Large-scale Multi-output Gaussian Processes for Clinical Decision Support

Abstract

In the scenario of real-time monitoring of hospital patients, high-quality inference of patients’ health status using all information available from clinical covariates and lab tests is essential to enable successful medical interventions and improve patient outcomes. Developing a computational framework that can learn from observational large-scale electronic health records (EHRs) and make accurate real-time predictions is a critical step in achieving this goal. However, existing EHRs pose several challenges for conventional methods. For instance, many covariates are sparsely sampled in time across patients. In addition, there are substantial uncertainties in patient condition and disease progression at any time. These properties make inferring the physiological status of a patient or joint analysis of time series across patients challenging.

This dissertation first presents MedGP, a statistical framework that provides accurate real-time predictions of physiological states. MedGP is based on multi-output Gaussian processes, a Bayesian nonparametric model, to capture temporal structures between clinical covariates from noisy and irregularly sampled time series data. MedGP has a number of benefits over current methods. First, alignments of time series across patients are not required. Next, MedGP learns robust and interpretable relationships across covariates through regularization. Lastly, to tackle the computational bottlenecks, an approximate inference approach is derived and shown to achieve orders of speedup. The empirical results show significant improvements over baselines in making online predictions on two real-world medical data sets consist of tens of thousands of patients and millions of observations.

Finally, two extended frameworks based on MedGP are explored to enhance clinical decision-making processes, including unifying with reinforcement learning algorithms for action recommendation and designing non-stationary kernels to learn the underlying dynamics of clinical treatments. Both methods demonstrate encouraging results on improving clinical practices and advancing towards precision medicine.