P-599. If You’re On-time, You’re Late: Early Detection of Common Respiratory Pathogens Using Wastewater Genomic Surveillance to Shape Hospital Preparedness

TL;DR

This study shows that wastewater surveillance can help predict hospital respiratory disease trends, especially for SARS-CoV-2, up to two weeks in advance.

Contribution

The study demonstrates wastewater genomic surveillance as a predictive tool for hospital preparedness using machine learning models.

Findings

XGBoost models predicted SARS-CoV-2 with the highest accuracy (MAE: 1.9%-2.8%) across time horizons.

Wastewater data showed weak-moderate correlation with clinical outcomes for SARS-CoV-2 and flu A, but not for RSV.

More data and model optimization are needed to improve predictions for RSV and flu A.

Abstract

Wastewater (WW) surveillance (WWS) offers a promising approach to early detection of pathogens. This study explores the relationships between common respiratory viruses detected in WW and clinical laboratory data, examining whether WW pathogen changes are a leading indicator of changes in respiratory disease in hospitals.

WW samples from 8 locations at Walter Reed National Military Medical Center were collected 3 times per week from October 2024 to March 2025. SARS-CoV-2, respiratory syncytial virus (RSV), and influenza (flu) A were quantified in the samples using digital polymerase chain reaction (dPCR). Daily percent positivity was calculated from de-identified WRNNMC laboratory results data, and cross-correlations between percent positivity and WW dPCR were assessed. Several modeling approaches were explored, including linear regression (lasso, ridge), tree-based methods (random forests, XGBoost), and neural networks. Further, multiple prediction horizons were examined (1, 3, 5, 7, and 14 days). Train and test sets (80/20%) were constructed from the data. Time-series cross-validation was implemented to evaluate feature selection, models, and model parameters. Mean average error (MAE) between true and predicted positivity rate was used to evaluate model performance.

Cross-correlation analyses between laboratory test results and WW dPCR showed very weak correlation for RSV (Figure 1). Weak-moderate correlations were observed for SARS-CoV-2 and Flu A, particularly with shorter time lags for Flu A (Figure 2-3). For forecasting, XGBoost was the best performing model for all pathogens. SARS-CoV-2 models exhibited the best performance across time horizons (MAE: 1.9%-2.8%). RSV and Flu-A models demonstrated less predictive capabilities across varying prediction horizons (MAE: 4.3-11.3% and 4.3%-9.5%, respectively). Further analyses with more time-series data, additional model features, and model optimization are required.

These preliminary findings support that SARS-CoV-2, RSV, and Flu A WW pathogen data may be predictive of clinical outcomes. As additional data are collected, relationships between WW pathogen levels and clinical outcomes will be further explored.

Marleen M. Welsh, Ph.D., Altria Group Inc: Stocks/Bonds (Public Company)|Johnson & Johnson: Stocks/Bonds (Public Company)|Merk & Co Inc: Stocks/Bonds (Public Company)|Pfizer Inc: Stocks/Bonds (Public Company)|Solventum Corp: Stocks/Bonds (Public Company) Valerie J. Morley, PhD, Ginkgo Bioworks: employee|Ginkgo Bioworks: Stocks/Bonds (Public Company) Dawn Gratalo, MS, Ginkgo Bioworks: Stocks/Bonds (Public Company) Casandra Philipson, PhD, PhD, Ginkgo Bioworks: Stocks/Bonds (Public Company)

Figures