Current methods for predicting patient outcomes based on time-series data, such as vital signs, typically rely on comparing individual measurements to a population-based “normal” model and triggering an alert if the measurements deviate significantly from that distribution.
This is known as novelty detection, and it overlooks the dynamics of multivariate time-series data by treating them as a series of point measurements.
The Oxford method leverages the accumulation of vast amounts of patient data, including multivariate vital-sign data and other time-series data from daily laboratory tests and other healthcare settings.
Compared to existing methods, this approach provides a significant performance boost by accurately predicting patient deterioration rather than simply identifying it. It is particularly useful for wearable sensors in the rapidly expanding non-clinical market, where existing methods lack robustness. It can also be applied to video-based monitoring data.