Researchers Enpei Chen and Xiong Yu have developed an adaptive machine learning framework that forecasts municipal COVID-19 clinical case surges three days in advance using sewage viral data and past clinical records. Published in AI in Civil Engineering, the study evaluated longitudinal monitoring data from the Ohio Wastewater Monitoring Network between July 2020 and October 2023 across four urban sewersheds in Cleveland, Youngstown, Cincinnati, and Akron. By integrating time series of SARS-CoV-2 N2 gene concentrations with past infection numbers, the computational framework achieved over ninety percent accuracy in predicting impending disease surges. The authors established that sewage-based pathogen tracking provides dependable real-time pre-warnings of viral outbreaks across evolving pandemic stages, offering public health authorities an actionable operational window before clinical hospitalizations and diagnoses escalate.
Wastewater-based epidemiology has emerged as an effective surveillance tool capable of detecting community pathogen circulation before patients receive formal clinical testing. Because infected individuals shed viral particles through sputum, urine, and feces regardless of symptoms, analyzing sewer influent captures both symptomatic and asymptomatic transmission across sewersheds. However, converting raw viral concentrations into quantitative clinical predictions has historically challenged public health analysts. Shifting viral shedding rates, varying transmission dynamics, emerging biological variants, and changing public behaviors generate complex relationships between wastewater pathogen measurements and recorded clinical cases. Conventional mathematical formulas and classic regression methods often experience numerical instability when infection counts drop to low levels, while earlier machine learning studies typically operated across static observation windows of only a few months, limiting their adaptability when mass vaccination campaigns, mutations, or policy shifts alter viral dissemination.
To overcome these forecasting challenges, the authors analyzed multi-year data from the Ohio Wastewater Monitoring Network, an inter-agency surveillance initiative established in June 2020 that monitors coronavirus gene copies across more than fifty Ohio counties. The study examined four municipal wastewater treatment plants representing distinct metropolitan settings: the NEORSD Westerly Wastewater Treatment Center in Cleveland, the Youngstown Wastewater Treatment Plant, the Mill Creek Wastewater Treatment Plant in Cincinnati, and the Akron Water Reclamation Facility. Participating facilities adhered to standardized collection protocols, gathering twenty-four-hour composite influent samples twice weekly. To quantify community viral loads, analytical laboratories amplified the N2 gene target region of the SARS-CoV-2 genome using nucleic acid amplification techniques, averaging gene counts across four duplicate samples while maintaining strict quality assurance and control standards across all participating testing institutions.
From summer 2020 to spring 2022, measured SARS-CoV-2 N2 gene concentrations in sewage correlated significantly with newly reported clinical infections across the four sewersheds, anticipating clinical infection trajectories by an average of three days. However, this lead time varied considerably between locations and pandemic waves. In December 2021 within the Akron sewershed, N2 concentrations reached an average of 347,977 copies per liter one week before clinical infections peaked, whereas in Mill Creek during the Omicron surge, wastewater viral loads and reported infections peaked nearly synchronously. Furthermore, between spring 2022 and late 2023, the correlation between wastewater viral levels and clinical cases decoupled noticeably as clinical testing participation declined and asymptomatic infections expanded, elevating the ratio of measured wastewater viral concentration to diagnosed clinical cases across communities.
To maintain predictive accuracy despite these shifting dynamics, the researchers implemented Long Short-Term Memory neural networks, a specialized recurrent architecture designed to learn long-term temporal dependencies without suffering from backpropagation limitations. In the network structure, an input matrix representing temporal features at a given timestep interacts with the hidden state from the prior timestep to regulate an input gate, a forget gate, an output gate, and an internal memory cell. The forget gate applies a sigmoid activation function to discard non-informative past fluctuations, while the input gate and a hyperbolic tangent cell function govern the entry of fresh observations. An element-wise Hadamard product updates the long-term cell state, and the output gate determines the updated hidden state carrying short-term memory forward, with weights and biases optimized through gradient descent.
Feature engineering served as an essential initial stage prior to executing the machine learning model. Because clinical infections were recorded daily while wastewater was sampled twice weekly, linear interpolation expanded wastewater viral concentration data to a daily time series, followed by Min-Max scaling to normalize input values. The network incorporated both wastewater viral measurements and historical clinical case counts as dual inputs to maintain robustness against sampling and transport uncertainties. Operating with a fifteen-day moving observation window, the model was configured to predict future clinical infections three days in advance. Hyperparameter optimization was conducted through grid search to minimize the loss function, identifying optimal values for learning rate, batch size, epoch quantity, Adam optimizer decay rate, hidden dimension, and layer dimension within the deep learning architecture.
Recognizing that static neural networks degrade when public health dynamics change, the study introduced an iterative training regime that updated model parameters every six months to match the unfolding epidemic. After achieving optimized performance on the initial twelve-month dataset spanning August 2020 to August 2021, the first trained model was evaluated on testing data from August 2021 through January 2022. At the conclusion of this initial six-month testing phase, newly acquired wastewater measurements and clinical tallies were incorporated into the historical archive and partitioned into updated training and validation sets to calibrate a second model iteration. This six-month updating cycle was repeated through four sequential stages, culminating in Model 4 completing predictions for the final testing window from April 2023 to October 2023, enabling the system to internalize shifting transmission patterns and maintain forecasting stability without manual restructuring.
When evaluated against seven-day moving averages of clinical cases, which smoothed day-to-day reporting irregularities, the iterative architecture demonstrated high predictive accuracy across all four municipal sewersheds over the extended evaluation period. For the initial model evaluated on NEORSD Westerly data, the framework attained a coefficient of determination of 86.66 percent with a mean absolute error of seven cases on the training dataset, and a coefficient of determination of 94.83 percent with a mean absolute error of one case on the validation dataset. Across the four combined testing stages from August 2021 through October 2023, the model produced coefficients of determination of 87.05 percent in NEORSD Westerly, 91.04 percent in Youngstown, 93.22 percent in Mill Creek, and 91.29 percent in Akron, tracking epidemiological curves with notable fidelity.
The model achieved its lowest relative mean absolute error in Akron at 16.91 percent, reflecting a mean absolute error of 12 cases against a mean of 72 cases, compared to 19.15 percent in Mill Creek, 23.15 percent in Youngstown, and 27.43 percent in NEORSD Westerly. Under daily case evaluation without smoothing, the models achieved coefficients of determination of 72.51 percent in NEORSD Westerly, 71.02 percent in Youngstown, 70.20 percent in Mill Creek, and 70.12 percent in Akron. Although the models exhibited slight underestimations during extreme peaks, such as the December 2021 Omicron wave when Mill Creek cases exceeded 1,200 and shedding durations contracted from seven days to three days, the network consistently traced community disease trajectories across all sewersheds.
To convert numerical forecasts into actionable decision-making tools for public health officials, the researchers introduced an early warning framework evaluated using confusion matrices. An impending outbreak warning was triggered if predicted daily case numbers exceeded the seventieth percentile of historical observations from a baseline window between August 2020 and August 2021, with a secondary statistical threshold defined by the ratio of variance to the mean. Across Youngstown and Mill Creek, the system achieved overall early warning accuracy exceeding ninety percent for moving average forecasts and approximately ninety percent for daily projections. In NEORSD Westerly, recall exceeded ninety percent and precision surpassed eighty percent, with false warnings outnumbering missed warnings due to slight initial overestimations at high case volumes, ensuring that over ninety percent of days requiring outbreak alerts were successfully captured.
To evaluate model generalizability across geographic regions, the authors conducted transfer learning experiments by deploying the pre-trained Akron base model directly to NEORSD Westerly, Youngstown, and Mill Creek without adjusting or retraining any internal model weights. Leveraging temporal patterns learned exclusively from Akron surveillance data, the transferred model generated accurate multi-timestep forecasts across the other three sewersheds, maintaining coefficients of determination near ninety percent and relative mean absolute errors around twenty percent. This successful cross-site deployment indicates that an established neural network model trained in one watershed can provide early warnings for separate communities that lack extensive historical data, conserving computational resources and accelerating surveillance setup in unmonitored municipalities, although the authors noted that future validation should incorporate wastewater monitoring datasets collected outside Ohio to confirm broader nationwide generalizability.
The researchers also demonstrated that the model reliably forecasted clinical case fluctuations influenced by public health policy mandates and societal events. In the NEORSD Westerly sewershed, the model captured infection surges following university autumn semester resumptions and major holidays, including Labor Day, Columbus Day, Thanksgiving, and Christmas, while tracing case declines linked to overnight curfews in November 2020 and expanding vaccine distribution in early 2021. The network also tracked infection rebounds following Presidents’ Day and subsequent surges in mid-2021 associated with the rescission of municipal emergency proclamations and mask mandates. Looking forward, the authors emphasized that combining time series artificial intelligence with municipal sewer monitoring not only strengthens proactive disease surveillance but also offers valuable engineering utility for wastewater infrastructure management, including dynamic hydraulic modeling and contaminant forecasting to improve urban system resilience.
Subject of Research: Machine learning for real-time wastewater-based epidemic forecasting
Article Title: A transferable machine learning model for real-time forecast of epidemic dynamics and pre-trigger event warning
Article References: Chen, E., & Yu, X. (2025). A transferable machine learning model for real-time forecast of epidemic dynamics and pre-trigger event warning. AI in Civil Engineering, 4(1), Article 18. https://doi.org/10.1007/s43503-025-00059-5
Image Credits: AI Generated
DOI: 10.1007/s43503-025-00059-5
Keywords: wastewater-based epidemiology, machine learning, LSTM neural networks, COVID-19 surveillance, epidemic forecasting, early warning systems, public health surveillance, transfer learning, transferable, machine, learning, model
News Source: Blake Davidson. (November 10, 2026). Machine Learning Model Forecasts Epidemic Outbreaks from Wastewater Days in Advance. Scienmag.



