Disease outbreaks and other public health emergencies involve a potentially wide range of different stakeholders. These stakeholders are typically interconnected by various types of interactions within and between themselves as well as with the pathogen and the respective social, natural and built environments. We propose to model the interactions among different stakeholders as a heterogeneous decision network where nodes denote different stakeholders and edges denote different types of interactions among them, while the interactions can be characterized based on the availability and sharing of data pertaining to said interactions, e.g. as to whether the pathogen is known, how it can be transmitted, or whether vaccination or treatment is available and how much it costs. We refer to the sharing of data or metadata as data flow, which can be modulated in several ways as one interaction triggers the next. For instance, a student might decide to change their behavior towards others based on the outcome of a diagnostic test, and authorities might decide to temporarily close the affected school or not.

For various reasons (e.g., societal, political, technical, or ethical), much of the data relevant for a complete path through such a decision network may not be directly accessible, posing challenges to investigating its potential and real decision chains. For example, outbreak propagation could be better characterized, understood, predicted and communicated if precise demographic information and migration information were available for individuals near the epicenter (McDonald et al. 1992,Pastor-Satorras and Vespignani 2001). The existence of such data, e.g. in government databases, does not generally imply its availability to other stakeholders, though sometimes, similar insights can be gleaned from other sources, e.g. mobility data from providers of transport services or fitness apps *4*5.

The modeling of the flow of data, metadata and related information through such a decision network enables us to create hypotheses to investigate the impact of the availability and quality of data on specific kinds of decisions, or with respect to specific stakeholders, locations or timing. In designing the model, we will initially focus on data from Public Health Emergencies of International Concern as well as the WHO’s R&D Blueprint*6, keeping in mind the applicability to other epidemiological contexts such as seasonal flu or the opioid crisis.

Once the structure of the decision network and the nature of potential data flows has been captured in the model, we can study causal relations between individual or aggregated decisions (such as the closure of one school or multiple schools), the associated data flows, and the spread of the pathogen and the disease. Such decision-making processes are often based on randomized controlled experiments (Guo et al. 2020, Kallus and Zhou 2019). However, they could be expensive in multiple dimensions, including in terms of time and financial resources, which makes them challenging to perform in outbreak contexts. One way to address this is to consider observational data of different stakeholders to infer causal effects between a specific decision (e.g., closure of school) and an important outcome (e.g. spread or containment of disease; Pearl 2009,Rubin 1978).

Learning such treatment effects from observational data as in the mobility example above requires us to handle confounding bias, which are unobserved variables that influence both the treatment and the outcome. For example, an individual’s poor socioeconomic status can affect their living conditions and may increase their chances to be infected, treated or cured. In addition to the observational data, we can include measures of the information flow between stakeholders (e.g. hygienic advice, or rumors) to infer the existence of hidden confounders (Hill 2011,Guo et al. 2019). To be able to adapt the decision network model to the nature of a given real or hypothetical outbreak scenario, data availability can be modeled generically as a systemic property, more granularly at the level of classes of stakeholders or decisions, and in a yet more fine-grained manner at the level of individual stakeholders or interactions if suitable data are available or can be inferred.

For a given outbreak management context, we can fine-tune the model parameters for that context in order to use the model to address questions that might arise during outbreak management. While classical epidemiological modeling provides information on expected outbreak dynamics and recommendations on outbreak management like how much of what to stockpile, our model would provide additional insights into whether, how and when details about preparedness and response or associated research should be shared and with whom, or what data quality requirements should be aimed for at which junctions in the network.

These details might range in scope from individual patients to triage protocols used by health workers to institutional or international policies about sharing diagnostic kits, material samples or computational pipelines. In short, the decision network model would behave much like a machine-actionable data management plan for the outbreak in question, which would also allow, for instance, to notify specific stakeholders of data-related outbreak developments relevant to them (Miksa et al. 2019).