Complex Agent Networks explaining the HIV epidemic among homosexual men in Amsterdam

Simulating the evolution of the Human Immunodeficiency Virus (HIV) epidemic requires a detailed description of the population network, especially for small populations in which individuals can be represented in detail and accuracy. In this paper, we introduce the concept of a Complex Agent Network(CAN) to model the HIV epidemics by combining agent-based modelling and complex networks, in which agents represent individuals that have sexual interactions. The applicability of CANs is demonstrated by constructing and executing a detailed HIV epidemic model for men who have sex with men (MSM) in Amsterdam, including a distinction between steady and casual relationships. We focus on MSM contacts because they play an important role in HIV epidemics and have been tracked in Amsterdam for a long time. Our experiments show good correspondence between the historical data of the Amsterdam cohort and the simulation results.

💡 Research Summary

This paper introduces a novel modeling framework called the Complex Agent Network (CAN) to simulate the spread of HIV with a high level of detail, especially in relatively small populations where individual behavior and network structure can be explicitly represented. The authors combine two well‑established approaches—multi‑agent systems (MAS) and complex network (CN) theory—into a hybrid model that captures both the heterogeneity of individual agents (age, infection stage, treatment status, risk behavior) and the non‑trivial topology of the contact network (scale‑free degree distribution, dynamic rewiring, distinction between steady and casual partnerships).

The case study focuses on men who have sex with men (MSM) in Amsterdam, a population that has been monitored for decades through the Amsterdam Cohort Study (ACS). The authors construct an undirected, scale‑free sexual contact network using the configuration model, assigning each vertex a degree drawn from a power‑law distribution p(k) ∝ k⁻ᵞ. Each agent can have at most one steady partner (lasting on average y years) and any number of casual partners; steady partnerships involve multiple sexual acts per year, while casual ones involve a single act.

HIV progression is modeled in three stages: primary infection (PI), asymptomatic period (AP), and AIDS. The PI stage, which lasts about three months, is split into two sub‑stages with distinct transmission probabilities (TP_PI,1 and TP_PI,2) to reflect the heightened infectivity during early infection. For each partnership, the per‑act transmission probability is calculated as

 P_TA_ij = F_P_j × F_R_ij × (F_T_i × TP_i),

where F_P_j is a partnership factor reflecting whether the partner has a steady relationship, F_R_ij captures risk‑behaviour adjustments, and F_T_i represents the reduction due to antiretroviral treatment. The number of acts per year (N_A_ij) and the type of partnership (steady vs. casual) further modulate the annual transmission probability. The overall probability that a susceptible agent i becomes infected in a given year is then

 P_TY_i = 1 – ∏{m=1}^{k_i} (1 – P_TY{mi}),

where the product runs over all of i’s partners.

The simulation proceeds in yearly time steps. At each step agents age, partnerships age, and the network evolves: (1) agents older than 65 are removed, (2) agents who reach the AIDS stage are removed, (3) a small fraction may migrate out, and (4) new susceptible agents are added to keep the population size constant. Partnerships are retained if they are steady and unexpired; casual ties are kept with a configurable probability (e.g., 0.2).

Implementation uses the Java‑based MASON agent framework for scheduling and the JUNG library for network handling, providing advanced scheduling, interactive visualization, and real‑time parameter tracking.

Results are compared with historical ACS data on yearly HIV prevalence, average number of partners, and age distribution. The CAN model reproduces the observed epidemic curves with high fidelity. Notably, the model captures the empirically observed reduction in transmission when the proportion of steady partnerships increases, reflecting the protective effect of reduced risk behaviour among individuals with long‑term partners. Sensitivity analyses show that treatment effectiveness, partnership duration, and risk‑behaviour factors substantially influence epidemic outcomes.

The paper’s contributions are threefold: (1) it proposes the CAN framework that integrates MAS and CN to enable fine‑grained epidemic modeling in small, heterogeneous populations; (2) it formalizes MSM‑specific partnership dynamics and stage‑dependent transmission probabilities, grounding them in empirical data; (3) it validates the model against a long‑standing cohort study, demonstrating both qualitative and quantitative agreement. The authors also outline extensions such as exploring other sexually transmitted infections, incorporating migration patterns, and testing public‑health interventions (e.g., partner‑number limits, treatment scale‑up).

In conclusion, the Complex Agent Network offers a powerful, flexible tool for simulating disease spread where individual heterogeneity and network topology are both critical. Its successful application to the Amsterdam MSM HIV epidemic suggests broad applicability to other contexts requiring detailed, data‑driven modeling of transmission dynamics.