Heart Rate Monitoring from Wrist-Type Photoplethysmographic (PPG) Signals During Intensive Physical Exercise Zhilin Zhang Samsung Research America – Dallas 1301 E. Lookout Dr., Richardson, TX 75082, USA zhilinzhang@ieee.org Abstract—Heart rate monitoring from wrist-type photoplethys- mographic (PPG) signals during subjects’ intensive exercise is a difficult problem, since the PPG signals are contaminated by extremely strong motion artifacts caused by subjects’ hand movements. In this work, we formulate the heart rate estimation problem as a sparse signal recovery problem, and use a sparse signal recovery algorithm to calculate high-resolution power spectra of PPG signals, from which heart rates are estimated by selecting corresponding spectrum peaks. To facilitate the use of sparse signal recovery, we propose using bandpass filtering, singular spectrum analysis, and temporal difference operation to partially remove motion artifacts and sparsify PPG spectra. The proposed method was tested on PPG recordings from 10 subjects who were fast running at the peak speed of 15km/hour. The results showed that the averaged absolute estimation error was only 2.56 Beats/Minute, or 1.94% error compared to ground- truth heart rates from simultaneously recorded ECG. Index Terms—Photoplethysmographic (PPG) Signals , Singular Spectrum Analysis (SSA), Sparse Signal Recovery, Heart Rate Estimation, Wearable Computing I. I NTRODUCTION Heartbeat rate monitoring during fitness is a key feature in many modern wearable devices such as Samsung Gear Fit. These devices generally record photoplethysmographic (PPG) signals [1] from wearers’ wrist, and then estimate heart rates from the PPG signals in real time. However, PPG signals are vulnerable to motion artifacts, which is adverse to heart rate monitoring during fitness (see Figure 1 for example). Many signal processing techniques have been proposed to remove motion artifact (MA) from raw PPG signals. One technique is independent component analysis (ICA). For example, Kim et al. [2] suggested to use a basic ICA algorithm and block interleaving to remove MA. Krishnan et al. [3] proposed using frequency-domain based ICA technique to remove MA. However, this technique has some limitations. One limitation is that the key assumption in basic ICA algorithms, namely statistical independence or uncorrelation, is not hold in PPG applications [4]. Thus MA removal is not satisfactory especially when MA is strong. Another popular technique is adaptive noise cancelation (ANC) [5], [6]. For example, Ram et al. [5] proposed using ANC to remove MA, where the reference signal was con- structed from fast fourier transform (FFT), singular value de- 0 200 400 600 800 1000 -40 -20 0 20 40 60 (a) PPG without Motion Artifacts Sample Index 0 70 140 210 0 50 100 150 (b) Periodogram of PPG in (a) Beat Per Minute (BPM) 0 200 400 600 800 1000 -150 -100 -50 0 50 100 (c) PPG with Motion Artifacts Sample Index 0 70 140 210 0 50 100 150 200 (d) Periodogram of PPG in (c) Beat Per Minute (BPM) Fig. 1. Comparison between an artifact-free PPG signal and an artifact- contaminated PPG signal. Plot (a) and Plot (b) show the waveform and the spectrum (calculated by Periodogram) of a PPG signal without motion artifacts, respectively. Plot (c) and Plot (d) show the waveform and the spectrum (calculated by Periodogram) of a PPG signal with strong motion artifacts, respectively. The circles in Plot (b) and Plot (d) indicate the spectrum peaks corresponding to the heart rate (calculated from simultaneously recorded ECG). From Plot (d) we see selecting the spectrum peak corresponding to the heart rate is difficult. composition, or ICA of the artifact-contaminated PPG signal. However, one limitation in this method is that the artifact- removal performance of ANC is sensitive to the reference signal, while reconstructing qualified reference signals is ex- tremely difficult when subjects are exercising. It is worth noting that most of these techniques were pro- posed for the scenarios when small motion movements were incurred, such as finger movements [3], [5], [6], walking [6], or slowly running (with the speed less than 8 km/hour) [7], [8]. Due to the limitations stated above, these proposed techniques may not be effective when subjects perform intensive physical exercise such as fast running. This work proposes an approach to estimate heart rates when subjects perform intensive physical exercise. We first formulate the heart rate estimation problem into a sparse signal reconstruction problem. To facilitate the use of sparse signal reconstruction algorithms, singular spectrum analysis (SSA) and temporal difference operation are proposed to partially GlobalSIP 2014: Advances in Signal Processing for Mixed-Signal and Optical Sensing: Hardware to Algorithms 978-1-4799-7088-9/14/$31.00 ©2014 IEEE 866