2016 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 64, NO.

9, SEPTEMBER 2017

Accurate Heart Rate Monitoring During Physical

Exercises Using PPG
Andriy Temko∗ , Senior Member, IEEE

Abstract—Objective: The challenging task of heart rate smart-watches or wristbands, HR monitoring can guide exercis-
(HR) estimation from the photoplethysmographic (PPG) sig- ers to adapt their training load and better match their training
nal, during intensive physical exercises, is tackled in this goals [26]. PPG signals have become a popular alternative to tra-
paper. Methods: The study presents a detailed analysis of
a novel algorithm (WFPV) that exploits a Wiener filter to at- ditional Electrocardiography (ECG) based HR estimation which
tenuate the motion artifacts, a phase vocoder to refine the measures the bio-potential generated by electrical signals that
HR estimate and user-adaptive post-processing to track control the expansion and contraction of heart chambers. How-
the subject physiology. Additionally, an offline version of ever, ECG requires the presence of ground and reference sensors
the HR estimation algorithm that uses Viterbi decoding is that must be attached to the chest. PPG-based HR monitoring at
designed for scenarios that do not require online HR mon-
itoring (WFPV+VD). The performance of the HR estimation the peripheral positions such as earlobes, fingertips or wrists is
systems is rigorously compared with existing algorithms seen as a much more convenient solution.
on the publically available database of 23 PPG recordings. The PPG signals [1]–[3], [29], [30] come from PPG sensors
Results: On the whole dataset of 23 PPG recordings, the al- which are embedded in these wearable devices. A PPG sensor
gorithms result in average absolute errors of 1.97 and 1.37 emits light to the skin and measures the changes of intensity of
BPM in the online and offline modes, respectively. On the
test dataset of 10 PPG recordings which were most cor- the light which is reflected or transmitted through the skin. The
rupted with motion artifacts, WFPV has an error of 2.95 BPM periodicity of these measurements in most cases corresponds
on its own and 2.32 BPM in an ensemble with two existing to the cardiac rhythm, and thus, HR can be estimated from the
algorithms. Conclusion: The error rate is significantly re- PPG signal.
duced when compared with the state-of-the art PPG-based Motion artifacts (MAs) are known to be a limiting factor
HR estimation methods. Significance: The proposed system
is shown to be accurate in the presence of strong motion ar- that prevents the straight-forward usage of PPG, especially in
tifacts and in contrast to existing alternatives has very few free living conditions. MAs are considered to result from sensor-
free parameters to tune. The algorithm has a low computa- tissue motion and sensor deformation. Strong movements during
tional cost and can be used for fitness tracking and health physical exercise make the HR estimate inaccurate as shown in
monitoring in wearable devices. The MATLAB implementa-
Fig. 1. Due to motion the sensors might situate far enough
tion of the algorithm is provided online.
apart from the skin that the true HR peak is absent in the PPG
Index Terms—Heart rate, photoplethysmographic, phase spectrum. A number of methods have been proposed to detect,
vocoder, spectrum estimation, viterbi decoding, wiener
filter. remove or attenuate MAs in PPG signals, including adaptive
filtering [4], [5], [22], [25], independent component analysis [6],
I. INTRODUCTION empirical mode decomposition [7], [23] or other decomposition
EARABLE devices have gradually increased their func- models [3], [27], spectral subtraction [8], [21], [24], and Kalman
W tionality over the last decades. Modern wearable de-
vices are equipped with a number of internal and external sen-
filtering [9].
A three-stage TROIKA method has recently been proposed
sors and can offer many useful fitness tracking features such as to estimate HR from PPG signals for scenarios where MAs
counting steps, calories, tracking sleep, etc. Photoplethysmog- are strong [2]. The method was based on signal decomposi-
raphy (PPG) based heart rate (HR) monitoring during physical tion, sparsity-based high-resolution spectrum estimation, and
exercise is one of these features [29], [30]. Implemented in spectral peak tracking and verification. The average absolute
error of 2.34 beats per minute (BPM) was reported on 12 PPG
Manuscript received January 4, 2017; revised February 23, 2017; ac- recordings. The TROIKA method was enhanced in [3] where the
cepted February 23, 2017. Date of publication March 1, 2017; date of spectra of PPG and acceleration signals were jointly estimated
current version August 18, 2017. This work was supported in part by a
Science Foundation Ireland Research Centre Award (12/RC/2272) and using a common sparsity constraint on the spectral coefficients
Wellcome Trust Seed Award in Science (200704/Z/16/Z). This paper was (JOSS). This was achieved by means of a multiple measurement
presented in part at EMBC 2015 and ICASSP 2017. Asterisk indicates vector model [10]. The error was reduced to 1.28 BPM when
corresponding author.
∗ A. Temko is with the Irish Centre for Fetal and Neonatal Transla- evaluated on the same 12 PPG recordings.
tional Research (INFANT), Department of Electrical and Electronic Engi- For the IEEE Signal Processing Cup 2015 the database
neeringUniversity College Cork, Cork, Ireland (e-mail: [email protected]). of 23 PPG recordings of people running on a tread-
This paper has supplementary downloadable material available at
http://ieeexplore.ieee.org (File size: <1 MB). mill or doing intensive physical exercises was made public
Digital Object Identifier 10.1109/TBME.2017.2676243 (http://zhilinzhang.com/spcup2015/data.html). The evaluation

0018-9294 © 2017 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.
See http://www.ieee.org/publications standards/publications/rights/index.html for more information.

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 TEMKO: ACCURATE HEART RATE MONITORING DURING PHYSICAL EXERCISES USING PPG 2017

Fig. 1. The challenge of HR estimation during physical exercises. Plot (a) shows a spectrogram of a 5 min PPG recording. Plots (b) and (d) show
examples of a PPG signal from the spectrogram in (a). Plot (c) and (e) show the spectral envelopes of the PPGs in (b) and (d), respectively. The
true HR is denoted with a circle. In the presence of MAs, plots (b) and (c), the highest peak of the spectral envelope does not coincide with the true
HR. Best viewed in color.

rules and metrics are defined which facilitates the comparison 4) Benefits of combining different HR estimation algorithms
between different approaches. In the period since the provi- are discussed and an ensemble is designed. Its superior
sion of this dataset, several HR estimation algorithms have been performance is presented for the first time.
designed and tested [2], [3], [21]–[25]. However, the reported 5) The Matlab implementation of the designed HR es-
improvements in performance are usually accompanied with timation algorithms is made available through online
the increased number of free parameters which may be a sign resources.
of overfitting given the fixed size of the dataset on which the This paper is organized as follows. The developed system
algorithms are both designed and tested. is described in detail in Section II. Section III describes the
In this paper a new approach to HR estimation which is based database, metrics and the performance assessment routine used
on Wiener filtering and the phase vocoder is detailed. In contrast in the study. The results on the provided data are presented in
to the previously presented systems, the proposed solution does Section IV and conclusions are drawn in Section V.
not rely on a set of heuristic rules and thresholds and requires
very few parameters to be tuned. The noise signature is estimated II. HEART RATE ESTIMATION ALGORITHM
from accelerometer signals and the Wiener filter is used to atten-
uate the noise components in the PPG signal. The phase vocoder The flowchart of the developed system is shown in Fig. 2.
is exploited to overcome the limited resolution of the discrete The system consists of 4 main blocks–pre-processing, signal
Fourier transform and to refine the initial dominant frequency de-noising to attenuate the influence of MAs, HR estimation,
estimation. A user-adaptive post-processing step is introduced and post-processing. The examples of signal transformations
and additionally an offline version of post-processing which is carried out at each stage are illustrated in Fig. 3.
based on Viterbi decoding is proposed that requires no tunable
parameters. A. Preprocessing
The main contributions of this study are: During the preprocessing stage, the two PPG signals and three
1) A detailed description of the HR estimation algorithm accelerometer signals are filtered with a 4th order Butterworth
(WFPV) including the proposed user-adaptive post- band-pass filter (0.4–4 Hz) as shown in Fig. 3(a). The two PPG
processing signals are then normalized to zero mean and unit variance (z-
2) An offline version that uses Viterbi decoding is designed score normalization) and averaged. The averaged PPG signal
(WFPV + VD) for applications where online HR moni- and the 3 accelerometer signals are down-sampled from 125 to
toring is not required (e.g. for fitness trackers for swim- 25 Hz for further processing. The signals are then subjected to
mers or offline fitness statistics). A significantly higher the Discrete Fourier Transform (DFT) with the number of bins
performance is obtained by trading online monitoring for set to 1024. Fig. 3(b) shows the spectral envelope of the PPG
accuracy. signal for HRs ranging from 60 to 180 BPM. This range of
3) The performance of the most accurate alternatives pub- HRs is chosen based on the specifics of database used in this
lished to date is summarised and thoroughly compared study which will be explained in Section III and can be changed
with both the online WFPV and offline WFPV + VD. accordingly for other corpuses. Overall, the pre-processing stage
Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 2018 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 64, NO. 9, SEPTEMBER 2017

Fig. 2. The flowchart of the developed HR estimation system.

Fig. 3. Signal transformation in the developed WFPV system. Plot (a) shows z-score normalised PPG from two channels, average PPG, and three
accelerometer signals after filtering, (b) shows the spectral envelopes of PPG and noise, measured as average accelerometer signals, and the peak
frequency of the PPG, (c) shows the processed spectral envelope and its maximum after MAs were attenuated with Wiener filtering, and (d) shows
the maximum of the spectral envelop before and after the phase vocoder.

in this work repeats that of other papers that report results on where S(f) and N(f) are the spectra of the clean PPG signal and
this dataset [1], [3], [6]. the MAs, respectively. The estimation of the clean signal can
then be obtained as:
 
B. De-noising Using Wiener Filtering N (f )
S̃ (f ) = X (f )−N (f ) = 1 − X (f ) = W (f ) X (f )
The Wiener filter is a common tool to estimate a desired signal X (f )
by linear time-invariant filtering of an observed noisy process (2)
[15], [31]. Assuming known stationary signal and additive noise For a signal observed in uncorrelated additive random noise,
spectra, the Wiener filter performs the minimum mean square the frequency-domain Wiener filter is given as:
error estimation of the desired signal given another related pro- PX X (f ) − PN N (f ) PS S (f )
cess. Causal Wiener filtering is applied here to estimate the W (f ) = = (3)
PX X (f ) PS S (f ) + PN N (f )
clean PPG signal from the observed PPG signal. The noisy PPG
signal, X(f ), is assumed to be corrupted by additive MA noise: where PS S (f ), PN N (f ) and PX X (f ) are the power spectrums
of the clean signal, noise and observed signal. The filter convolu-
X (f ) = S (f ) + N (f ) (1) tion in time domain is equivalent to multiplication in frequency

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 TEMKO: ACCURATE HEART RATE MONITORING DURING PHYSICAL EXERCISES USING PPG 2019

and thus the Wiener filter acts as an adaptive signal-to-noise de- and equals to 1/8 ∗ 60 = 7.5 BPM. Zero-padding before DFT
pendent attenuator, where frequencies which are more affected is used to interpolate the spectral envelope to other frequen-
by the noise are given less importance. cies thus decreasing the frequency spacing between neighboring
The filter Eq. (3) requires separate estimates of the noise and DFT bins. This does not create new information but allows for
signal power spectrums. The noise spectrum can be directly a better revelation of the existing information in the signal.
estimated from the accelerometer signals which is done by av- The phase vocoder is the technique that is used in audio
eraging the spectrum of the 3 accelerometer signals. The clean processing to manipulate audio length without changing its pitch
PPG spectrum, PS S (f ), can be estimated as a subtraction of the or to change its pitch without affecting its length, by preserving
noise signal from the observed signal, PX X (f ) − PN N (f ), the coherence of phase information. The phase vocoder uses a
or recursively from previous filter outputs. Depending on how polar representation of the DFT and the instantaneous frequency
the power spectrum of the clean PPG signal is estimated, two estimation is computed as a discrete derivative of the phase.
Wiener filters are implemented, with frequency domain filter When analyzing the signal with multiple overlapping windows
coefficients given as: the individual signal components (sinusoids) will be correlated
in time and spread over multiple adjacent DFT frequency bins
PN N (t, k)
w1 (t, k) = 1 − 1
t (4) (spectral leakage). The deviation of the true frequency from
C i=t−C +1 PX X (i,k ) the bin center frequency is encoded in phase changes of two
1
t−1 consecutive frames, so that the instantaneous frequency can be
i=t−C w2 (i, k) PX X (i, k)
w2 (t, k) = 1
C
t−1 (5) given as:
C i=t−C w2 (i, k) PX X (i, k) + PN N (t,k )
1 dθ (t)
where w(t,k) is the weight of the k-th frequency bin at time, t. f (t) = (6)
2π dt
The power spectrums of noise and PPG in Eq. (5) and Eq. (6)
The DFT phases, θ2 , θ1 , from the current and previous frames,
are normalized by their maximum values to be commensurable.
of the chosen frequency peak in the magnitude spectrum, f , are
It can be seen that in both equations the power spectrum of the
used to refine the initial frequency estimation:
observed signal is averaged over the past C spectral envelopes
(C = 15, in this work). In Eq. (4), the power spectrum of the   (θ2 − θ1 + 2πn)
arg min f˜ (n) − f ; f˜ (n) = , ∀n ∈ N
clean PPG signal is estimated by subtracting the observed noise n (2π (t2 − t1 ))
from the observed PPG signal. If C = 1, then the Wiener filter (7)
in Eq. (4) performs a simple version of spectral subtraction where n is a positive integer, t2 , t1 are the time stamps of
[16], [17]. In Eq. (5), the spectrum of the clean PPG signal is the two frames, here t2 − t1 = 2 s which is the window shift.
computed recursively by averaging the previous filtered signal The series, f˜(n), is computed for several n using Eq. (7), and
outputs. the value of f˜ which is closest to the initial frequency estima-
The spectral envelop of the cleaned PPG signal is then ob- tion, f , is chosen. As a result, the previous dominant frequency
tained by multiplying the spectral envelope of the observed value is refined to the new value, f ← f˜. This is illustrated in
signal, PX X (t, k), with the calculated filter coefficients, w(t, k). Fig. 3(d), where the estimated DFT HR of 161.3 BPM was
It can be seen that in the current implementation the Wiener refined to 158.2 BPM, with the true HR being of 159.1 BPM.
filter requires only one parameters to be specified, C. It can be seen that the phase vocoder technique requires no
The spectral envelopes processed with the two designed filters parameters to set.
are scaled by their standard deviation because unlike w2 , w1
can have negative values, when the scaled power of observed D. Post-processing
noise is larger than the scaled power of the observed signal
1) Online Post-processing With Heuristic Rules
for certain frequencies. The resultant signals are averaged to
and Thresholds: The post-processing steps include history
give a final spectral envelope of the cleaned PPG signal. The
tracking and smoothing. The history of the past HR estimation is
dominant frequency (the frequency with the highest magnitude)
preserved and used to guide the search range for the maximum
is converted to the HR estimate in BPM as shown in Fig. 3(c).
DFT magnitude in the current frame. For instance, if the past
HR estimation was 125 BPM, then the current HR is expected
C. HR Estimation and Refinement Using Phase Vocoder to be within a certain range around 125 BPM. Here it is set to
In this work, the phase vocoder technique [11]–[13], [32] is the maximum absolute HR difference between consecutive HR
employed to refine the initial HR estimate through the estimation estimates, f (t), observed so far for this user:
of the instantaneous frequency as the rate of change of phase
τi = max{|f (t) − f (t − 1)| : k < t < i − 1} (8)
angle at time [33]. For signals that are not truly sinusoidal or for
nonstationary signals one needs to account for the time-varying The search range, ± τi , t ≤ k is initialized to be wide enough
nature of the process and that can be done with estimation of (±25 BPM) for the first 30 s–1 min of each recording, and then
the instantaneous frequency. adapts to the specifics of the user’s physiology. This eliminates
The effective frequency resolution (the minimum frequency the need to tune another threshold in the post-processing.
that can be estimated, the Rayleigh frequency) of the data is For the final smoothing, the weighted average between the
limited by the size of the window of the analyzed data (8s) current estimate and its prediction using linear regression is

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 2020 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 64, NO. 9, SEPTEMBER 2017

computed if the difference between the current and previous TABLE I

DATABASE OF 23 PPG RECORDINGS FROM IEEE SP CUP
HR estimates is above 5 BPM:
f˜ = α f + (1 − α) fL in R eg (9)
where fL in R eg is a prediction given by a regression line which
is fitted over the past 6 HR estimates using a least-squares error,
and α = 0.8.

E. Offline Post-processing With Viterbi Decoding

The offline version of WFPV substitutes the original post-
processing steps with a probabilistic framework using Viterbi
decoding [28]. The time-frequency plane (spectrogram) of a
complete recording which is composed of DFT magnitudes after
Wiener filtering is considered as a N-by-T state-space map of
emission probabilities, B, for N states (discrete values of HR)
and T observations (time windows), where bj t is a magnitude
value of the jth DFT bin for the tth time window. The N-
by-N matrix of transition probabilities, A, where aij represents
the probability of changing from the ith HR to the jth HR, is
estimated from the ground truth automatically by counting the
transitions using the leave-one-recording-out procedure. In this
manner, the ground truth of the testing recording is never used
but the ground truths of all other recordings are used to estimate
the transition probability matrix. The process of estimation of
the transition probability matrix is fully automatic, data-driven
and requires no user-tunable parameters.
The Viterbi algorithm is then applied to recursively estimate
the most likely path (the path with the highest cumulative prob- two PPG sensors with green LEDs (wavelength: 515 nm). The
ability) through the time dimension, t, using emission and tran- acceleration signal was also recorded from the wrist using a
sition probability matrices, B, A: three-axis accelerometer. Both the PPG sensors and the ac-
Ψt (j) = argmax [δt−1 (i) aij ] , (10) celerometer were embedded in a comfortable wristband. The
1≤i≤N ECG signal was recorded simultaneously from the chest using
where δt (j) = max1≤i≤N [δt−1 (i)aij ]bj t , 2 ≤ t ≤ T, 1 ≤ j ≤ wet ECG sensors. All signals were sampled at 125 Hz and sent
N , δ1 (i) = πi bi1 , 1 ≤ i ≤ N , and πi are the prior probabili- to a nearby computer via Bluetooth.
ties. After the recursion is computed the state sequence is back- Three types of activities were performed. Type 1 (T1) activity
tracked as: involved walking or running on a treadmill for intervals 0.5–1–
1–1–1–0.5 min with the speed of 1–2, 6–8, 12–15, 6–8, 12–15,
iT = arg max [δT (i)] 1–2 km/h, respectively. The subjects were asked to use the hand
1≤i≤N
with the wristband to pull clothes, wipe sweat on forehead, and
it = Ψt+1 (it+1 ) , t = T − 1, T − 2, . . . , 1 (11) push buttons on the treadmill. Type 2 activity included various
The state sequence is converted to HR estimates which are forearm and upper arm exercise which are common in arm
then smoothed with a non-causal (central) moving average filter. rehabilitation (e.g. shake hands, stretch, push, running, jump,
The Viterbi decoding effectively performs the post-processing and push-ups). Type 3 activity consisted of intensive forearm
in a threshold-free probabilistic manner. and upper arm movements (e.g. boxing).
The Matlab implementation of the online and offline WFPV The synchronously recorded ECG was used to provide the
HR estimators along with the main results are available online ground truth HR in BPM as described in detail in [2] and [3].
(https://github.com/andtem2000/PPG). The ECG-based HRs were calculated for every 8 s window with
a 2 s shift. The same window length and shift are suggested
III. DATABASE AND METRICS for HR estimation from PPG to have the same number of HR
estimates and true HRs. The ground truth HRs were then used
A. Database to assess the performance of the developed algorithms.
The data used were provided for the IEEE Signal Processing
Cup 2015 and are detailed in Table I. The dataset consists of
B. Metrics
23 5-min recordings which were collected from 18 to 58 years
old subjects performing various physical exercises. For each The metrics which are usually computed in other studies [3]
subject, the PPG signals were recorded from the wrist using are reported here. The Absolute Error (AE) is used to evaluate

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 TEMKO: ACCURATE HEART RATE MONITORING DURING PHYSICAL EXERCISES USING PPG 2021

TABLE II agreement area) occur more frequently in the low HR region,

PERFORMANCE OF WFPV
centered at ∼ 80 BPM. At the same time, the most frequent
HR according to Fig. 4(c) is much higher, ∼ 140 BPM, but the
B B & SS B & WF1 B & WF2 B & WF1& WF2 B & WF&PV
algorithm makes fewer large errors in that region. This indicates
All avAE 13.11 5.08 4.19 3.30 2.21 1.97 the range of HR which are mostly affected by MA. Fig. 5 shows
sdAE 10.13 7.13 5.39 4.70 2.61 2.48 an example of the true and estimated HR for recording 9–the
recording with the best performance achieved.
Table III details the HR scores for each of the 23 recordings in
the accuracy of each HR estimate: the database, for avAE and avRE metrics, and also summarizes
the performance for a group of recordings. It can be seen that the
AEi = |fest (i) − ftr u e (i)| (12) error in HR estimation heavily depends on the type of physical
activity. Running on a treadmill results in an avAE of 1.02 BPM,
where fest (i) and ftr u e (i) denote the estimated and the true
whereas arm exercises and intensive arm exercises result in an
HR value in the i-th time window in BPM, respectively.
avAE of 3.01 BPM, respectively.
The following three metrics are used to evaluate the performance
of the developed algorithm, Average Absolute Error (avAE),
Standard Deviation of the Absolute Error (sdAE) and Average B. Comparison With Existing Alternatives
Relative Error (avRA):
Table III also provides the performance of alternative HR es-
1 
N timation algorithms tested on the same dataset or a part of it.
avAE = AEi (13) Apart from TROIKA [2] and JOSS [3] the results are reported
N i=1
for several methods that have been recently published in journals

1 N 2 such as: a time-varying spectral filtering algorithm for recon-
sdAE = (AEi − avAE) (14) struction of motion artifact (SpaMa, [21]), an algorithm based on
N i=1
ensemble empirical mode decomposition (EEMD, [23]), a HR
1  AEi
N
estimation algorithm based on asymmetric least squares spec-
avRE = (15) trum subtraction and Bayesian decision theory (Spectrap, [24]),
N i=1 ftr u e(i)
an iterative method with adaptive thresholds (IMAT, [25]), and
where N is the total number of estimates (number of windows). a multi-channel spectral matrix decomposition method [MC-
These metrics are reported for each of the 23 recordings. SMD, 27].
The first 12 (T1 activity) of the 23 recordings of the dataset
IV. RESULTS AND DISCUSSION were used in [2] where the scores of 2.34, 1.82, and 2.47 BPM
were reported for avAE, avRE, and sdAE, respectively. Subse-
A. Performance Assessment
quently, on the same dataset the avAE and avRE were reduced
The performance of the proposed HR estimation system is to 1.28 and 1.01 BPM, with an increased sdAE of 2.61 BPM in
summarized in Table II. On the database of 23 recordings, [3]. Evaluating on the same 12 recordings, the developed sys-
WFPV results in an avAE of 1.97 BPM with sdAE of 2.48 BPM. tem obtains scores of 1.02, 0.81 and 1.25 BPM for avAE, avRE,
Table II also shows the contribution of each system constituent and sdAE, respectively, which corresponds to a relative error
towards the final performance. The Baseline denotes the HR reduction of 20% and 52% for avAE and sdAE. It can be seen
estimator as shown in Fig. 2, with all the preprocessing (fil- that the error was significantly reduced for the worst performing
tering, normalization), DFT, and the post-processing (history recordings in the dataset which is reflected in the lower sdAE.
tracking and smoothing), as described in Section III, but with- The correlation coefficient was 0.997, as compared to 0.993
out Wiener filtering and the phase vocoder. The baseline is the in [3]. On the first 12 PPG recordings the WFPV algorithm
system that takes the maximum DFT magnitude to determine outperforms all of the listed approaches with the exception of
the HR and has no routine to compensate for the presence of SpaMa [21].
MA. The Baseline performance reaches an avAE of 13.11 BPM. On the last 10 PPG recordings (rec. 14–23), WFPV also
If a simple spectral subtraction is added (Eq. (4), C = 1), then outperforms all the listed approaches including SpaMa [21]
the performance improves to 5.08 BPM (B&SS, Table II). Us- with the exception of Spectrap [24] which employs non-causal
ing the Wiener filter as described in Eq. (4) (WF1) improves the smoothing and cannot be seen as an online algorithm. Because
performance to 4.19 BPM. Using the recursive implementation most studies do not report results for rec. 13, the majority of
of the Wiener filter (WF2) reduces the error from 5.08 to 3.30 approaches can be compared using the average of the results for
BPM, and the combination of both filter, WF1 and WF2, results rec. 1–12 and 14–23. With an avAE of 1.90 BPM, the proposed
in an error of 2.21 BPM. A refinement introduced by the phase approach outperforms all the listed alternatives.
vocoder reduces the error to 1.97 BPM. A similar reduction in Looking at the performance of the offline version (WFPV
sdAE is observed. + VD), it can be seen that the incorporation of Viterbi de-
Fig. 4 shows the correlation (a), Bland-Altman plot (b) and coding and non-causal smoothing has a significant effect on
the distribution of HR in the database. Interestingly, Fig. 4(b) the performance, reducing the avAE from 1.97 to 1.37 BPM,
indicates that the largest errors (the points outside the limit of outperforming all the listed methods including the offline

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 2022 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 64, NO. 9, SEPTEMBER 2017

Fig. 4. (a) Pearson correlation between the estimated HR and the ground truth HR, (b) Bland-Altman plot, and (c) the distribution of the HR in the
DB. Correlation coefficient of the WFPV algorithm is 0.9908.

on the whole dataset of all 23 recordings. It is worth noting

that since the ground truth for the whole dataset is available,
there is no difference between the training and testing data from
the performance assessment perspective. In this context, the
number of the free parameters of the developed system is the
only indicator of the potential overfit. It can also be seen from
Table III that although the whole database of 23 recordings is
available since early 2015, several very recent works [23], [25],
[27] still prefer to report results on the ‘easier’ part of the dataset,
T1 activity only.
The method presented in this study requires only a few pa-
rameters to be tuned. The number of frequency bins in the DFT
was set to 1024 (with zero-padding). Likewise, C = 15 was
used in the Wiener filters. The range of the search space for
the maximum in DFT magnitudes for the next frame is deter-
Fig. 5. The true and estimated heart rate for recording 9. mined adaptively using Eq. (8). Viterbi decoding for the offline
algorithm does not require any thresholds to tune. The proposed
algorithm takes under 10s to process the whole PPG dataset of
Spectrap [24]. This indicates that the performance comparison 23 recordings (Matlab R2013b @ Intel Core E7200 2.5 GHz)
is only sensible among the algorithms that perform either offline which compares favorably with techniques reported in [2], [3],
or online processing. [22], [23], and [25]. It is reported in [23] that in order to es-
timate the HR for the first 12 recordings TROIKA [2] takes
C. Analysis several hours, JOSS [3] takes 300 s, EEMD [23] takes 200 s,
The approaches reported in [2], [3], [23]–[25], and [27] rely and IMAT [25] takes several hours as tested in this study using
on a number of heuristic rules and thresholds (see Table IV). the provided implementation.
For instance, the post-processing step alone in [2], [3], [23], and Reproducibility is an essential principle of the scientific pro-
[27] requires a dozen parameters to be specified. The relevance cess. The availability of the data and the algorithm code will
of the knowledge extracted from the data (in terms of the rules encourage more researchers to explore the area of PPG signal
and threshold values) is tested on the same data and the best processing and build novel technical solutions. Unfortunately,
achievable results are usually reported. Increasing the number among the solutions discussed in Table III only the implemen-
of degrees of freedom of the designed system will inevitably tation of the IMAT algorithm [25] is available online. Repro-
lead to an improved performance but it comes at the cost of ducibility of results will allow for verification of genuine tech-
an increased risk of poor generalization on the unseen data, nical contributions.
especially if fine-tuning is performed in the post-processing to
correct the errors of the core algorithm.
As can be seen from Table III, some studies report results D. Combination With Other Approaches
only on the first 12 recordings (T1 activity); others in addition It is well known that a better performance can be obtained
report performance on the data that were used as test data in by blending complementary approaches [34]. Table V shows
SP Cup (rec. 14–23), and only a few report the performance the performance of the system that ensembles the TROIKA [2],

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 TEMKO: ACCURATE HEART RATE MONITORING DURING PHYSICAL EXERCISES USING PPG 2023

TABLE III
COMPARISON OF THE HR ESTIMATION ERRORS ON 23 PPG RECORDINGS OF VARIOUS ALGORITHMS PROPOSED IN THE LITERATURE

Rec Activity TROIKA∗ [2] JOSS∗ [3] SpaMa [21] EEMD [23] Spectrap [24] (Offline) IMAT [25] MC-SMD [27] WFPV This study WFPV+VD This study (Offline)

1 T1 2.29 | 2.18 1.33 | 1.19 1.23 | 1.14 1.70 | - 1.18 | 1.04 1.72 | 1.50 1.16 | 0.91 1.25 | 1.15 0.93 | 0.89
2 T1 2.19 | 2.37 1.75 | 1.66 1.59 | 1.30 0.84 | - 2.42 | 2.33 1.33 | 1.30 1.07 | 0.87 1.41 | 1.30 0.82 | 0.73
3 T1 2.00|1.50 1.47 | 1.27 0.57 | 0.45 0.56 | - 0.86 | 0.66 0.90 | 0.75 0.80 | 0.62 0.71 | 0.59 0.64 | 0.54
4 T1 2.15 | 2.00 1.48 | 1.41 0.44 | 0.31 1.15 | - 1.38 | 1.31 1.28 | 1.20 1.13 | 0.84 0.97 | 0.88 0.83 | 0.80
5 T1 2.01 | 1.22 0.69 | 0.51 0.47 | 0.31 0.77 | - 0.92 | 0.74 0.93 | 0.69 0.98 | 0.68 0.75 | 0.57 0.50 | 0.38
6 T1 2.76 | 2.51 1.32 | 1.09 0.61 | 0.45 1.06 | - 1.37 | 1.14 1.41 | 1.20 1.29 | 0.96 0.92 | 0.75 0.78 | 0.61
7 T1 1.67 | 1.27 0.71 | 0.54 0.54 | 0.40 0.63 | - 1.53 | 1.36 0.61 | 0.50 0.88 | 0.65 0.65 | 0.50 0.50 | 0.40
8 T1 1.93 | 1.47 0.56 | 0.47 0.40 | 0.33 0.53 | - 0.64 | 0.55 0.88 | 0.80 0.81 | 0.64 0.97 | 0.83 0.67 | 0.56
9 T1 1.86 | 1.28 0.49 | 0.41 0.40 | 0.42 0.52 | - 0.60 | 0.52 0.59 | 0.50 0.55 | 0.43 0.55 | 0.48 0.45 | 0.38
10 T1 4.70 | 2.49 3.81 | 2.43 2.63 |1.59 2.56 | - 3.65 | 2.27 3.78 | 2.40 3.18 | 1.95 2.06 | 1.29 1.43 | 0.90
11 T1 1.72 | 1.29 0.78 | 0.51 0.64 | 0.42 1.05 | - 0.92 | 0.65 0.85 | 0.60 0.79 | 0.51 1.03 | 0.68 0.74 | 0.48
12 T1 2.84 | 2.30 1.04 | 0.81 1.20 | 0.86 0.91 | - 1.25 | 1.02 0.71 | 0.50 0.72 | 0.53 0.99 | 0.70 0.75 | 0.53
13 T2 - - 3.41 | 4.25 - - - - 3.54 | 4.08 2.77 | 3.19
14 T2 6.63 | 8.76 8.07 | 10.9 7.29 | 9.80 - 4.89 | 6.29 - 9.59 | 12.2 8.68 | 10.9
15 T2 1.94 | 2.56 1.61 | 2.01 2.73 | 2.21 - 1.58 | 1.98 - 2.57 | 3.16 1.99 | 2.43
16 T3 1.35 | 1.04 3.10 | 2.69 3.18 | 2.11 - 1.83 | 1.49 - 2.25 | 1.87 1.83 | 1.51
17 T3 7.82 | 4.88 7.01 | 4.49 3.01 | 2.52 - 3.05 | 2.00 - 3.01 | 1.99 2.22 | 1.49
18 T3 2.46 | 2.00 2.99 | 2.52 4.46 | 3.23 - 1.62 | 1.36 - 2.73 | 2.29 2.01 | 1.70
19 T3 1.73 | 1.27 1.67 | 1.23 3.58 | 3.98 - 1.24 | 0.92 - 1.57 | 1.15 1.23 | 0.90
20 T2 3.33 | 3.90 2.80 | 3.46 1.94 | 1.66 - 2.04 | 2.23 - 2.10 | 2.41 1.53 | 1.78
21 T3 3.41 | 2.43 1.88 | 1.32 2.56 | 2.02 - 2.49 | 1.81 - 3.44 | 2.45 2.74 | 1.96
22 T3 2.69 | 2.12 0.92 | 0.74 3.12 | 3.28 - 1.16 | 0.92 - 1.61 | 1.26 1.02 | 0.80
23 T2 0.51 | 0.59 0.49 | 0.57 1.72 | 1.97 - 0.66 | 0.79 - 0.75 | 0.88 0.51 | 0.59
Mean T1 avAE 2.34 1.28 0.89 1.02 1.50 1.25 1.11 1.02 0.65
Rec avRE 1.82 1.01 0.65 - 1.12 0.99 0.80 0.81 0.55
1-12 sdAE 2.47 2.61 - 1.79 1.95 - 1.99 1.25 1.00
T2-T3 avAE - - 3.36 - - - - 3.01 2.16
Rec avRE - - 3.33 - - - - 3.06 2.21
13-23 sdAE - - - - - - - 3.83 2.89
Test avAE 3.19 3.05 3.35 - 2.13 - - 2.95 2.11
Rec avRE 2.95 3.00 3.27 - 2.77 - - 2.96 2.12
14-23 sdAE 3.61 3.35 - - 2.04 - - 3.71 2.82
Rec avAE 2.73 2.08 2.01 - 1.79 - - 1.90 1.31
1-12, avRE 2.33 1.91 1.84 - 1.87 - - 1.98 1.26
14-23 sdAE 2.99 2.79 - - 1.99 - - 2.37 1.83
All avAE - - 2.07 - - - - 1.97 1.37
Rec avRE - - 1.95 - - - - 1.89 1.34
1-23 sdAE - - - - - - - 2.48 1.91

∗
The HRs generated by TROIKA and JOSS on recordings 14-23 are obtained from https://sites.google.com/site/researchbyzhang/publications.

TABLE IV
THE NUMBER OF USER-TUNABLE PARAMETERS

# tunable threshhold TROIKA [2] JOSS [3] SpaMa [21] EEMD [23] Spectrap [24] IMAT [25] MC-SMD [27] WFPV This study WFPV + VD This study

De-noising& HR detection∗ > 10 5 6 > 10 6 > 10 > 10 2 2

Post-processing > 10 > 10 5 > 10 4 > 10 > 10 4 1

∗
The number of parameters does not include the preprocessing parameters such as filter length, cut-off points, down-sampling, etc.

TABLE V in all reported metrics. In fact, Bland-Altman plots that show the
PERFORMANCE OF THE ENSEMBLE OF THE THREE HR ESTIMATION
ALGORITHMS ON RECORDINGS 14–23
distribution of errors for a given algorithm (as shown in Fig. 4(b)
for WFPV) can be used to assess the level of complementarity
of various approaches—for instance, algorithms that produce
TROIKA [2] JOSS [3] WFPV Ensemble
most errors in the region of high HR would be good candidates
Rec.14-23 avAE 3.19 3.05 2.95 2.32 to form ensemble with WFPV. For this purpose, the availability
avRE 2.95 3.00 2.96 2.29
sdAE 3.61 3.35 3.71 2.37
of the algorithms implementation for building more accurate
solutions is essential.

V. CONCLUSIONS
JOSS [3] and WFPV methods by taking a simple average of
their HR estimates. It can be seen that even a late decision-level An algorithm based on the Wiener filtering and the phase
combination of the HR estimates significantly reduces the error vocoder is proposed. It provides a simple but effective solution

Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.
 2024 IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 64, NO. 9, SEPTEMBER 2017

to the PPG-based HR estimation and can serve as a baseline [20] A. Temko, “Estimation of heart rate from photoplethysmography during
performance for further studies. The algorithm has low compu- physical exercise using wiener filtering and the phase vocoder,” in Proc.
Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Aug. 2015, pp. 1500–1503.
tational cost and is well suited for fitness tracking and health [21] S. Salehizadeh et al., “A novel time-varying spectral filtering algorithm
monitoring in wearable devices [18], [19]. An offline version of for reconstruction of motion artifact corrupted heart rate signals during
the algorithm with a higher accuracy is proposed that utilizes intense physical activities using a wearable photoplethysmogram sensor,”
Sensors, vol. 16, no. 1, 2016.
Viterbi decoding to post-process HR estimates. The Matlab im- [22] H. Pan et al., “HeartBEAT: Heart beat estimation through adaptive
plementation for both algorithms is provided. tracking,” in Proc. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., 2016,
Future research may concentrate on the usage of other spectral pp. 587–590.
[23] E. Khan et al., “A robust heart rate monitoring scheme using photo-
estimation methods. The Wiener filter can be applied in the time plethysmographic signals corrupted by intense motion artifacts,” IEEE
domain to facilitate this option. Additionally, prior information Trans. Biomed. Eng., vol. 63, no. 3, pp. 550–562, Mar. 2016.
available about the user, such as the user weight or age may [24] B. Sun and Z. Zhang, “Photoplethysmography-based heart rate monitor-
ing using asymmetric least squares spectrum subtraction and Bayesian
be incorporated to estimate the maximum HR expected. The decision theory,” IEEE Sens. J., vol. 15, no. 12, pp. 7161–7168,
probabilistic framework with Viterbi decoding can be adapted Dec. 2015.
to online monitoring with a predefined delay. Finally, electric [25] M. Mashhadi et al., “Heart rate tracking using wrist-type photoplethys-
mographic (PPG) signals during physical exercise with simultaneous ac-
potential wearable sensors can be used to complement the PPG celerometry,” IEEE Signal Process. Lett., vol. 23, no. 2, pp. 227–231,
ones [35]. Feb. 2016.
[26] C. Garber et al., “Quantity and quality of exercise for developing and
REFERENCES maintaining cardiorespiratory, musculoskeletal, and neuromotor fitness in
apparently healthy adults: Guidance for prescribing exercise,” Med. Sci.
[1] J. Allen, “Photoplethysmography and its application in clinical physio- Sports Exercise, vol. 43, pp. 1334–1359, 2011.
logical measurement,” Phys. Meas., vol. 28, pp. 1–39, 2007. [27] J. Xiong et al., “Spectral matrix decomposition-based motion artifacts
[2] Z. Zhang et al., “TROIKA: A general framework for heart rate monitoring removal in multi-channel PPG sensor signals,” IEEE Access, vol. 4,
using wrist-type photoplethysmographic signals during intensive physi- pp. 3076–3086, Jun. 2016.
cal exercise,” IEEE Trans. Biomed. Eng., vol. 62, no. 2, pp. 522–531, [28] L. Rabiner, “A tutorial on hidden Markov models and selected applications
Feb. 2015. in speech recognition,” Proc. IEEE, vol. 77, no. 2, pp. 257–286, Feb. 1989.
[3] Z. Zhang, “Photoplethysmography-based heart rate monitoring in physical [29] A. Challoner, “Photoelectric plethysmography for estimating cutaneous
activities via joint sparse spectrum reconstruction,” IEEE Trans. Biomed. blood flow,” in Non-invasive Physiological Measurement, vol. 1, P. Rolfe,
Eng., vol. 62, no. 8, pp. 1902–1910, Aug. 2015. Ed. Oxford, U.K.: Academic, 1979, pp. 127–151.
[4] R. Yousefi et al., “A motion-tolerant adaptive algorithm for wearable [30] A. Kamal et al., “Skin photoplethysmography—A review,” Comput. Meth-
photoplethysmographic biosensors,” IEEE J. Biomed. Health, vol. 18, ods Programs Biomed., vol. 28, pp. 257–269, 1989.
no. 2, pp. 670–681, Mar. 2014. [31] S. Vaseghi, Advanced Digital Signal Processing and Noise Reduction.
[5] M. Ram et al., “A novel approach for motion artifact reduction in PPG New York, NY, USA: Wiley, 2000, ch. 6, pp. 178–203.
signals based on AS-LMS adaptive filter,” IEEE Trans. Instrum. Meas., [32] M. Puckette and J. Brown, “Accuracy of frequency estimate using the
vol. 61, no. 5, pp. 1445–1457, May 2012. phase vocoder,” IEEE Trans. Speech Audio Process., vol. 6, no. 2,
[6] B. Kim and S. Yoo, “Motion artifact reduction in photoplethysmogra- pp. 166–176, Mar. 1998.
phy using independent component analysis,” IEEE Trans. Biomed. Eng., [33] B. Boashash, “Estimating and interpreting the instantaneous frequency
vol. 53, no. 3, pp. 566–568, Mar. 2006. of a signal. I. Fundamentals,” Proc. IEEE, vol. 80, no. 4, pp. 520–538,
[7] X. Sun et al., “Robust heart beat detection from photoplethysmography Apr. 1992.
interlaced with motion artifacts based on empirical mode decomposi- [34] E. Thomas et al., “Discriminative and generative classification techniques
tion,” in Proc. IEEE-EMBS Int. Conf. Biomed. Health Informat., 2012, applied to automated neonatal seizure detection,” IEEE J. Biomed. Health
pp. 775–778. Informat., vol. 17, no. 2, pp. 297–304, Mar. 2013.
[8] H. Fukushima et al., “Estimating heart rate using wrist-type photoplethys- [35] M. Magno et al., “A low power wireless node for contact and contactless
mography and acceleration sensor while running,” in Proc. Annu. Int. heart monitoring,” Microelectronics, vol. 45, no. 12, pp. 1656–1664, 2014.
Conf. IEEE Eng. Med. Biol. Soc., 2012, pp. 2901–2904.
[9] B. Lee et al., “Improved elimination of motion artifacts from a pho-
toplethysmographic signal using a Kalman smoother with simultaneous
accelerometry,” Physiol. Meas., vol. 31, no. 12, pp. 1585–1603, 2010. Andriy Temko (M’05–SM’12) received the Engi-
[10] S. Cotter et al., “Sparse solutions to linear inverse problems with mul- neering degree in computer science from Dnipro
tiple measurement vectors,” IEEE Trans Signal Process., vol. 53, no. 7, National University, Dnipro, Ukraine, in 2002,
pp. 2477–2488, Jul. 2005. and the Ph.D. degree in telecommunication
[11] J. Flanagan and R. Golden, “Phase vocoder,” Bell Syst. Tech. J., vol. 45, from the Universitat Politècnica de Catalunya,
pp. 1493–1509, 1966. Barcelona, Spain, in 2008. During 2006–2007,
[12] T. Dutoit and F. Marquès, Applied Signal Processing—A MATLAB-Based he was a Task Leader in detection and classifica-
Proof of Concept. New York, NY, USA: Springer, 2009. tion of acoustic events within the EU-funded in-
[13] A. Temko et al., “Clinical implementation of a neonatal seizure detection ternational evaluation campaigns on detection of
algorithm,” Decision Support Syst., vol. 70, pp. 86–96, 2015. events, activities, and their relationships. Since
[14] Z. Zhang, “Heart rate monitoring from wrist-type photoplethysmographic late 2008, he has been with the Irish Centre for
(PPG) signals during intensive physical exercise,” in Proc. IEEE Global Fetal and Neonatal Translational Research University College Cork,
Conf. Signal Inf. Process., 2014, pp. 866–870. Ireland, working on algorithms for the detection of brain injuries
[15] R. Brown and P. Hwang, Introduction to Random Signals and Applied in the newborn. He is the author of more than 70 peer-reviewed
Kalman Filtering, 3 ed. New York, NY, USA: Wiley, 1996. publications and two patents. He developed and patented a novel
[16] S. Boll, “Suppression of acoustic noise in speech using spectral subtrac- cutting-edge neonatal seizure detection system which is currently
tion,” IEEE Trans. Acoust., Speech, Signal Process, vol. ASSP-27, no. 2, undergoing European multi-center clinical trial toward its regulatory
pp. 113–120, Apr. 1979. approval and clinical adoption. In 2016, he was named as win-
[17] A. Temko et al., “Robust neonatal EEG seizure detection through adap- ner of the Kaggle challenge, and a $10,000 first prize, for his
tive background modelling,” Int. J. Neural Syst., vol. 23, no. 4, 2013, work in predicting seizures in the human brain through long-term
Art. no. 1350018. EEG recordings, organized by National Institute of Health, Ameri-
[18] M. Magno et al., “Wearable low power dry surface wireless sensor node for can Epilepsy Society and Melbourne University. His research inter-
healthcare monitoring application,” in Proc. IEEE 9th Int. Conf. Wireless ests include acoustic and physiological signal processing, clinical de-
Mobile Comput., 2013, pp. 200–206. cision support tools, and applications of machine learning for signal
[19] A. Temko et al., “REACT: Real-time EEG analysis for event detection,” in processing.
Proc. AMA-IEEE Med. Technol. Conf. Individualized Healthcare, 2010.
Authorized licensed use limited to: Chulalongkorn University provided by UniNet. Downloaded on October 21,2024 at 04:58:14 UTC from IEEE Xplore. Restrictions apply.