Last updated on 03/11/2021
Tang C. (corresponding author). Submitted. “HitCompl: Non-Equidistant Dynamic Bayesian Networks for Risk Prediction Using Electronic Health Records.” Data Mining and Knowledge Discovery.
Patients with chronic diseases are reported to be at risk for unexpected complications which usually cause worsening disease severity: disabilities and even death. This study constructs an unsupervised framework on electronic health records (EHRs), which we call HitCompl, for understanding complication risks to reinforce the patients in managing disease progression. We first retrieve patients with a targeted chronic disease to cluster their exact diagnoses into several groups. Based on non-equidistant dynamic Bayesian networks, we then address the problem of tracking disease severity over time. That is, our approach models the time course of progressive disease status as the irregular key time steps, and then discovers causality among almost all complications relating to the targeted chronic disease. Experiments on real-world EHRs of 9,484 patients with diabetes derived interesting clinical insights on diabetes complications. Our results also demonstrate that the HitCompl framework is often on par with deep learning models in both accuracy and efficiency for training and evaluation.