Bayesian Network Inference on Diabetes Risk Prediction Data

Diabetes mellitus is a common and serious disorder that affects multiple organs in humans. The detection of early signs and symptoms are important to control the side effects of diabetes. In our study, we investigate cause-effect relations between symptoms of diabetes and disease status using Bayesian networks (BNs). BNs are directed acyclic graphs (DAGs) that gives information about regulator-regulated relations. We applied greedy climbing hill algorithm (GCHA) as a score-based BN inference algorithm, grow shrink (GS), and iterative associate Markov blanket (IAMB) algorithms as constraint-based BN inference algorithms on early-stage diabetes risk prediction dataset, which consists of 16 possible symptoms and diabetes status. There are two main objectives of the study: to infer relationships between variables and to compare the performance results of the score-based and constraint-based algorithms. The most dense BN structure is obtained by GCHA-based DAG and the most loosely coupled BN structure is obtained by GS-based DAG. Though GCHA infers many cause-effect relations, the performance of constraint-based algorithms are higher than GCHA. The results of our study are also compatible with the results of different studies in the literature.