A Framework for PPG-to-ECG Feature Estimation Based on Continuous Wavelet Transform


Student thesis: Doctoral Thesis

View graph of relations


Related Research Unit(s)


Awarding Institution
Award date4 Jun 2020


We have developed a framework for the estimation of electrocardiography (ECG) signal features via the photoplethysmography (PPG) signal. ECG signal features are the gold standard for the diagnosis of many heart problems, but limitations such as their large size, high cost and inconvenience of use render existing ECG measurement devices unsuitable for long-term daily use. PPG uses an optical method to detect a pulse signal in the human body. Numerous advantages such as their compact structure, low price and ease of operation allow the extensive use of PPG measurement systems in portable devices for real-time continuous heart rate monitoring. Nevertheless, the heart rate alone cannot provide details about cardiac function. One obvious solution to this dilemma is to determine the correlations between PPG and ECG signal features and use the PPG signal as a surrogate for the ECG signal. 
In this framework, the detection of both PPG and ECG feature points is based upon single-scale continuous wavelet transform (CWT) with Mexican hat (mexh) wavelet (mexh-CWT). The algorithm for detection of PPG feature points includes three main steps: mexh-CWT of the original PPG signal at scale 20, normalization of the CWT coefficients into the range [-1, 1] and application of the decision rules to the normalized CWT coefficients. The mexh wavelet at scale 20 is the optimal choice for the PPG signals sampled at 500 Hz in the Vortal Dataset; the PPG signals sampled at 300 Hz in the CapnoBase Dataset must first be resampled to 500 Hz. The main steps in detection of ECG feature points are similar to those for PPG feature points except for the optimal wavelet scale. The optimal wavelet scale is 8 for the ECG signals sampled at 500 Hz in the Vortal Dataset. The ECG signals sampled at 300 Hz in the CapnoBase Dataset will also be resampled to 500 Hz.
We first investigated the correlations between the three PPG feature points (i.e., Foot point [Ft], Rapid ejection peak [Pk1] and Reduced ejection peak [Pk2]) and the five ECG feature points (i.e., P wave onset [Pon], P wave peak [Ppk], R wave peak [Rpk], T wave peak [Tpk] and T wave offset [Toff]) in the 26 healthy young subjects of the Vortal Dataset. Due to the commonalities in physiological conditions and measurement environments, the data normalized by the RR interval (i.e., ‘Rpk - Rpk interval’) show a high degree of consistency between the 26 subjects. We therefore chose to combine the normalized data from all 26 subjects and analyze them within a unified framework. Based upon these normalized data, 13 strong linear relationships (Pearson correlation coefficients > 0.8) were found between the paired PPG and ECG features. For example, the Pearson correlation coefficient between the normalized ‘Pk1 - Pk2 interval’ and the normalized ‘Rpk - Toff interval’ is 0.92. Better yet, we found a way to estimate the pulse transit time (PTT; i.e., ‘Rpk - Pk1 interval’) via the PPG signal alone: normalized PTT = 2.13 ∙ normalized ‘Ft - Pk1 interval’ + 0.09. The Pearson correlation coefficient between the normalized PTT and the normalized ‘Ft - Pk1 interval’ is 0.91. 
We also closely studied the 6 records in the CapnoBase Dataset. Unlike the Vortal Dataset, simultaneous PPG and ECG signals in the CapnoBase Dataset were recorded from clinical patients with diverse physiological conditions and measurement environments, so very large differences can be seen in the normalized data of various patients. For instance, whilst a strong linear relationship exists between the normalized PTT and the normalized ‘Ft - Pk1 interval’ for every patient, the fitting equation of the two features differs greatly among the 6 patients: for Patient-0029 (age, 10 years; weight, 46 kg), normalized PTT = 4.18 ∙ normalized ‘Ft - Pk1 interval’ - 0.09; for Patient-0330 (age, 59 years; weight, 70 kg), normalized PTT = 1.40 ∙ normalized ‘Ft - Pk1 interval’ + 0.09. We therefore chose to display and analyze the normalized data of the 6 patients in a separate way.
A comparative analysis between the Vortal Dataset and the CapnoBase Dataset reveals two common points: 1) the PPG features generally have a strong linear relationship with the ECG features in the ‘Rpk - Toff interval’ (i.e., left ventricular systole period), but not with the ECG features in the ‘Pon - Rpk interval’ (i.e., left atrial systole period); and 2) a strong linear relationship always exists between the normalized PTT and the normalized ‘Ft - Pk1 interval’.
Like many previous researchers, we validated our algorithm for detection of ECG feature points with the MIT-BIH Arrhythmia Database (MITDB). The results show that our algorithm can correctly detect 99.66% of the R wave peaks; the detection results herein are considered to be correct if within ± 75 ms of the expert annotations.
As we know, the correlations between PPG signal features and cardiovascular characteristics (e.g., cardiac output and arterial compliance) have been studied thoroughly for decades by many scientific organizations, and a variety of commercially available devices have been based upon PPG technology that can provide multiple forms of information on the cardiovascular system (e.g., heart rate and blood oxygen saturation). However, no study has yet precisely estimated ECG signal features via PPG signals on a beat-to-beat basis. In addition, a method to estimate PTT via a single PPG signal is not found in any published paper. Our research fills these two gaps. Furthermore, we believe that our findings will help to separate the cardiac factors from the vascular factors in the PPG signal, which will allow future PPG studies of various purposes to be more targeted.