Modeling COVID-19 disease processes by remote elicitation of causal Bayesian networks from medical experts

Background COVID-19 is a new multi-organ disease, caused by the SARS-CoV-2 virus, resulting in considerable worldwide morbidity and mortality. While many recognized pathophysiological mechanisms are involved, their exact causal relationships remain opaque. A better understanding is needed for predicting their progression, targeting therapeutic approaches, and improving patient outcomes. While many mathematical causal models describe COVID-19 epidemiology, none have been developed for its pathophysiology. The virus's rapid and extensive spread and therapeutic responses made this particularly difficult. Initially, no large patient datasets were publicly available, and their data remains limited. The medical literature was flooded with unfiltered, technical and sometimes conflicting pre-review reports. Clinicians in many countries had little time for academic consultations, and in-person meetings were unsafe. Methods and Findings In early 2020, we began a major project to develop causal models of the pathophysiological processes underlying the disease's clinical manifestations. We used Bayesian network (BN) models, because they provide both powerful tools for calculation and clear maps of probabilistic causal influence between semantically meaningful variables, as directed acyclic graphs (DAGs). Hence, they can incorporate expert opinion and numerical data, and produce explainable results. Dynamic causal BNs, which represent successive "time slices" of the system, can capture feedback loops and long-term disease progression. To obtain the likely causal structures, we used extensive elicitation of expert opinion in structured online sessions. Centered in Australia, with its exceptionally low COVID-19 burden, we managed to obtain many consultation hours. Groups of clinical and other subject matter specialists, all independent volunteers, were enlisted to filter, interpret and discuss the literature and develop a current consensus. We aimed to capture the experts' understanding, so we encouraged discussion and inclusion of theoretically salient latent (i.e., unobservable) variables, documented supporting literature while noting controversies, and allowed experts to propose mechanisms by extrapolation from other diseases. Intermediary experts with some combined expertise facilitated the exchange of knowledge to BN modelers and vice versa. Our method was iterative and incremental: we systematically refined and checked the group output with one-on-one follow-up meetings with the original and new experts to validate previous results. In total, 35 experts contributed 126 face-to-face hours, and could review our products. Conclusions Our method demonstrates and describes an improved procedure for developing BNs via expert elicitation, which can be implemented rapidly by other teams modeling emergent complex phenomena. The results presented are two key models, for the initial infection of the respiratory tract and the possible progression to complications, as causal DAGs and BNs with corresponding verbal descriptions, dictionaries and sources. These are the first published causal models of COVID-19 pathophysiology, with three anticipated applications: (i) making expert knowledge freely available in a readily understandable and updatable form; (ii) guiding design and analysis of observational and clinical studies, by identifying potential mediators, confounders, and modifiers of treatment effects; (iii) developing and validating parameterized automated tools for causal reasoning and decision support, in clinical and policy settings. We are currently developing such tools for the initial diagnosis, resource management, and prognosis of COVID-19, parameterized using the ISARIC and LEOSS databases