An Interpretable Distance Measure for Multivariate Non-Stationary Physiological Signals

We introduce d
symb
, a novel distance measure for comparing multivariate non-stationary physiological signals. Unlike most distance measures on multivariate signals such as variants of Dynamic Time Warping (DTW), d
symb
can take into account their non-stationarity thanks to a symbolization step. This step is based on a change-point detection procedure, that splits a non-stationary signal into several stationary segments, followed by quantization using K-means clustering. The proposed distance measure leverages the general edit distance that is applied to the symbolic sequences. The performance of d
symb
compared to two commonly used DTW variants is illustrated by applying it to physiological signals recorded during walking protocols. In particular, d
symb
is shown to be interpretable: its symbolization detects the segments that correspond to salient behaviors. An open source GitHub repository is made available to reproduce all the experiments in Python.