Local Gaussian Process Features for Clinical Sensor Time Series
C. Scott Brown, Ryan Benton · 2019
Time series obtained from clinical sensors such as EEG and Accelerometers often contain mixtures of complex signals that can be difficult to interpret and require advanced methods for automated analysis. In this work, we introduce a flexible and extensible method for transforming clinical time series into new, simpler features that may aid in automated analysis and interpretation. The method is based on information extracted from locally-trained Gaussian Processes. In this work, we describe the procedure, illustrate the idea on contrived data and demonstrate its effectiveness at improving existing methods for both epilepsy detection and activity classification. We show that our feature extraction technique generalizes, and can be used as a preprocessing step for arbitrary machine learning and statistical methods. We further illustrate that the techniques evaluated in our experiments form only a small subset of a very broad class of feature extraction methods on arbitrary data that might form the basis for further study.