Participatory Agent-Based Modelling of Emergency Department Patient Flow
Emergency health care is in crisis; the core “4-hour” KPI has not been met since July 2015. Emergency Departments (EDs) are socio-technical organisational systems with complex internal interactions between a wide range of actors as well as complex interactions with their urban environment, such as the capacity for local city authorities to provide care in the home and avoid readmission. As a result, it is very difficult to understand the potential effects of any changes in practice on the overall effectiveness of the ED and the twin goals of meeting the 4-hour KPI while ensuring patient safety and quality of care. Over the past two years, and in collaboration with contacts at King’s College Hospital, we have been developing agent-based models (ABMs) of EDs and the interactions of staff and patients therein. ABMs are an AI technique particularly suited to this problem as they provide explainable analysis of the emergent behaviour of complex systems formed from the lower-level interaction of large numbers of agents (e.g., staff in the ED case).
However, there is a critical issue of trust by hospital staff in how ABMs are currently employed: the correspondence of the simulation with the details of a real hospital can be opaque to users and out of their direct control, so they can mistrust the results. ABMs currently are implemented directly in languages such as Java or C++, meaning they are too technical and too far removed from the domain terminology to be directly understood and manipulated by domain stakeholders (hospital managers, clinical staff). Any engagement of domain stakeholders with the models and their outputs must necessarily always be mediated by a technical stakeholder‚ typically, a software developer with expertise in agent-based modelling, but who is not normally an expert in the emergency-health domain. This may involve high-level modelling notations, but these are informal and have no direct technical connection to the final model and simulations. As a result, domain stakeholders find it difficult to trust that the final agent-based models and the simulation results adequately reflect clinical reality, and findings from ABM-based analyses are often not translated into interventions to actual practice.
The aim of this PhD project is to study how using domain-specific modelling languages (DSMLs) including a variety of notations (graphical, textual, tabular) closely aligned with domain stakeholders-conceptualisation of the ED environment will affect trust in agent-based models and acceptance of ABM technology in this field. We expect this to increase uptake and impact of this promising AI technology.
Relevant research questions include:
- What do these DSMLs need to look like to appeal to domain stakeholders?
- What concepts do we need to capture to generate meaningful AB simulations?
- How do we measure domain stakeholders-trust in ABM models and simulation results?
patient flow; agent-based modelling; domain-specific languages; participatory modelling