The combination of expert knowledge and Bayesian networks to identify the relevant factors in diabetic kidney disease (DKD)

Type II diabetes mellitus (T2DM) is a major health issue in Europe, with and increasing number of affected individuals developing diabetic kidney disease (DKD). The consequences of DKD can be very severe and challenging to manage, impacting patients’ quality of life and increasing the cost of care and hospitalization. DKD in T2DM is complex, with many different components characterized by inter-individual and intra-individual heterogeneity of pathophysiology at the molecular level and in its phenotype. Therefore, it is crucial to derive in-depth information on what characterizes its pathophysiology and pattern of disease progression. Statistical models can assist in this task by uncovering the interconnections among variables associated with the disease. In this work we used Bayesian networks, a class of probabilistic graphical models, to learn from available data the robust relationships among a set of variables and to determine which variables and combinations of variables are important for disease diagnosis, treatment choice and prognosis. By incorporating expert knowledge in the modeling phase, we demonstrated that the synergy between data and expert prior information is a great source of valuable help in gaining new insights into complex systems and in understanding DKD with respect to a predefined target of interest. Our results indicated that more than half of the variables identified as relevant to the target of interest are not part of the a priori pathophysiology theoretical framework identified by clinical experts. This has provided new evidence regarding the pathophysiology of the disease, highlighting how statistical methods and data-driven technologies can support expert decisions with new knowledge.