Identifying Latent Variables in Dynamic Bayesian Networks with Bootstrapping Applied to Type 2 Diabetes Complication Prediction

Abstract

Background: Type 2 Diabetes is a chronic disease with an onset that is commonly associated with multiple life-threatening co morbidities (complications). Early prediction of diabetic complications while discovering the behaviour of associated aggressive risk factors can reduce the patients’ suffering time. Therefore, models of the time series diabetic data (which are often imbalanced, incomplete and involve complex interactions) are needed to better manage diabetic complications. Aims: The aim of this work is to both deals with imbalanced clinical data using a bootstrapping approach, whilst determining the precise position of latent variables within probabilistic networks generated from the observations. The main motivation behind this paper is to stratify patient groups by means of latent variables to discover how complications in diabetes interact. Methods: We propose a time series bootstrapping method for building Dynamic Bayesian Networks that includes hidden / latent variables, applied to a case for predicting T2DM complications. A combination of the IC* algorithm on time series bootstrapped data is utilized to identify the latent variables within Bayesian model. Then, an exploration of inference methods assessed the influences of these latent variables. Results: Our promising findings show how this targeted use of latent variables improves prediction accuracy, specificity, and sensitivity over standard approaches as well as aiding the understanding of relationships between these latent variable sand disease complications / risk factors. The contribution of this paper compared to the previous papers in which time series bootstrapping is used for re-balancing the data and providing confidence in the prediction results. Conclusion: Our results showed that our re-balancing approach by the use of Time Series bootstrapping method for an unequal number of time series visits demonstrated an improvement in the prediction performance. Additionally, the most highlighted contribution of this paper gained insight by interpreting the latent states (looking at the associated distributions of complications), which led to a better understanding of risk factors and patient-specific interventions: here the fact that the latent variable demonstrated that a patient falls into a sub-group that is hypertensive but not suffering from retinopathy.