Going through directional changes: evolving human movement classifiers using an event based encoding

Directional changes (DC) is an event based encoding for time series data that has become popular in financial analysis, particularly within the evolutionary algorithm community. In this paper, we apply DC to a medical analytics problem, using it to identify and summarise the periods of opposing directional trends present within a set of accelerometry time series recordings. The summarised time series data are then used to train classifiers that can discriminate between different kinds of movement. As a case study, we consider the problem of discriminating the movements of Parkinson's disease patients when they are experiencing a common effect of medication called levodopa-induced dyskinesia. Our results suggest that a DC encoding is competitive against the window-based segmentation and frequency domain encodings that are often used when solving this kind of problem, but offers added benefits in the form of faster training and increased interpretability.