Linking Bayesian Network and Intensive Care Units Data: A Glycemic Control Study

Health informatics in glycemic control is visibly a promising research area. However, this applied science requires more intelligent mechanisms by which user requirements for more accurate prediction can be fulfilled. Such mechanisms must provide very flexible and user friendly procedures to enable complicated decision support functions. This article presents the linking process of per-patient demographic and admission to intensive care unit data with their glycemic control performance using probabilistic causal Bayesian Network models (BNs). Data from two glycemic control protocols are exploited to test the feasibility. The identified steps crucial in building a dependable model are variable selection, state discretization, and structure learning. Different BNs can be generated with more than 83.73% overall precision rate and 93.4% overall calibration index with the combination of these steps. A network with a 95.36% precision was obtained with an equal distance discretization algorithm dataset and Maximum Weight Tree Spanning unsupervised structure learning. The study was the first testing phase in which the results generated by selected data and process is proposed as a benchmark. The resulting network is centred on 'Hypertension' status to predict BG mean and number of measurements as a result of the prediction interest. This co-morbidity is proposed to be considered systematically in the modelling of any glycemic control to optimize its function in the intensive care units.