Nowcasting influenza from wastewater with machine learning
A machine-learning study, in collaboration with the Luxembourg Institute of Science and Technology, that estimates influenza cases from viral concentrations in wastewater — turning environmental signals into an early epidemic indicator.
The problem
Traditional influenza surveillance lags behind real transmission. The COVID-19 pandemic showed that wastewater carries an early epidemiological signal — the question was whether that signal could reliably estimate influenza cases in Luxembourg.
Approach
I reviewed existing statistical methods, weighed their trade-offs, and benchmarked a wide field of predictive models against real data — training on 80% and validating on 20% — to find which best recovered case counts from viral load.
- Reviewed and compared candidate statistical methods
- Benchmarked 12 models (LOESS, KNN, Random Forest, Prophet, GAM, SVM, and more)
- 80/20 train–validation split, evaluated by MAE
- LOESS performed best, closely followed by KNN
Data & method
Weekly samples were collected over four years (March 2020–March 2024) from four wastewater treatment plants covering southern Luxembourg, analyzed via droplet-digital PCR. Influenza A and B viral concentrations were combined and modelled against reported cases.
- dd-PCR viral concentration measurements
- Influenza A + B loads combined
- A Statistical Analysis Plan guiding the pipeline
Outcome
The work fed into the VIRALERT research programme and associated scientific publications, establishing that machine-learning models can recover epidemic trends from wastewater data.
Public-health value
Wastewater-based nowcasting offers a low-cost, population-level early-warning signal for influenza — complementing clinical surveillance and giving health authorities more lead time to respond.
Lessons learned
Rigorous benchmarking beats intuition: the simplest models (LOESS, KNN) outperformed more complex ones on this signal.
A clear Statistical Analysis Plan up front makes results defensible and reproducible.
Environmental data can be a powerful, underused source of public-health intelligence.