Contributors

This work is led by the Institut National de Santé Publique (INSP) (Pierre Akilimali, Adelard Lofungola) and the Institut National de Recherche Biomédicale (INRB) Kinshasa/One Health Institute for Africa (INOHA) Kinshasa (Dav Ebengo, Placide Mbala-Kingebeni and Tania Bishola) in collaboration with the AfricaCDC and partners across the University of Oxford and Northeastern University; please contact dav.ebengo@umie-inrb.org or pierre.akilimali@insp.cd for further information.

Data

The entire dashboard was built using open-access, publicly available data.

Data sources

Data: https://github.com/INRB-UMIE/Ebola_DRC_2026

Administrative boundaries are the DRC Ministry of Health zones de santé (519 zones) based on publicly available shape files from the Humanitarian Data Exchange, joined to a separate set of 300-dimensional Population Dynamics Foundation Model (PDFM) embeddings (520 zones) via DHIS2 codes and OpenStreatMaps reference IDs.

Per-region populations are sums of the GRID3 v4.4 population raster (spatial resolution of 3 arc-seconds, ~100-metre grid cells) within the boundary of each region per the HDX shapefile. Health-facility counts come from GRID3 COD - Health Facilities v8.0, joined by DHIS2 codes. Refugee/IDP site locations were pulled from OpenStreetMap via Overpass.

Pairwise driving travel times are obtained from the public OSRM service. Pairwise great-circle (haversine) distances are computed directly from the health zone centroids. PCR testing capacity data was made available by AfricaCDC. Outbreak counts (suspected/confirmed cases and deaths) are extracted from the latest INSP sit rep.

Methods

Risk assessment via stochastic metapopulation SEIR modeling

We implemented a discrete-time tau-leap Susceptible Exposed Infectious Removed (SEIR) metapopulation model that allows for overdispersed transmission. We partitioned DRC into 519 health zones. Each zone i is treated as its own well-mixed sub-population with a fixed total population N_i (from GRID3 v4.4) and its own per-zone S_i / E_i / I_i / R_i compartments. Zones interact through a mobility kernel M_ij. The state evolves in discrete daily time steps. At each step, every susceptible in zone i has a probability 1 − exp(−λ_i Δt) of becoming exposed, with the per-zone force of infection

λ_i = (1 − ε) · B_i / N_i + ε · Σ_j M_ij · B_j / N_j,

where B_i is the summed per-individual transmission rate over all infectious cohorts currently in I_i, and a mixing parameter ε ∈ [0, 1] sets the fraction of contacts drawn from outside the zone. Individuals exit the latent period (E→I) following a gamma-distribution with mean 10 days and shape parameter n_E = 2 (coefficient of variation ≈ 0.71). We implement the gamma distribution by splitting E into n_E sequential exponential sub-stages each with rate n_E·σ, where σ = 1/(10 days), such that the total time governing the E→I transition is Erlang(n_E, n_E·σ) ~ Gamma(shape=n_E, mean=10 d). The infectious period (I→R) is assumed to be exponential with a mean of 7 days.

Overdispersion is implemented at the individual level such that rate β_j is drawn from Gamma(k, R0/(k·D_inf)) when a transition happens from E→I. The result is that the per-individual distribution of secondary infections is roughly NegBinom(R0, k). Individuals entering I in the same time step are tracked as a cohort, and recoveries split the cohort's transmission rate exactly via the Beta thinning identity Gamma(a+b) → Beta(a, b)-fraction.

Between-zone coupling M_ij is a row-normalized convex combination of three independently constructed components. Each component is a normalized origin→destination matrix. The three components are a long-range gravity model, short-range gravity model, and empirical flow model. We use driving times to calibrate the gravity models and Flowminder data for empirical flows and combine them such that

M_gravity_ij = (1 − w_short) · M_long_ij + w_short · M_short_ij

if row i has empirical outflow:

M_ij = (1 − w_emp) · M_gravity_ij + w_emp · M_emp_ij

else:

M_ij = M_gravity_ij;

where M_long_ij = N_j^α · exp(−D_ij / scale_long), M_short_ij = N_j · exp(−D_ij / scale_short), α is the gravity model exponent, scale_long and scale_short are normalization factors, M_emp_ij are the empirical movements, w_emp a weighting factor, and D_ij is the pairwise distance, and N_j the destination population.

We consider a range of different initial conditions. From each set of initial conditions, an outbreak is simulated forward to the date of the latest situation report. We compute the per-zone reporting rate such that the observed case fatality rate is 30%. We run a grid search over R0 × k × ε × and emp_w. Each combination is run 1,000 times via stochastic simulation. Scoring is log-RMSE between the observed and reporting-rate-scaled simulated case counts. We pool across initial conditions and parameters using Approximate Bayesian Computation to identify an overall most-favored scenario and calculate credible intervals.

DRC Ebola Bundibugyo 2026