Effectiveness of Curvature and Signal Derivatives in Fast Curve Segmentation

Gordon Dickers (University of Wales Trinity Saint David, United Kingdom (Great Britain)); John Rees (University of Wales Trinity St David, United Kingdom (Great Britain)); Tim Bashford (University of Wales Trinity Saint David, United Kingdom (Great Britain)); Oluwakemi Ademoye (University of Kent, United Kingdom (Great Britain))

Segmentation of a plane curve or times series data is an important step in classification and region localisation. Here we investigate the effectiveness of curvature and derivatives of curves as features for a time series classifier. We evaluate their effect on the accuracy of a long short term memory recurrent neural network used to segment electrocardiogram signals and identify useful techniques to improve run-time efficiency. We compare our results with existing features pre-processors that have been used in the literature and present summary results of their relative speed and accuracy. We find that using curvature, first and second Gaussian derivatives can produce a significant speed up when pre-filtering of a sampled dataset is required. When used in combination, first and second derivatives improve classification accuracy by up to 24% when compared with the un-processed signal and, in terms of speed, they outperform by orders of magnitude the second best classifier's execution time.

Journal: International Journal of Simulation- Systems, Science and Technology- IJSSST V22

Published: Oct 7, 2021

DOI: 10.5013/IJSSST.a.22.01.12