Simulating the spread of COVID-19 via spatially-resolved susceptible-exposed-infected-recovered-deceased (SEIRD) model with heterogeneous diffusion

• Reliable models that geo-localize outbreaks are needed to support forecasts and more effective allocation of medical resources.Source: §1 · p. 1
• The proposed PDE-based SEIRD model uses inhomogeneous diffusion to represent continuous spatio-temporal spread and deterministic movement shaped by local natural or social features.The authors note that local population densities vary in space and time, and defer modeling non-local dynamics with fractional differential operators to future work.Source: §1 · p. 2
• In Lombardy’s February–April 2020 outbreak, simulations showed remarkable qualitative agreement with reported epidemiological data, while reopening scenarios yielded contrasting results underscoring local population densities and contagion dynamics.Lombardy’s dense population and transportation network made it suitable for this mesoscales-focused approach.Source: §1 · p. 2

Model

• The model tracks susceptible, exposed, infected, recovered, and deceased populations with coupled PDEs over a spatial domain and time interval.The living population is defined as the sum of the susceptible, exposed, infected, and recovered populations.Source: §2 · p. 2
• An Allee effect represents outbreak clustering toward large population centers, and model parameters may vary with time, space, or population compartment.The Allee effect is characterized by parameter A.Source: §2 · p. 2
• Exposed asymptomatic patients can transmit at contact rate βe, some move directly to recovered at rate φe, and recovered patients are assumed immune over the modeled months.The model includes no backflow from recovered to susceptible.Source: §2 · pp. 2–3
• Spatial movement follows an inhomogeneous random walk whose diffusion limit is a second-order differential operator; population-proportional diffusivity can reflect local geographic or human inhomogeneities.Source: §2 · p. 3

3. Numerical implementation

• The Lombardy domain is discretized into 30,407 triangles; time stepping uses backward Euler with fully implicit Picard iteration, and linear systems use GMRES with a Jacobi preconditioner.A mesh convergence analysis was performed, but its data are not shown.Source: §3 · p. 3
• Initial conditions use population- and area-weighted Gaussian functions at municipalities with at least 10,000 inhabitants, initialized from 27 February 2020 Lab24 data.The initial data capture a severe Lodi outbreak and moderate exposed and infected counts in Bergamo, Brescia, and Cremona; homogeneous Neumann boundaries mimic complete regional isolation, acknowledged as unrealistic if lockdowns relax.Source: §3 · p. 3
• The model assumes σ=1/7, φr=1/24, φm=1/160, and φe=1/6 day−1 from literature data, and sets births and non-COVID-19 mortality to zero.The zero birth and mortality assumptions reflect the simulations’ months-long time scale.Source: §3 · p. 3
• Parameters are calibrated first by fitting a 0D SEIRD model to outbreak timing, then by recursive spatial simulations against spatiotemporal data, assessed with R2 and RMSE.The spatial refinement uses Eqs. (1)–(5).Source: §3 · pp. 3–4
• Deaths are the quantitative calibration target because case data may be unreliable amid incomplete testing and asymptomatic infections; other compartments are matched qualitatively, with forecasts focused on exposed and infected cases.The latter are highlighted as relevant to resource allocation and contagion-prevention decisions.Source: §3 · p. 4
• Symptomatic and asymptomatic infectivity are assumed equal, with βi,e decreasing across four lockdown phases; symptomatic diffusion remains fixed at νi=1.0 · 10−4 km2·day−1.βi,e is 3.3 · 10−4 on 27 February, 8.5 · 10−5 after 9 March, 6.275 · 10−5 after 22 March, and 4.125 · 10−5 after 28 March 2020; νs,e,r decreases from 0.0435 to 0.0198, 0.0090, and 0.0075 km2·day−1 over those phases. The Allee term A is 1,000 individuals·km−2, and low symptomatic diffusion reflects the assumption that symptomatic people are largely immobile.Source: §3 · p. 4

4.1. Forecasting the spatio-temporal dynamics of the COVID-19 pandemic in Lombardy

• The outbreak began in Lodi, Cremona, Bergamo, and Brescia, then moved north from Lodi through Milan’s southern suburbs to its center; Milan’s onset was later.Lodi and Cremona were initially most affected but quickly improved, unlike the explosive growth in Milan, Bergamo, and Brescia; Cremona was somewhat underpredicted, possibly because nearby Piacenza was excluded.Source: §4.1 · p. 4, Figure 2 · p. 5
• Calibration to deceased cases yielded more predicted infections than reported; the subgroup fit had R2=0.972 and range-normalized RMSE=7.6%.The authors suggest infection underreporting may explain the difference, while noting that mortality data depend on uncertain transmission rates; scaling infections to reported magnitudes was for visualization only.Source: §4.1 · p. 4, Figure 2 · p. 5

4.2. Exploring alternative reopening scenarios

• Four illustrative four-month scenarios after 27 February 2020 compared maintained restrictions, region-wide relaxation on 3 May, two relaxation assumptions, and keeping Milan restricted while relaxing elsewhere.Scenario A assumed early-outbreak contact rates; scenario B assumed greater success limiting contacts despite increased mobility, attributed to greater public awareness of prevention.Source: §4.2 · pp. 4–5

4.1. Forecasting the spatio-temporal dynamics of the COVID-19 pandemic in Lombardy

• R2 was 0.997 for all Lombardy, 0.977 for Bergamo, 0.976 for Brescia, and 0.998 for Milan, with qualitative agreement in the three main affected areas.Source: §4.1 · p. 4, Figure 2 · p. 5

4.2. Exploring alternative reopening scenarios

• Region-wide relaxation risked severe, rapid growth in Milan, whereas keeping Milan restricted while relaxing elsewhere produced more favorable outbreak dynamics.Brescia and Bergamo showed only marginal growth increases under region-wide relaxation and retained favorable trends.Source: §4.2 · p. 5

Discussion

• The model qualitatively reproduced Lombardy’s spatiotemporal outbreak, while reopening simulations favored locally tailored restrictions over a one-size-fits-all policy.Simulations suggested that maintaining strict lockdown outside Milan offered little benefit, while keeping restrictions in Milan might prevent explosive growth.Source: §5 · p. 6, Figure 3 · p. 6
• The proof-of-concept model needs richer calibration, boundary conditions, and spatial processes, especially for applications across larger regions.Proposed extensions include data assimilation and fitting additional compartments; links to neighboring regions; geography-dependent diffusion; and non-local effects.Source: §5 · p. 6
• Adding hospitalization, intensive-care, age, or sex compartments could improve resource-allocation support, while future analyses could incorporate lockdown costs.The authors also note that current results underscore the need for restrictions; socioeconomic costs are not included in the present analysis.Source: §5 · p. 6