Research Article: State-dependent Gaussian kernel-based power spectrum modification for accurate instantaneous heart rate estimation

Date Published: April 5, 2019

Publisher: Public Library of Science

Author(s): Heewon Chung, Hooseok Lee, Jinseok Lee, Lars Kaderali.


Accurate estimation of the instantaneous heart rate (HR) using a reflectance-type photoplethysmography (PPG) sensor is challenging because the dominant frequency observed in the PPG signal corrupted by motion artifacts (MAs) does not usually overlap the true HR, especially during high-intensity exercise. Recent studies have proposed various MA cancellation and HR estimation algorithms that use simultaneously measured acceleration signals as noise references for accurate HR estimation. These algorithms provide accurate results with a mean absolute error (MAE) of approximately 2 beats per minute (bpm). However, some of their results deviate significantly from the true HRs by more than 5 bpm. To overcome this problem, the present study modifies the power spectrum of the PPG signal by emphasizing the power of the frequency corresponding to the true HR. The modified power spectrum is obtained using a Gaussian kernel function and a previous estimate of the instantaneous HR. Because the modification is effective only when the previous estimate is accurate, a recently reported finite state machine framework is used for real-time validation of each instantaneous HR result. The power spectrum of the PPG signal is modified only when the previous estimate is validated. Finally, the proposed algorithm is verified by rigorous comparison of its results with those of existing algorithms using the ISPC dataset (n = 23). Compared to the method without MA cancellation, the proposed algorithm decreases the MAE value significantly from 6.73 bpm to 1.20 bpm (p < 0.001). Furthermore, the resultant MAE value is lower than that obtained by any other state-of-the-art method. Significant reduction (from 10.89 bpm to 2.14 bpm, p < 0.001) is also shown in a separate experiment with 24 subjects.

Partial Text

In recent years, instantaneous heart rate (HR) estimation has attracted considerable attention owing to the advent of wearable devices such as wristwatches and bands that can be used to obtain photoplethysmographs (PPGs). At present, various commercially available reflectance-type wrist-worn PPG devices, such as Apple Watch, Fitbit Surge, and Samsung Gear, are capable of producing instantaneous HR estimates. However, the accuracy of most of these devices is limited to situations in which the wearer is at rest, walking, or performing low-intensity exercise. During high-intensity exercise, the measured PPG signals are severely corrupted by motion artifacts (MAs) that are shaped similarly to pure pulses, which cause the dominant frequency in the PPG signal to deviate from the true HR. In addition, it is challenging to detect the pure pulse peak for estimating the HR. Here, severe corruption implies that motion artifacts with larger amplitude than the pure pulse are coupled with the PPG signal, making it difficult to distinguish the actual pure pulse [1–3].

In the pre-processing stage, we down-sampled the signals to 25 Hz. Regarding the effects of the sampling rate, it was reported that the HR estimation performance was nearly the same but the computational time was drastically reduced by down-sampling the signals [20]. Moreover, in [2], the HR estimation performances were compared at different sampling frequencies of 25, 125, 250, and 500 Hz. The results showed that the HR estimation results were similarly accurate for the first 12 subjects of the ISPC dataset, i.e., MAEs of 1.02 bpm (25 Hz), 1.06 bpm (125 Hz), 1.10 bpm (250 Hz), and 1.12 bpm (500 Hz). Our results also showed consistent trends. Even with a different sampling rate, the HR estimation performance was nearly the same. The MAE values with a sampling rate of 125 Hz were 1.21 bpm and 2.15 bpm for the ISPC and BAMI datasets, respectively. Note that the MAE values with a sampling rate of 25 Hz were 1.20 bpm and 2.14 bpm for the ISPC and BAMI datasets, respectively. As down-sampling reduces the computational load without accuracy degradation, most existing algorithms including our method down-sampled the signals to around 25 Hz [1–3, 20–25, 27].