13 June 2024

Cam Zachreson: A comparison of three ABMs

In this meeting Cam gave a presentation about the relative merits and trade-offs of three different approached for agent-based models (ABMs).

Attendance: 7 in person, 13 online.

Theoretical frameworks

Key message

Each framework is built upon different assumptions about space, contacts, and transmission.

Cam introduced three theoretical frameworks for disease transmission, which he used in constructing infectious disease models for three different projects. Note that all three models use the same within-host model for individual dynamics.

1. Border quarantine for COVID-19: international arrivals, quarantine workers, and the local community are divided into mixing groups within which there is close contact. There is also weaker contact between these mixing goups.

2. Social isolation in residential aged care facilites: the transmission is a multigraph that explicitly simulates contact between individuals. The graph is dynamic: workers and worker-room assignments can change every day. Workers share N edges when they service N rooms in common.

3. A single hospital ward (work in progress): a shared space model represents spatial structure as a network of separate spaces (i.e., nodes). Nurses and patients are associated with spaces according to schedules. Each space has its own viral concentration, driven by shedding from infectious people and ventilation (the air changes around 6 times per hour). Residence in a space results in a net viral dose, which confers a probability of infection (using the Wells-Riley model).

Question

Are many short interactions equivalent to one long interaction?

Pros and cons of model structures

Key message

Each framework has unique strengths and weaknesses.

As shown in the slide below, Cam identified a range of pros and cons for each modelling framework. Some of the key trade-offs between these frameworks are:

• The ease of validation (aged care and hospital ward) versus the ease of communication (quarantine);

• Having explicit physical parameters and units (hospital ward) versus having vague and/or phenomenological parameters (quarantine and aged care); and

• Being simple to construct and efficient to run (quarantine and aged care) versus being complex to construct and computationally expensive (hospital ward).

Pros and cons of the three approaches.

Constructing complex models

Key message

Complex models typically have complex data requirements.

Data requirements can also present a challenge when constructing complex models. For example, behaviour models are good for highly-structured environments such as hospital wards, where nurses have scheduled tasks that are performed in specific spaces. However, the data required to construct the behaviour model can be very hard to collect, access, and use. Even if nurses wear sensors, the sensor data are never sufficiently clean or complete to use without substantial cleaning and processing.

Airflow between spaces in a highly-structured environment is also complex to model. Air exchange can involve diffusion between adjacent spaces, but also airflow between non-adjacent spaces through ventilation systems. These flows can be difficult to identify, and are computationally expensive to simulate (computational fluid dynamics).

Cam concluded by observing that existing hospitals wards tend to have a design flaw for infection control:

There are many shared spaces in which infection can spread among individuals via air transmission.

Reproducibility in stochastic models

Key message

These models rely on random number generators (RNGs), whose outputs can be controlled by defining their initial seed. Using a separate RNG for each process in the model provides further advantages (see below).

In contrast to agent-based models of much larger populations, these models are small enough that they can be run as single-threaded code, and multiple simulations can be run in parallel. The bulk of computational cost is usually sweeping over many combinations of parameter values.

The aged care (multigraph) and hospital ward (shared space) models decouple the population RNG from the transmission dynamics RNG. An advantage of using multiple RNGs is that we can independently control and modify these processes. For example, by using separate RNGs for infections and testing, we can adjust testing practices without affecting the infection process.

Topic for a Masters project

Question

Does anyone know a suitable Masters student?

Cam is looking for a Masters student to undertake a project that will look at individual-level counterfactual scenarios. The key idea is to identify sets of preconditions (e.g., salient details of the event history and/or current epidemic context) and ensure that the model will always generate the same outcome when given these preconditions. An open question is how far back in the event history is necessary/sufficient.