State-space approximations to Gaussian processes for patient vital-sign monitoring in computationally-constrained clinical environments

Gaussian processes (GPs) define a probability distribution over a space of functions from which a set of observed data are assumed to be generated. When applied to a time-series of patient vital-sign data, GP models (i) can encode prior clinical knowledge of the dynamics of the data; (ii) are patient-specific; and (iii) can be learned in real-time. The clinical value of GPs [1], [2] has been demonstrated by their superior performance in advanced warning of deterioration compared to the current clinical practice of heuristic thresholding methods, as well as in comparison to methods based on kernel density estimates (KDEs) [3]. The latter [3], which represents the current state-of-the-art in clinical practice assume that vital-sign measurements are independently and identically distributed samples from a population of healthy patients. Despite the three advantages listed above, GPs are computationally intensive, which precludes their use in computationally-constrained clinical environments. Such clinical environments may include remote or at-home patient monitoring, ambulatory patient monitoring, and monitoring in rural [4] or low-income country (LIC) [5] healthcare settings. State-space approximations of GPs (implemented as Kalman filters) are proposed to provide the aforementioned benefits of GPs while being computationally lightweight such that their use is feasible in computationally-constrained environments. Results demonstrate that our proposed approach to patient monitoring can be further exploited to allow for patient-specific model selection in real-time (which is difficult to achieve with standard Gaussian processes)