Heart rate variability (HRV) is a powerful measure to gain information on the activation of the central nervous system and is thus a strong indicator for the overall health and emotional state of a person. Currently, the gold standard for HRV analysis is the examination of R-peaks in electrocardiograms (ECG), which requires a placement of electrodes on the torso. This is often impracticable, especially for the use in daily routines or 24/7 measurements. Photoplethysmograms (PPG) are an alternative to ECG assessment and are easier to acquire, e.g. by using fitness trackers or smart watches. Nevertheless, PPG data is more susceptible to motion artifacts. Hence, goal of this work is to develop and evaluate an artificial neural network (ANN) approach to estimate the R-peak locations in complex PPG signals. Public data collections were used as benchmark to compare our ANN-based approach to state-of-the-art methods. Results show that ANNs can improve HRV estimation during motion. HRV estimations from baseline methods (decision-tree based and automatic multiscalebased peak detection) were compared with the best performing neural network (3L-GRU) using the TROIKA dataset with respect to reference parameters obtained from a manual selection of the peaks in ECG data. In most cases, the neural network based HRV estimation was closer to the reference HRV compared to baseline methods (lower μ and σ ) Also, σ is smaller for the best performing ANN approach across most HRV parameters. Inclusion of another PPG or acceleration channel did not affect HRV estimation. Although, the neural network learning approach outperforms conventional methods, the examined PPG-based HRV estimation has still accuracy limitations. Nonetheless, the proposed estimation approach opens up new directions for further improvement.
© 2020 by Walter de Gruyter Berlin/Boston
This work is licensed under the Creative Commons Attribution 4.0 International License.