The authors described epymorph as a Python package that builds spatial epidemic models by composing three separately specified components: transmission dynamics, movement processes, and geographic structure. This modular design forces explicit separation of model assumptions that are often intertwined in custom metapopulation code. The package retrieves the specific data each component requires (authors provided frozen copies in the manuscript repository), so investigators can swap movement models or contact structures without rewriting the entire model. epymorph therefore enables targeted comparisons where a change to a single component can be evaluated in isolation rather than bundled inside bespoke code.
epymorph includes built-in Bayesian filters to estimate time- and space-varying parameters from surveillance inputs. Those parameter estimates feed directly into probabilistic forecasts. The study team emphasized that the package moves from data assimilation to forecast generation in a single pipeline, reducing the friction between model fitting and operational prediction. The abstract indicates that parameter estimation and forecasting are implemented natively; the source did not report specific filtering algorithms, tuning choices, or prior specifications in the abstract itself, and those methodological details are available in the manuscript repository referenced by the authors.
Investigators applied epymorph across the 2025–2026 U.S. influenza season to test how demographic structure affects admissions forecasts. The cohort showed improved forecasts of hospital admissions when the population was stratified by age — but only when an empirically derived contact matrix was used. The authors reported that using an empirically derived contact matrix produced the forecast gains, whereas relying on synthetic mixing-pattern inputs did not produce the same improvement. The team used publicly available CDC Weekly Hospital Respiratory Data for influenza admissions and census-derived population by age; the abstract lists the data sources and notes that frozen inputs are archived with the code.
For the 2020–2021 winter wave of SARS-CoV-2, the investigators used epymorph to test the impact of commuting data and geographic aggregation on hospital census forecasts. Data indicated that incorporating county-to-county commuting flows improved county-level forecasts of hospital census. At the same time, the study team observed that a model aggregated at the state level provided the best state-level forecasts. Those findings show that adding realistic movement structure can matter at finer spatial scales, while coarser aggregation may out-perform detailed county models for statewide forecasts. The abstract does not report numeric forecast skill metrics or the precise form of the commuting model; those specifics are available in the manuscript and accompanying repositories.
The authors archived frozen copies of all inputs and provided links to the analysis repository and the epymorph codebase. Data sources listed in the manuscript included CDC and HHS hospital metrics, county-level COVID-19 death counts, U.S. Census American Community Survey population and commuting flows, TIGER/Line geographic files, and a contact-matrix repository. The team noted that epymorph can automatically pull the data components require and that all case-study inputs used for the two experiments are included in the analysis repository. The authors also declared no competing interests and confirmed that all data were aggregate, publicly available, and de-identified.
The abstract frames epymorph as a tool for controlled computational hypothesis testing but does not provide peer review validation at the time of posting; the article is a medRxiv preprint and has not been certified by peer review. The abstract did not report the numerical forecast performance measures, filter hyperparameters, or the exact model parametrizations used in each experiment; readers must consult the manuscript and archived code for those details. Unanswered practical questions described by the authors include which specific filter choices or priors are optimal for different pathogens, how transferable empirical contact matrices are across regions, and whether additional mobility layers (beyond commuting flows) materially improve forecasts in other seasons or pathogens. These remain empirical questions to be addressed with the reproducible experiments that epymorph facilitates.
The investigators made the code and frozen inputs available at the linked GitHub repositories, and they listed the public surveillance and census sources used for the influenza and COVID-19 case studies. That transparency allows other groups to reproduce the controlled model comparisons described here and to extend them to new regions, alternative movement datasets, or other infectious diseases. The manuscript includes further methodological detail and the archived inputs required to replicate the experiments reported in the abstract.