Acute respiratory infections (ARIs) spread rapidly and present persistent public health challenges. Traditional surveillance that depends on sentinel hospitals and laboratory testing can lack timeliness and complete population coverage, limiting early detection and rapid response. To address these gaps, the authors developed an integrated framework that combines wastewater-based epidemiology (WBE) with machine learning to establish a community-level, fever prevalence–based early warning model for ARIs. The stated objective was to evaluate whether wastewater measurements of common antipyretic and analgesic consumption could serve as a proxy for community fever prevalence and whether machine learning–based anomaly detection could identify epidemic signals earlier or more reliably than conventional surveillance.
The study used wastewater data on paracetamol and ibuprofen consumption as markers of population-level febrile illness. A total of 603 wastewater samples (n = 603) were collected from multiple cities in China, including Dalian, Beijing, and Guangzhou, among others. These measurements were used to construct a fever prevalence indicator based on downstream signal processing and aggregation of drug-consumption signals in wastewater. The abstract describes the wastewater-derived indicator and its intended role as a proxy for community fever prevalence; detailed sample collection protocols, normalization methods, and exact indicator construction steps were not reported in the abstract.
To detect anomalies in the wastewater-derived fever indicator, the authors evaluated a range of approaches spanning statistical, unsupervised, and supervised learning methods. In total, 11 models were tested for anomaly detection performance. Specific model types named in the abstract that demonstrated favorable performance include the interquartile range (a statistical thresholding method), isolation forest (an unsupervised anomaly detection algorithm), neural networks (supervised or unsupervised depending on implementation), and logistic regression (a supervised classification approach). The abstract does not provide the full list of all 11 models, nor does it provide hyperparameter settings, training-validation splits, or other modelling details.
Model validation and evaluation identified several approaches with favorable performance. The interquartile range and isolation forest methods were notable because the warning thresholds they produced aligned with external surveillance data: the interquartile range yielded a warning threshold of 3.3%, and the isolation forest produced a threshold of 3.5%. Using these thresholds, the models classified November 2021 as a non-epidemic period and November 2024 as an epidemic period. The abstract reports these threshold values and the corresponding classification of those months but does not include additional metrics such as sensitivity, specificity, lead time, or cross-validation results in the abstract.
The study compared the wastewater- and machine learning–derived early warning outputs with influenza-like illness surveillance from sentinel hospitals. The interquartile range and isolation forest thresholds produced epidemic/non-epidemic determinations that were consistent with sentinel surveillance: November 2021 was identified as non-epidemic and November 2024 as epidemic. This concordance is presented as supporting evidence that the WBE-plus–machine learning approach can reflect community-level fever dynamics and align with established surveillance indicators.
According to the authors, coupling WBE with machine learning offers a timely and precise method for monitoring community-level fever prevalence. The combined approach is proposed to provide early-warning capability for fever-dominant emerging respiratory infections and for previously unidentified etiological agents, referenced in the abstract as “Disease X.” The authors argue this framework addresses limitations of conventional surveillance—specifically, incomplete coverage and delayed response—and provides a scalable, evidence-based tool to strengthen public health preparedness and response at the community level.
The authors declared no competing interests in the Conflict of Interest statement. The abstract provides key high-level findings, sample count (n = 603), named cities sampled (including Dalian, Beijing, and Guangzhou), the list of some evaluated models, and the two specific warning thresholds (3.3% for interquartile range and 3.5% for isolation forest) that matched sentinel surveillance classifications for November 2021 and November 2024. Detailed methodological descriptions, full model lists and settings, sensitivity and specificity metrics, and other granular results were not reported in the abstract and would require consultation of the full text for comprehensive appraisal.
This study presents an integrated WBE and machine learning framework that uses wastewater-derived paracetamol and ibuprofen signals to estimate community fever prevalence and detect anomalies indicative of ARI epidemics. Of 11 tested models, several performed well, and thresholding by interquartile range and isolation forest produced epidemic designations aligned with sentinel hospital influenza-like illness surveillance. The authors conclude that the approach can deliver a scalable, timely early-warning tool to augment conventional surveillance and support public health preparedness.