approach for both training and testing data (Fig. 1).

In this approach, we predict the hospitalization rate for a seven-day period (prediction window) based on a preceding set of data points (context window). The size of the prediction window is set to seven days because of the time lag between wastewater viral load and hospitalization rate observed via correlation analysis (see Results section). After each prediction, both windows are shifted forward by seven days, ensuring no overlap between successive prediction windows. Applying this method throughout the entire period for which both wastewater and hospitalization data are available results in 72 testing windows.

This approach, which is similar to one used in our previous study (17), simulates a scenario where the models are trained on past data, as if they were being used at the time the data was originally collected.

Given that the models operate in fundamentally different ways, we assigned them context windows of varying sizes (Fig. 3). For tree-based models like XGBoost and Random Forest, which benefit from larger context windows, we used a window size of 70 days to be consistent with a previously published study (17). Additionally, for model training we shifted the time series by 7 days to obtain the corresponding target vector. Similarly, ARIMA, which identifies patterns in the data, also requires a larger sample size, so it was assigned a 70-day context window as well. Here shifting the data is not necessary as ARIMA employs a different learning strategy (see methods). While linear regression generally improves with more data, using a large context window can smooth out smaller trends. In fact, using a 70-day window for linear regression in this case could be counterproductive, as it may obscure short-term trends relevant to the seven-day forecast. Instead, we opted to fit the linear regression using only the last seven days of training data to better capture recent changes. For a valid comparison between the models, it is essential that they forecast the same days and have the same number of forecasting windows. To achieve this, we introduced offsets at the beginning of the time series, ensuring that the initial test forecasting windows are aligned across all three models.

2.5 Model Evaluation and Comparison

To assess the performance of the models on the testing windows, we used the mean absolute error (MAE) (34) and the mean absolute percentage error (MAPE) (35) as metrics:

$$\:MAE\:=\frac{1}{n}{\sum\:}_{i=1}^{n}\:\left|{Y}_{i}-{\widehat{Y}}_{i}\right|\:\:;\:MAPE\:=\frac{1}{n}{\sum\:}_{i=1}^{n}\left|\frac{{Y}_{i}-{\widehat{Y}}_{i}}{{Y}_{i}}\right|*100\:$$

where Y represents the observed values, Ŷ represents the predicted values, and n is the number of data points, which in our case corresponds to the length of the testing window (7 days). The MAPE expresses the deviation of the prediction from the observed data as a percentage, making it a more intuitive measure compared to the MAE. However, due to this normalization the MAPE is high for deviations in small scales, e.g. 50% if the predicted value is 1 but the observed value is 2. Therefore we decided to record both the MAE and the MAPE. All models were evaluated on the original scale (non log-transformed).

In order to test for statistical differences between two models, we declare them to be statistically different if their 95% confidence interval of their mean MAPE does not overlap.

3.1 Wastewater Viral Load as an Early Warning Indicator

In Fig. 3 we plotted the normalized hospitalization rates and the LOESS interpolated and normalized daily wastewater viral load for the time period of June 2022 to March 2023. It can be seen that the time series follow the same up- and down-trends. The plot also reveals that the relation between the scales of both time series for hospitalization and wastewater viral load changed throughout the observed period. While hospitalization showed its maximal peak at the end of September 2022, wastewater viral load peaked most in March 2023. Next we calculated the correlation coefficients between hospitalization rates and viral load. The results can be seen in Table 1. Spearman analysis suggests a high correlation of around 80%. Comparing the results for the weekly and daily viral load shows similar results, daily viral load had a smaller confidence interval, though. Additionally, we wanted to derive the time lag with the highest correlation, i.e. the number of days the time series need to be shifted against each other to achieve maximal correlation. Here we found the best lag to be at around 6 to 7 days. In Fig. 4 we also show how Spearman correlation coefficient changes with different number of days shifted. Here it can be seen that a negative lag, i.e. shifting the hospitalization more towards the future, drastically reduces the correlation while a positive lag increases the correlation up to its maximum at 6 to 7 days and then slowly decreases towards 14 days. Altogether our analysis suggests wastewater viral load to be an early warning indicator for later hospitalization.

Additionally, we calculated Spearman correlation coefficients between the ICD-10 coded diagnoses of respiratory syndromes and wastewater viral load (see Table 2). Since the diagnoses are reported only on a weekly basis, we used the weekly viral load data for this analysis as well. As anticipated, we found a strong correlation (0.89) between wastewater viral load and COVID-19 diagnoses. Furthermore, diagnoses associated with non-specific COVID-19 symptoms also yielded high correlation coefficients. This is further supported by the correlation coefficients between other diagnoses and COVID-19, which are also around 90%. Given that the diagnoses are available only on a weekly resolution and thus have a reduced sample size, we were neither able to analyze the maximal time lag relative to wastewater viral load nor could we utilize these data for short-term forecasting of hospitalization rates.

Table 1

**Spearman Correlation between hospitalization rate (by 6 days shifted hospitalization rate) and wastewater viral load, and 95% confidence intervals.**
	Viral Load (Weekly)	Viral Load (Daily)
Hospitalization	0.80 ; (0.70, 0.86)	0.81 ; (0.78, 0.84)
Shifted Hospitalization	0.84 ; (0.77, 0.90)	0.84 ; (0.82, 0.87)

Table 2

**Spearman Correlation between ICD-10 coded diagnoses of respiratory syndromes and wastewater viral load, and 95% confidence intervals.**
	Viral Load (Weekly)	U0.7.1 COVID-19
U0.7.1 COVID-19	0.89 ; (0.82, 0.94)
J00-J06 Acute upper respiratory infections	0.90 ; (0.84, 0.94)	0.89 ; (0.81, 0.94)
J10-J18 Influenza and pneumonia	0.80 ; (0.67, 0.88)	0.94 ; (0.89, 0.96)
J20-J22 Other acute respiratory infections	0.85 ; (0.76, 0.91)	0.91 ; (0.86, 0.95)

3.2 Limited Benefit of Wastewater Viral Load for Forecasting of Hospitalization

The model performances on forecasting the hospitalization 7 days ahead are displayed in Table 3. The results are given as the mean MAE and mean MAPE over the 72 testing windows including the corresponding 95% confidence interval. All models were tested both autoregressive and with the inclusion of the wastewater viral load. Altogether the ARIMA model turned out to perform significantly better than all tested alternatives. Inclusion of wastewater viral load as an exogenous predictor did not improve prediction performance. To assess the extent to which wastewater data contributes to the prediction compared to hospitalization incidences, we calculated the proportion of the model's reliance on each data source on the example of ARIMA. This was done by summing the absolute values of the coefficients from the AR and MA components and comparing them to the absolute coefficient of the exogenous wastewater time series across all 72 test windows. Before averaging the AR/MA coefficients and the wastewater coefficients, we converted all 72 ratios into percentages to mitigate any distortions caused by unusually high or low coefficients. The results showed that, on average, the AR and MA components accounted for 78.35% of the model's total coefficient weight, while the wastewater data contributed 21.65%. For Random Forest and XGBoost we further tested whether the inclusion of the federal state wastewater viral load and hospitalization rates could improve the model performances. Here we only found a mean MAPE of 11.19% for Random Forest and of 11.92% for XGBoost, thus not improving the models that were solely trained on country level data.

Finally, we also built cross-modal Random Forest and XGBoost models, where the hospitalization was forecasted only based on the wastewater viral load as a predictor. Here we found a Random Forest performance of 7.13% and XGBoost performance of 8.87% mean MAPE. From this we can derive that both models can perform as good as the models that were either trained purely autoregressive or also including hospitalization data.

Table 3

**Model performances on forecasting hospitalization rates 7 days ahead.** Models were evaluated on 72 testing windows using the MAE and MAPE. Here the mean MAE and mean MAPE are displayed with the corresponding 95% confidence interval. Models were trained and tested autoregressive (AR) yes or no. The best performing model is highlighted in bold letters.
Model	AR	MAE	MAPE in %
Linear Regression	X	0.57 ; (0.40, 0.74)	8.51 ; (7.05, 9.97)
Linear Regression		0.71 ; (0.49, 0.92)	11.85 ; (8.74, 14.96)
Ridge Regression	X	0.54 ; (0.39, 0.68)	8.10 ; (6.77, 9.43)
Ridge Regression		0.54 ; (0.39, 0.68)	8.10 ; (6.77, 9.43)
ARIMA	X	0.32 ; (0.23, 0.40)	4.69 ; (3.88, 5.51)
ARIMA		0.30 ; (0.22, 0.39)	4.82 ; (3.97, 5.68)
Random Forest	X	0.61 ; (0.38, 0.85)	8.07 ; (6.63, 9.52)
Random Forest		0.49 ; (0.35, 0.63)	7.06 ; (5.86, 8.26)
XGBoost	X	0.96 ; (0.56, 1.37)	10.32 ; (7.58, 13.07)
XGBoost		0.60 ; (0.39, 0.80)	9.05 ; (6.53, 11.58)

Table 4

**Decision tree based cross-modal model performances on forecasting hospitalization rates 7 days ahead.** Models were evaluated on 72 testing windows using the MAE and MAPE. Here the mean MAE and mean MAPE are displayed in percent with the corresponding 95% confidence interval. Models were trained cross-modal, i.e. wastewater viral load was used as sole predictor for hospitalization rates.
Model	MAE	MAPE in %
Random Forest - Cross-Modal	0.50 ; (0.36, 0.64)	7.13 ; (5.91, 8.35)
XGBoost - Cross-Modal	0.67 ; (0.43, 0.92)	8.87 ; (6.48, 11.26)

Table 5

**Decision tree based model performances on forecasting hospitalization rates 7 days ahead including regional wastewater viral load and hospitalization data.** Models were evaluated on 72 testing windows using the MAE and MAPE. Here the mean MAE and mean MAPE are displayed with the corresponding 95% confidence interval. Models were trained including regional wastewater viral load data as well as regional hospitalization rates from 5 German federal states.
Model	MAE	MAPE in %
Random Forest - Regional	0.75 ; (0.53, 0.96)	11.19 ; (9.56, 12.83)
XGBoost - Regional	0.83 ; (0.58, 1.08)	11.92 ; (9.98, 13.86)

3.3 Model Performance is robust over time

In Fig. 5 we plotted the model performances (MAE and MAPE) on the test windows as a function of time and fitted a robust trend regression line to them on the example of ARIMA including wastewater viral load. The result shows that no significant increase in model prediction errors for the hospitalization rate could be observed. This demonstrates that our approach is robust against changes of the epidemic process in both the recording of hospitalization rate and wastewater viral load, as well as exogenous influence factors during this period.

The COVID-19 pandemic underlined the need for reliable measures and tools to effectively monitor and forecast the spread of the disease. Wastewater viral load has emerged as a promising indicator for the early detection of pandemic outbreaks and trends as well as a potential predictor for pandemic-related surveillance metrics, such as the hospitalization rate. In this study, we evaluated the wastewater viral load data collected by the AMELAG project in Germany, examining its correlation with COVID-19 hospitalization and incorporating it as a predictor in otherwise autoregressive models to forecast hospitalization rates 7 days ahead.

The correlation analysis showed a high Spearman correlation (around 0.8). These measures were even improved by shifting the time series against each other, using a time lag of 6 to 7 days. The observed changes in the relationship between hospitalization and wastewater viral load over time, with each time series peaking at different stages (see Fig. 3), suggest that factors such as variations in wastewater testing procedures or hospital recording methods, and reporting in general may have influenced the data. Despite these fluctuations, the high Spearman correlation (above 0.8) indicates a strong rank-order, monotonous relationship between the two variables, suggesting that as one increases or decreases, the other tends to follow a similar trend. The lag analysis further confirmed that wastewater viral load precedes hospitalization rate by approximately 6 to 7 days, indicating a potential lead time advantage in using wastewater data for early warning of increasing hospitalization. Similarly, the correlation analysis on the ICD-10 coded diagnoses showed high correlations indicating a strong relationship between wastewater viral load and diagnoses related to respiratory syndromes given from general practitioners.

The observed time lag between wastewater viral load and hospitalization rates motivated us to evaluate, in how far forecasting models could benefit from including wastewater surveillance as an exogenous predictor. However, our analysis could not show any significant improvement. In general, ARIMA outperformed all other alternative methods for short-term forecasting, which is in alignment with findings in our previous study (17). At the same time, including wastewater viral load as an exogenous predictor did not significantly enhance forecasting performance. This is also in agreement with the observation that the wastewater viral load had a limited influence of around 20% to the ARIMA forecast. A possible explanation is that both data modalities, wastewater viral load and hospitalization rates, have a high level of redundancy relative to each other, as reflected by the strong correlation. This could also be the reason why our cross-modal models performed similarly well as the ones including the hospitalization rate. Overall, this finding supports the idea that wastewater surveillance could be used for deriving an early warning indicator as well as predictor of hospitalization rates.

Models using regional wastewater viral load data did not outperform those based on aggregated country-level data, despite the expectation that more granular data would provide additional insights. One possible explanation is that wastewater sampling is conducted at distinct treatment plants, which are regionally not necessarily distributed in the same way as the population. This may lead to regional discrepancies in the time lag between viral load peaks and according hospitalization rates. In contrast, country-level data tend to be more smoothed, reducing the impact of such differences between regions.

Finally, we showed that the model performance is robust over time. A robust prediction performance is generally hard to achieve for forecasting models, since the recording of hospitalization and the measurement of wastewater viral load could change over time, but also exogenous factors like the pathogen variant, vaccination rate or availability of efficacious drugs may affect the hospitalization rate but differently wastewater viral load. This change in relationship was confirmed in the normalized plot of both time series (Fig. 2), which further reinforces the decision to use a sliding window approach for model development. This method allowed us to focus on short time segments where global changes are less likely to impact the analysis. It further ensures that the model remains adaptive to the local dynamics of the data, providing a more reliable prediction of trends within each window, regardless of broader fluctuations over time, and with this the confirmed robustness over time.

In their study on wastewater surveillance (36), Shah et al. pointed out that wastewater monitoring can identify the presence of SARS-CoV-2 even before any clinical cases are reported, which aligns with our findings. They further underscore the potential of wastewater-based epidemiology to deliver prompt insights for public health decision-making. These observations were corroborated by German studies (5, 7, 8). Hill et al. (37) discovered that incorporating wastewater data into a multivariate model for predicting hospitalization across 56 counties in New York State enhanced the model's performance by 15%. Similarly, Li et al. (38) demonstrated that wastewater data could effectively forecast hospitalization in the USA using a Random Forest approach. On the other hand, Pilz et al. (11) showed that while wastewater data qualitatively predicted hospitalization waves, it was inadequate for quantitatively predicting prevalence on city-level. Their study relied on wastewater viral load data from a single German federal state. Although our study did not observe significant enhancements from including wastewater viral load, our findings are consistent with the literature and emphasize the potential of wastewater viral load as an early detection indicator and as a sole predictor for cross-modal forecasting.

Our study does have its limitations. In our predictive modeling, we used daily interpolated wastewater viral load as a predictor. This interpolation, generated through a LOESS fit, may have introduced a slight data leakage into the analysis. Utilizing the original weekly wastewater viral load data presents statistical challenges due to the considerably smaller sample size, reduced by a factor of seven. Additionally, our focus was on short-term, seven-day forecasts. While longer-term forecasting and scenario planning are crucial during pandemics, short-term forecasts can still offer valuable insights, such as in the allocation of hospital resources or use of protective measures to reduce viral spread.

The COVID-19 pandemic highlighted the need for reliable monitoring tools, with wastewater viral load emerging as a promising early indicator of outbreaks. This study evaluated wastewater data from the AMELAG project in Germany, assessing its correlation with COVID-19 hospitalization and its potential as a predictor in forecasting models. Although strong correlations were found, particularly with a time lag of 6 to 7 days, incorporating wastewater viral load into predictive models did not significantly improve forecasting accuracy. On the other hand, using wastewater viral load as a sole predictor of hospitalization rates demonstrated a prediction performance comparable to using hospitalization rates alone. This suggests that wastewater data is valuable as an early indicator as well as predictor of epidemiological surveillance, especially as it provides information on community level and is much more cost effective than individually testing large populations. At the same time, a combination of wastewater information with hospitalization rates provides limited benefit.

Models using regional wastewater data did not outperform those based on national data, likely due to discrepancies between the regional distribution of wastewater plants and population density. Altogether, despite limitations, including data interpolation and a focus on short-term forecasts, our study reinforces the potential of wastewater surveillance as an early warning tool and predictor of epidemiological parameters. Future research should explore its broader application for other pathogens, potentially in combination with other data modalities such as social media (39).

MAE

Mean Absolute Error

MAPE

Mean Absolute Percentage Error

XGBoost

eXtreme Gradient Boosting

Ethics approval and consent to participate:

Not applicable

Consent for publication:

Not applicable

Availability of data and materials:

The hospitalization rates and the wastewater viral load data on GitHub: https://github.com/robert-koch-institut. For access to the diagnosis dataset contact docmetric. The code is available at: https://github.com/SCAI-BIO/wasterwater_analysis.

Competing interests:

The authors JB, ST, AK and HF declare that they have no competing interests. The authors AA and PR are employees of the commercial company docmetric, which did not influence the results of this study.

Funding:

This work has been supported by the AIOLOS (Artificial Intelligence Tools for Outbreak Detection and Response) project. The project was supported by the French State and the German Federal Ministry for Economic Affairs and Climate Action (grant number 01MJ22005A) and the French Ministry of Economy and Finance in the context of the France 2030 initiative and the Franco-German call on Artificial Intelligence technologies for risk prevention, crisis management, and resilience.

Authors’ contributions:

Conceptualization, methodology, supervision, project administration, and funding acquisition: AK, HF. Data curation, formal analysis, visualization, investigation, validation, and writing—original draft: JB, ST, AA, PR and HF Writing—review and editing: JB, ST, AA, PR, AK and HF. All authors contributed to the article and approved the submitted version.

Acknowledgements:

Not applicable

Fauci AS, Folkers GK. Pandemic Preparedness and Response: Lessons From COVID-19. J Infect Dis. 2023;228(4):422–5.
Botz J, Wang D, Lambert N, Wagner N, Génin M, Thommes E, et al. Modeling approaches for early warning and monitoring of pandemic situations as well as decision support. Front Public Health. 2022;10:994949.
National Academies of Sciences, Engineering, and, Health M, and Medicine Division; Division on Earth and Life Studies; Board on Population Health and Public Health Practice; Water Science and Technology Board; Committee on Community Wastewater-based Infectious Disease Surveillance. Wastewater-based Disease Surveillance for Public Health Action [Internet]. Washington (DC): National Academies Press (US); 2023 [cited 2024 Aug 21]. http://www.ncbi.nlm.nih.gov/books/NBK591712/
CDC. Advanced Molecular Detection (AMD). 2024 [cited 2024 Aug 21]. Wastewater Surveillance: A New Frontier for Public Health. https://www.cdc.gov/advanced-molecular-detection/php/success-stories/wastewater-surveillance.html
Wilhelm A, Schoth J, Meinert-Berning C, Agrawal S, Bastian D, Orschler L, et al. Wastewater surveillance allows early detection of SARS-CoV-2 omicron in North Rhine-Westphalia, Germany. Sci Total Environ. 2022;846:157375.
Vo V, Tillett RL, Papp K, Shen S, Gu R, Gorzalski A, et al. Use of wastewater surveillance for early detection of Alpha and Epsilon SARS-CoV-2 variants of concern and estimation of overall COVID-19 infection burden. Sci Total Environ. 2022;835:155410.
Agrawal S, Orschler L, Lackner S. Long-term monitoring of SARS-CoV-2 RNA in wastewater of the Frankfurt metropolitan area in Southern Germany. Sci Rep. 2021;11(1):5372.
Ho J, Stange C, Suhrborg R, Wurzbacher C, Drewes JE, Tiehm A. SARS-CoV-2 wastewater surveillance in Germany: Long-term RT-digital droplet PCR monitoring, suitability of primer/probe combinations and biomarker stability. Water Res. 2022;210:117977.
Joseph-Duran B, Serra-Compte A, Sàrrias M, Gonzalez S, López D, Prats C, et al. Assessing wastewater-based epidemiology for the prediction of SARS-CoV-2 incidence in Catalonia. Sci Rep. 2022;12:15073.
McManus O, Christiansen LE, Nauta M, Krogsgaard LW, Bahrenscheer NS, von Kappelgaard L, et al. Predicting COVID-19 Incidence Using Wastewater Surveillance Data, Denmark, October 2021–June 2022. Emerg Infect Dis. 2023;29(8):1589–97.
Pilz M, Küfer KH, Mohring J, Münch J, Wlazło J, Leithäuser N. Statistical analysis of three data sources for Covid-19 monitoring in Rhineland-Palatinate. Ger Sci Rep. 2024;14(1):10245.
Robert Koch-Institut. COVID-19-Hospitalisierungen in Deutschland [Internet]. Zenodo. 2024 [cited 2024 Aug 21]. https://zenodo.org/doi/10.5281/zenodo.13352753
Robert Koch-Institut F. 32. Abwassersurveillance AMELAG [Internet]. Zenodo; 2024 [cited 2024 Aug 21]. https://zenodo.org/doi/10.5281/zenodo.13354024
Cleveland WS. Robust Locally Weighted Regression and Smoothing Scatterplots. J Am Stat Assoc. 1979;74(368):829–36.
ICD-10 Version. 2019 [Internet]. [cited 2024 Sep 12]. https://icd.who.int/browse10/2019/en
Spearman Rank Correlation (Spearman’s Rho). Definition and How to Calculate it - Statistics How To [Internet]. [cited 2024 Aug 21]. https://www.statisticshowto.com/probability-and-statistics/correlation-coefficient-formula/spearman-rank-correlation-definition-calculate/
Botz J, Valderrama D, Guski J, Fröhlich H. A dynamic ensemble model for short-term forecasting in pandemic situations. PLOS Glob Public Health. 2024;4(8):e0003058.
Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, et al. Scikit-learn: Machine Learning in Python. J Mach Learn Res. 2011;12(85):2825–30.
What Is Ridge Regression? | IBM [Internet]. 2023 [cited 2024 Aug 22]. https://www.ibm.com/topics/ridge-regression
Introduction to ARIMA models [Internet]. [cited 2024 Aug 22]. https://people.duke.edu/~rnau/411arim.htm#arima010
Time Series Differencing. A Complete Guide | InfluxData [Internet]. [cited 2024 Aug 22]. https://www.influxdata.com/blog/time-series-differencing-complete-guide-influxdb/
pmdarima. ARIMA estimators for Python — pmdarima 2.0.4 documentation [Internet]. [cited 2024 Aug 22]. http://alkaline-ml.com/pmdarima/
GeeksforGeeks [Internet]. 2024 [cited 2024 Aug 22]. What Is an ARIMAX Model? https://www.geeksforgeeks.org/what-is-an-arimax-model/
Biau G, Scornet E. A random forest guided tour. TEST. 2016;25(2):197–227.
Chen T, Guestrin C, XGBoost:. A Scalable Tree Boosting System. In: Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining [Internet]. New York, NY, USA: Association for Computing Machinery; 2016 [cited 2024 Aug 22]. pp. 785–94. (KDD ’16). https://dl.acm.org/doi/10.1145/2939672.2939785
Luo J, Zhang Z, Fu Y, Rao F. Time series prediction of COVID-19 transmission in America using LSTM and XGBoost algorithms. Results Phys. 2021;27:104462.
Fang Zgang, Yang S, qin, Lv C, xia, An S yi, Wu W. Original research: Application of a data-driven XGBoost model for the prediction of COVID-19 in the USA: a time-series study. BMJ Open [Internet]. 2022 [cited 2024 Mar 4];12(7). https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9251895/
Galasso J, Cao DM, Hochberg R. A random forest model for forecasting regional COVID-19 cases utilizing reproduction number estimates and demographic data. Chaos Solitons Fractals. 2022;156:111779.
Özen F. Random forest regression for prediction of Covid-19 daily cases and deaths in Turkey. Heliyon. 2024;10(4):e25746.
Lv CX, An SY, Qiao BJ, Wu W. Time series analysis of hemorrhagic fever with renal syndrome in mainland China by using an XGBoost forecasting model. BMC Infect Dis. 2021;21(1):839.
Masini RP, Medeiros MC, Mendes EF. Machine learning advances for time series forecasting. J Econ Surv. 2023;37(1):76–111.
Shrivastava S. Cross Validation in Time Series [Internet]. Medium. 2020 [cited 2024 Feb 15]. https://medium.com/@soumyachess1496/cross-validation-in-time-series-566ae4981ce4
Akiba T, Sano S, Yanase T, Ohta T, Koyama M, Optuna. A Next-generation Hyperparameter Optimization Framework. In: Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining [Internet]. New York, NY, USA: Association for Computing Machinery; 2019 [cited 2024 Aug 23]. pp. 2623–31. (KDD ’19). https://doi.org/10.1145/3292500.3330701
Arize AI, [Internet]. [cited 2024 Sep 12]. Mean Absolute Error In Machine Learning: What You Need To Know. https://arize.com/blog-course/mean-absolute-error-in-machine-learning-what-you-need-to-know/
Mean Absolute Percentage Error (MAPE). What You Need To Know - Arize AI [Internet]. [cited 2024 Aug 22]. https://arize.com/blog-course/mean-absolute-percentage-error-mape-what-you-need-to-know/
Shah S, Gwee SXW, Ng JQX, Lau N, Koh J, Pang J. Wastewater surveillance to infer COVID-19 transmission: A systematic review. Sci Total Environ. 2022;804:150060.
Hill DT, Alazawi MA, Moran EJ, Bennett LJ, Bradley I, Collins MB, et al. Wastewater surveillance provides 10-days forecasting of COVID-19 hospitalizations superior to cases and test positivity: A prediction study. Infect Dis Model. 2023;8(4):1138–50.
Li X, Liu H, Gao L, Sherchan SP, Zhou T, Khan SJ, et al. Wastewater-based epidemiology predicts COVID-19-induced weekly new hospital admissions in over 150 USA counties. Nat Commun. 2023;14(1):4548.
Wang D, Lentzen M, Botz J, Valderrama D, Deplante L, Perrio J, et al. Development of an early alert model for pandemic situations in Germany. Sci Rep. 2023;13(1):20780.

No competing interests reported.

supplements.docx

Download PDF

Reviewers agreed at journal
02 Oct, 2024
Reviewers invited by journal
27 Sep, 2024
Editor assigned by journal
26 Sep, 2024
Submission checks completed at journal
24 Sep, 2024
First submitted to journal
21 Sep, 2024

You are reading this latest preprint version

Wastewater as an Early Indicator for Short-Term Forecasting COVID-19 Hospitalization in Germany

Status:

Version 1

Abstract

Background

Methods

Results

Conclusion

Figures

1. Background

2. Methods

2.1 Surveillance Data

2.2 Correlation Analysis

2.3 Forecasting Models

2.4 Sliding Window Approach for Model Training and Testing

2.5 Model Evaluation and Comparison

3. Results

3.1 Wastewater Viral Load as an Early Warning Indicator

3.2 Limited Benefit of Wastewater Viral Load for Forecasting of Hospitalization

3.3 Model Performance is robust over time

4. Discussion

5. Conclusion

Abbreviations

Declarations

References

Additional Declarations

Supplementary Files

Status:

Version 1