Online Program

Return to main conference page
Thursday, May 17
Bayesian Modeling
Thu, May 17, 3:00 PM - 3:45 PM
Regency Ballroom B
 

Bayesian Optimization of Personalized Models for Real-Time Patient Monitoring (304721)

*Glen Wright Colopy, Oxford University  

Keywords: Machine Learning, Healthcare, Personalized Healthcare, Time series, Bayesian, Bayesian Optimization, Optimization, Gaussian Processes

Personalized patient modeling promises many improvements over population-based models. Patient vital-sign time series are already monitored as standard practice in many clinical settings. Gaussian process regression (GPR) can flexibly model time series for personalized clinical inference that is lost on population-based approaches. Personalized time series modeling has its own challenges, since these models must (i) accommodate the plausible physiology of any patient in the population (via a wide-range of parameter values or vague regularizing priors), but also (ii) expeditiously learn personalized parametrization or regularizers for timely clinical inference. We demonstrate how Bayesian optimization can sequentially build personalized GP models over patient time series via covariance kernels and their parameters. The Bayesian optimization approach efficiently samples the model search space, which is expensive, non-convex, and black-box with high effective dimensionality. The resulting personalized models are superior at useful clinical tasks, such as forecasting and advanced detection of clinical emergencies.