Epidemics in a Synthetic Urban Population with Multiple Levels of Mixing

Technical Summary: Epidemics in a Synthetic Urban Population with Multiple Levels of Mixing

Problem Statement
The prediction and control of infectious diseases require models that move beyond simple compartmental schemes (e.g., SIR) which assume fully mixed populations. While recent individual-based models incorporate heterogeneous contact patterns, there remains a need to understand how specific socio-demographic and geographic features of a host population impact epidemic diffusion. Specifically, the interplay between stable social relations (households, acquaintances) and rare, fortuitous interactions (e.g., in transit or public spaces) within a realistic urban setting is not fully characterized. The paper addresses the question of how different levels of mixing, dependent on age and distance, influence the speed, pervasiveness, and predictability of outbreaks in a medium-sized urban environment.

Methodology
The authors construct a data-informed, agent-based Susceptible-Infected-Removed (SIR) model applied to a synthetic population of approximately 60,000 agents representing the municipality of Viterbo, Italy.

• Synthetic Population: Agents are geo-referenced and age-stratified into four groups (children, young people, adults, elderly) using WorldPop density estimates and provincial census data. The territory is partitioned into a grid of 500m square tiles.
• Network Construction: The urban social network (USN) is an unweighted undirected graph $G = (V, E)$ formed by the union of two sub-graphs:
  1. Household Graph ($G_H$): A set of cliques inferred from census data, representing daily contacts.
  2. Acquaintance Graph ($G_A$): Edges representing frequent contacts (friends, coworkers) are added probabilistically based on age-based social mixing matrices (from SOCRATES data), geographic distance ($d^{-1}$), and an intrinsic "social fitness" parameter.
• Epidemic Simulation: The model simulates disease propagation where infected individuals transmit the disease to susceptible contacts with probability $\beta$ and recover with rate $\mu$. The model is calibrated to Influenza-Like Illness (ILI) parameters ($R_0 \approx 1.3$, recovery time 3 days).
• Experimental Configurations: Six distinct mixing configurations are compared to isolate the effects of contact patterns:
  1. Homogeneous Mixing (HM): Erdős-Rényi random graph (baseline).
  2. Static Network (SN): Only stable household and acquaintance relations; no fortuitous contacts.
  3. Homogeneous Noise (HN): Stable relations plus random fortuitous contacts.
  4. Age-based Noise (AN): Fortuitous contacts depend on age-based social mixing.
  5. Distance-based Noise (DN): Fortuitous contacts depend on geographic distance.
  6. Age- and Distance-based Noise (ADN): Fortuitous contacts depend on both age and distance.

Key Results

• Impact of Age Structure: Introducing age-structured contact patterns (SN, AN, ADN) results in significantly faster and more pervasive outbreaks compared to homogeneous mixing. These configurations produce degree distributions with longer right tails, leading to earlier epidemic peaks and higher attack rates.
• Role of Distance: Assuming that interaction frequency decays with distance ($d^{-1}$) has a negligible effect on high-level diffusion patterns (speed and total attack rate) compared to age-structured mixing. However, distance does influence the frequency of inter-tile infections, affecting local incidence.
• Epidemic Threshold and Reproduction Number: The epidemic threshold ($\beta_c$) is lower (making the system more vulnerable) in configurations with heterogeneous, age-driven contacts. The basic reproduction number ($R_0$) is generally higher in data-driven configurations. The emergence of an outbreak is strongly associated with the index case being a socially active individual.
• Geographic Spreading and Density: The study identifies two distinct regimes based on tile population density:
  • Low Density (< ~100 inhabitants): Sub-logarithmic time to peak and constant attack rates.
  • High Density: Approximately logarithmically increasing time to peak and attack rates.
  • Infection reaches densely populated tiles within 10–15 days regardless of configuration, but the timing for low-density areas varies significantly based on the contact model.
• Predictability: The epidemic is geographically heterogeneous (entropy < 1), with some tiles being consistently more dangerous. However, the system's predictability is low at the onset of the epidemic and only becomes high near the epidemic peak, when the infection is already widespread. No dominant "backbone" or preferred epidemic pathways were identified; the wide range of alternative routes prevents the formation of a single predictable trajectory.

Significance and Claims
The paper claims to provide a framework for understanding epidemic diffusion in realistic urban populations without requiring high-resolution mobility or transportation data, relying instead on census and survey data to construct the contact network.

• Hierarchical Diffusion: Preliminary evidence suggests the existence of hierarchical spatial diffusion patterns in urban areas, characterized by two regimes for epidemic spread in low- and high-density regions. This echoes findings observed at the country/city scale.
• Modeling Trade-offs: The study demonstrates that while distance-dependent mixing is empirically grounded, its impact on macro-scale epidemic outcomes is minimal compared to the structural impact of age-stratified social mixing.
• Limitations and Future Work: The authors note the absence of dominant epidemic pathways in their urban model, contrasting with meta-population models (e.g., airline networks) where traffic flows create dominant connections. They suggest that future work should investigate whether this absence is inherent to the urban scale or if incorporating additional data (mobility, phone calls, transportation) would reveal such structures.

The work underscores that accurate epidemic modeling requires moving beyond mean-field assumptions to account for the specific topological structure of social networks, particularly the role of age-structured contacts in driving outbreak dynamics.