Deep PPG: Large-Scale Heart Rate Estimation with Convolutional Neural Networks

Abstract

Photoplethysmography (PPG)-based continuous heart rate monitoring is essential in a number of domains, e.g., for healthcare or fitness applications. Recently, methods based on time-frequency spectra emerged to address the challenges of motion artefact compensation. However, existing approaches are highly parametrised and optimised for specific scenarios of small, public datasets. We address this fragmentation by contributing research into the robustness and generalisation capabilities of PPG-based heart rate estimation approaches. First, we introduce a novel large-scale dataset (called PPG-DaLiA), including a wide range of activities performed under close to real-life conditions. Second, we extend a state-of-the-art algorithm, significantly improving its performance on several datasets. Third, we introduce deep learning to this domain, and investigate various convolutional neural network architectures. Our end-to-end learning approach takes the time-frequency spectra of synchronised PPG- and accelerometer-signals as input, and provides the estimated heart rate as output. Finally, we compare the novel deep learning approach to classical methods, performing evaluation on four public datasets. We show that on large datasets the deep learning model significantly outperforms other methods: The mean absolute error could be reduced by 31 % on the new dataset PPG-DaLiA, and by 21 % on the dataset WESAD.

Photoplethysmographic Time-Domain Heart Rate Measurement Algorithm for Resource-Constrained Wearable Devices and its Implementation

“…The proposed simplified peak detection algorithm was compared to the original AMPD algorithm presented in [19]. For this purpose, the PPG signals from the PPG-DaLia database [15] were used, and the detected peaks were compared. As the original AMPD peak detection algorithm uses a random variable to fill the LMS matrix, its results are slightly different from run to run, depending on the seed value of the random number generator.…”

“…For the purpose of evaluating the algorithm, the PPG-DaLia database [15], containing more than 35 h of data recorded from 15 persons, was used. The database contains signals collected from a PPG sensor, accelerometer, and ECG, where the ECG is used as ground truth.…”

“…The database contains signals collected from a PPG sensor, accelerometer, and ECG, where the ECG is used as ground truth. As described in [15], ground truth heart rate values were obtained from an ECG signal processed by R-peak detection algorithm [21]. Then, the detection results were manually inspected and corrected, mainly for a few cases where significant motion was observed.…”

“…is the reference heart rate obtained from the ECG ground truth signal for the same window j. This evaluation method is commonly used in related work [14,15,22,23].…”

“…A neural network adaptive whitening filter to model the lower frequencies of the signal is presented in [12]. Much recent work is concentrated on methods using time-frequency spectra [13,14] and deep-learning approaches [15][16][17]. A review of the current state-of-the-art signal-processing techniques for HR estimation from a wrist PPG signal can be found, for example, in [4,18].…”