Performance of Spectral, Autocorrelation and Peak Count Based PR Estimation Methods Under Normal/Abnormal PPG for Wearable Devices

In most wearable devices, pulse rate (PR) is measured from photoplethysmogram (PPG) signal by using fast Fourier transform (FFT), autoregressive (AR), autocorrelation function (ACF) and peak counting methods. In this paper, we study accuracy and reliability of PR estimation by highlighting major drawbacks of these methods and optimal selection of measurement duration (10, 15, 30 or 60 s) for accurate measurement of PR from different morphological pulsatile patterns of normal and abnormal PPG signals. Evaluation results on the PPG signals taken from seven databases with two sampling rates (125 Hz and 25 Hz) demonstrate that the amplitude threshold (ATh) based PR estimation method is more accurate and lightweight as compared to the existing FFT-, AR- and ACF-based methods in terms of mean absolute error (MAE), Pearson coefficient (PC), Bland-Altman ratio (BAR), ve error ranges (ERs) (0-1, 2-5, 6-10, 11-25,>25 bpm) and computational load. The ATh- based method had the performance of MAE=0.61 bpm, PC=0.996, BAR=3.7% and ER>5 bpm of 1.57%, and outperforms the MAE, PC, BAR and ER>5 bpm values respectively of 2.74 bpm, 0.9432, 14.4% and 11.75% for FFT-based, the values of 3.96 bpm, 0.8028, 36.8%, and 8.07% for AR-based and the values of 2.15 bpm, 0.9607, 12.1% and 7.94% for ACF-based method. Results showed that the sampling rate and measurement duration impact the accuracy and computational load.