Dynamic Time Warping of Pulse Wave Curves

Abstract

Abstract Being able to non-obtrusively and continuously monitor arterial blood pressure is of great interest, particularly in the context of wearable sensors. A common limitation is the need for dedicated hardware, which is either obtrusive or expensive. In our current work, we investigate unimodal pulse waves from three handily accessible heterogeneous sources: photoplethysmography, bioimpedance and pulse applanation tonometry. We derive and evaluate multiple parameters regarding their correlation to reference blood pressure curves. These parameters stem from features of the warping paths resulting from dynamic and derivative dynamic time warping. The warping is performed between adjacent pulses or to a reference waveform. Spearman Rho coefficients of up to 0.98 and averaging 0.77 at highly significant p-values are recorded for single parameters. We record mean absolute deviation values of 0.08 across subjects. The results indicate there are negligible lags between reference and parameter curves. The sign of the correlation coefficients is consistent only for a small subset of parameters; the underlying cause could not yet be identified. We conclude that the warping path approach seems a promising way to go, yet still needs refinement. In particular, developing a time and amplitude warping method is paramount. Since warping quantizes all the morphological changes in the pulse wave without fiducial point detection, it could become a powerful tool for future investigations.