33. Graphical Models in Clinical Epidemiology: Literature Review and Case Study

150 150 TECHNA Symposium

Zheng Jing Jimmy Hu 1, Nicholas Mitsakakis 2

Dalla Lana School of Public Health, University of Toronto
Nicholas Mitsakakis, IHPME, University of Toronto; BRU and TGHRI, UHN

Graphical models (GM) are methods at the intersection of Statistics and AI that use the language of a graph in order to represent and examine dependences and probabilistic relationships between variables. In a graphical model, nodes represent variables and edges represent any dependencies among them. Research questions in Clinical Epidemiology and other areas of health sciences concern associations and relationships among a large number of variables (socio-demographic, clinical factors, intermediate outcomes, final outcomes etc.) that interact in a highly complex manner. Although GM is a promising tool for deciphering and investigating these relationships, their application to this type of studies is limited. The objective of this study is to demonstrate the utility of GM in clinical research studies. First, a comprehensive literature review was conducted in order to examine and understand the current use of graphical models in clinical and epidemiologic studies. Second, a case study was conducted using Bayesian Networks, a type of “directed” GM, to investigate the relationship between patient characteristics, sleep disturbance and quality of life among patients post-liver transplant. The literature review, executed through OVID MedLine, identified 68 articles up to February 2018, with most articles published after 2013 and involving Bayesian Network methodology. Case study showed that depression and fatigue predicted poor mental health related quality of life while depression and insomnia predicted poor physical health related quality of life. In addition, diagrams from trained Bayesian Networks enable clinicians to examine the strength and direction of the causal relationship between any two variables. Altogether, the literature review and case study suggest that GM and especially Bayesian Networks is an effective tool for exploring relationships in complex healthcare studies.