A Dynamic Bayesian Network model for long-term simulation of clinical complications in type 1 diabetes

Display Omitted We use DCCT and EDIC longitudinal studies to design Dynamic Bayesian Network models for type 1 diabetes.The structure of our networks, as well as their conditional probability tables, are learned from the data.Both clinical values (HDL, LDL, BMI, etc.) and complications (CVD and nephropathy) are simulated.The simulations span over 15years, error on test data is below 10%.We present two models: one is entirely data driven, the other is mediated by medical knowledge. The increasing prevalence of diabetes and its related complications is raising the need for effective methods to predict patient evolution and for stratifying cohorts in terms of risk of developing diabetes-related complications. In this paper, we present a novel approach to the simulation of a type 1 diabetes population, based on Dynamic Bayesian Networks, which combines literature knowledge with data mining of a rich longitudinal cohort of type 1 diabetes patients, the DCCT/EDIC study. In particular, in our approach we simulate the patient health state and complications through discretized variables. Two types of models are presented, one entirely learned from the data and the other partially driven by literature derived knowledge. The whole cohort is simulated for fifteen years, and the simulation error (i.e. for each variable, the percentage of patients predicted in the wrong state) is calculated every year on independent test data. For each variable, the population predicted in the wrong state is below 10% on both models over time. Furthermore, the distributions of real vs. simulated patients greatly overlap. Thus, the proposed models are viable tools to support decision making in type 1 diabetes.