Evaluation of HRV from Repeated Measurements of PPG and Arterial Blood Pressure Signals

Andrejs Fedjajevs1, Willemijn Groenendaal1, Lars Grieten2, Carlos Agell1, Pieter Vandervoort2, Evelien Hermeling3
1imec, 2Future Health Department, Ziekenhuis Oost-Limburg, 3imec NL


Background: heart rate variability (HRV) is a valuable non-invasive indicator of autonomic nervous system. Guidelines prescribe that HRV is calculated from RR intervals of the ECG signal. Nowadays with the development of wearables and the need for unobtrusiveness photoplethysmography (PPG) inter-beat intervals (IBI) are considered as a viable HRV source.

Aims: in this work we compare short-time HRV from ECG and two pulse sources: PPG and arterial blood pressure (ABP) – the gold reference invasive pulse signal. We define the optimum characteristic point to use for IBI extraction from each pulse signal. We calculate errors for individual time- and frequency-domain metrics, compare those between pulse modalities and relate to physiological variations across multiple measurements.

Methods: 69 five min rest data (ECG, PPG, ABP) was extracted from 25 healthy volunteers measured three times within two hours. Each signal was annotated by automatic beat detector followed by visual inspection. Pearson’s correlation, mean absolute\relative errors and Bland-Altman analysis were used to study the level of agreement between traditional and surrogate HRV metrics.

Results: The upstroke is the best location for both pulse signals yielding the mean IBI error of 3.7 and 3.0 ms. ABP is superior to PPG, but both can be used for HRV of healthy subjects at rest, though not all metrics are equally reliable. High frequency power and pnn50 exhibit the highest relative error difference between the two modalities: 22% and 30% for PPG vs 16% and 12% for ABP, accordingly. Other frequency- and time-domain metrics are less than 4% different with relative errors below 8%. Bland-Altman plots reveal steeper regression trend for PPG-based HRV and tendency to overestimate low RMSSD and SDNN values more. Intra-subject variation (between sessions) on average exceeds surrogate HRV differences: RMSSD and SDNN may change by 30% and absolute power features – more than 100%.