Identifying patient-specific trajectories in haemodialysis using Bayesian Hierarchical Gaussian Processes

In the domain of chronic disease monitoring, continuous time-series vital sign data (such as blood pressure) can be collected through low-cost wearable devices. Approximately 300 million people have chronic kidney disease globally. These patients undergo multiple haemodialysis sessions per week. In addition, patients are at risk of intra-dialytic hypotension, which leads to chronic heart disease and a high incidence of mortality. We propose Bayesian Hierarchical Gaussian Processes (HGPs) to model changes of systolic blood pressure (SBP) over time for continuously monitored patients. Furthermore, we use Bayesian HGPs to infer the hidden latent structure of the SBP time-series pattern/trajectory for each individual patient. To identify normal versus abnormal patterns, we further propose symmetric Kullback-Leibler divergence of multivariate normal distributions to provide a metric to identify deviations from normal latent trajectories. We apply Bayesian HGPs to a dataset of patients undergoing haemodialysis monitoring, and demonstrate its superiority in identifying abnormal SBP patterns compared to alternative state-of-the-art clustering algorithms.