Gaussian Processes for Advanced Warning of Patient Deterioration

• Activity Number: 552
• Type: Contributed
• Date/Time: Wednesday, August 12, 2015 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistical Learning and Data Mining
• Author(s): Glen Colopy* and Stephen J. Roberts and David A. Clifton
• Companies: University of Oxford and University of Oxford and IBME - University of Oxford
• Keywords: Gaussian Processes ; Nonparametric ; Bayesian ; Medical ; Statistical Learning ; Machine Learning

Abstract

1-in-7 electronically monitored patients in the stepdown ward suffer cardiac arrest or emergency readmission to ICU. Advanced detection of patient deterioration could prevent many of these adverse events. The current state-of-the-art relies on heuristic track-and-trigger Early Warning Scores, or Parzen Window novelty thresholds on a patient's vital signal measurements (VSMs). Both methods assume an i.i.d. relationship between a patient's VSMs and can only assess a patient's present risk but not forecast future risk. Gaussian Process Regression (GPR) provides a solution for both the i.i.d. assumption and the need to forecast: The dependency between VSMs within a timeseries are modelled by the GPR covariance function. The plausible range of future VSM values are represented as a distribution over a function of time. A robust fitting of the data can be further improved by placing prior distributions over the parameters of the covariance function, and augmenting the covariance function to decipher physiological trends from measurement noise. Once fit to a patient VSMs, GPR is amenable to both traditional regression or timeseries-clustering to improve upon the current state-of-the art.