Motion Artifact Reduction In Photoplethysmography
`For Reliable Signal Selection
`
`Runyu Mao1, Mackenzie Tweardy2, Stephan W. Wegerich2,
`Craig J. Goergen3, George R. Wodicka3 and Fengqing Zhu1
`
`Fig. 1: Comparison between PPG collected from a subject
`doing Activities of Daily Living (ADL) and the same subject
`in rest state.
`
`the vital information has been precisely recorded in these
`corrupted PPG signals. On the other hand, PPG collected
`from the subject at rest is relatively clean and stable. Due
`to this reason, many existing PPG datasets [4], [5] collect
`reliable and clean PPG signals in lab settings, where subjects
`are in static states. Synthesized noises are then added to the
`clean PPG signal to simulate the noisy PPG signals [6], [7],
`[8], [9], [10]. Nevertheless, both signal and noise are very
`different from actual PPG signals collected during ADL.
`PPG signals can also be collected where subjects are
`instructed to perform certain high-intensity exercises, and
`various methods were proposed to analyze these PPG signals,
`e.g., strong MA removal and heart rate estimation [11], [12],
`[13], [14]. However, similar to the static PPG signal, these
`data collections were conducted in controlled settings. Thus,
`they struggle to accurately capture the noise due to ADL or
`non-period motions.
`In naturalistic settings, the quality of PPG collected during
`ADL is unpredictable. Some PPG segments may be highly
`corrupted or may not contain useful vital information. Thus,
`it is crucial to select reliable PPG signals collected from
`ADL for meaningful analysis. To estimate the quality of
`PPG signals, Signal Quality Index (SQI), a binary sequence
`to indicate the quality of the corresponding PPG interval,
`has been developed based on waveform features and used
`in some studies [15], [16], [17]. Instead of focusing on
`the morphology of the signal, accuracy of the derived heart
`
`Abstract— Photoplethysmography (PPG) is a non-invasive
`and economical technique to extract vital signs of the human
`body. Although it has been widely used in consumer and
`research grade wrist devices to track a user’s physiology, the
`PPG signal
`is very sensitive to motion which can corrupt
`the signal’s quality. Existing Motion Artifact (MA) reduction
`techniques have been developed and evaluated using either
`synthetic noisy signals or signals collected during high-intensity
`activities - both of which are difficult to generalize for real-
`life scenarios. Therefore, it is valuable to collect realistic PPG
`signals while performing Activities of Daily Living (ADL) to
`develop practical signal denoising and analysis methods. In this
`work, we propose an automatic pseudo clean PPG generation
`process for reliable PPG signal selection. For each noisy PPG
`segment, the corresponding pseudo clean PPG reduces the MAs
`and contains rich temporal details depicting cardiac features.
`Our experimental results show that 71% of the pseudo clean
`PPG collected from ADL can be considered as high quality
`segment where the derived MAE of heart rate and respiration
`rate are 1.46 BPM and 3.93 BrPM, respectively. Therefore, our
`proposed method can determine the reliability of the raw noisy
`PPG by considering quality of the corresponding pseudo clean
`PPG signal.
`
`I. INTRODUCTION
`Photoplethysmogram (PPG) is a non-invasive signal mea-
`sured by transmitting light through the skin and measuring
`the reflected absorption of that light from the capillaries to
`assess human physiological information [1]. Both heartbeat
`and respiratory activities are recorded in PPG which can be
`used to derive physiological features, including vital signs
`of heart rate and respiration rate [2]. With the increasing
`popularity of wearable devices,
`the use of a wrist-worn
`sensor can provide a low-cost, low-burden mechanism for
`collecting PPG signals for personal health monitoring.
`However, PPG signals are sensitive to various Motion
`Artifacts (MA). For example, PPG collected by a wrist-based
`sensor may be highly corrupted due to hand motion, walking,
`etc [3]. As shown in Figure 1, PPG collected from wrist
`devices during Activities of Daily Living (ADL) is corrupted
`by different types of noise. Both high frequency and low
`frequency noise occur non-periodically which contaminates
`vital features in PPG. It is difficult to determine whether
`
`1Runyu Mao and Fengqing Zhu are with the School of Electrical and
`Computer Engineering, Purdue University, West Lafayette, IN 47907, USA
`{mao111,zhu0}@purdue.edu
`2Mackenzie
`Tweardy
`and
`Stephan
`W.
`Wegerich
`are
`with
`PhysIQ,
`Chicago,
`IL
`60606,
`USA
`{mackenzie.tweardy,stephan.wegerich}@physiq.com
`3Craig J. Goergen and George R. Wodicka are with the Weldon School
`of Biomedical Engineering, Purdue University, West Lafayette, IN 47907,
`USA {cgoergen,wodicka}@purdue.edu
`
`arXiv:2109.02755v1 [eess.SP] 6 Sep 2021
`
`AliveCor Ex. 2003 - Page 1
`
`

`

`rate from the PPG segment is proposed to determine the
`PPG’s reliability [18]. Although heart rate is one of the most
`essential vital signs for cardiac activity monitoring, crucial
`temporal information, e.g., heart rate variability, cannot be
`revealed. In addition, for different health applications, the
`quality of the PPG signal requirement may be different.
`For example, highly corrupted PPG signal may provide
`precise heart rate estimation but could yield an incorrect
`respiratory rate. In this case, the PPG quality can be treated
`as acceptable for heart rate analysis but “unreliable” for
`respiratory analysis. Therefore, a good PPG signal selection
`should be based on quality measurement mechanisms that
`contain rich temporal information and could dynamically
`adapt to different application requirements.
`In addition, temporal information of real noisy ADL PPG
`could also benefit learning-based approaches for subsequent
`processing and analysis, e.g., deep neural networks based
`denoising. Roy et al. [19] and Lee et al. [20] proposed two
`different auto-encoder networks for PPG denoising. Due to
`the scarcity of clean ADL PPG signals, synthetic noisy PPG
`and corresponding clean PPG signal are used for training.
`Although promising denoising performance are reported, it
`is difficult to generalize the trained model to real-life settings
`where noise and signal distributions may be significantly
`different from the training data.
`In this work, we propose an automatic, reliable ADL PPG
`selection framework leveraging features of electrocardiogram
`(ECG). Our goal is to select reliable PPG segments from
`collected data where both wrist PPG, chest ECG and associ-
`ated features are available. For each raw PPG segment, we
`produce a time-domain aligned pseudo clean PPG signal that
`contains rich temporal information for vital sign estimation,
`including heart rate and respiration rate. The quality of
`the raw PPG signal is determined by the vital information
`embedded in the pseudo clean PPG. Our contributions are
`summarized as follows:
`• We propose a high quality PPG selection system that
`leverages band-pass filter design followed by Principal
`Component Analysis refinement
`to generate pseudo
`clean PPG signal.
`• Instead of providing a binary PPG signal quality indica-
`tor, our system provides pseudo clean PPG which can be
`used to derive vital information to assess signal quality
`for different applications.
`• Our proposed PPG selection system can collect reliable
`PPG signals from ADL, as opposed to PPG signals with
`synthetic noise added.
`
`II. METHOD
`An overview of our proposed high quality PPG selection
`system is illustrated in Figure 2. The raw single-channel
`PPG is collected from a wrist device, i.e., Samsung Galaxy
`WatchTM. Additionally, simultaneous single-lead ECG and
`accelerometer are collected through the accelerateIQTM plat-
`form using the VitalPatchTM device. Three ECG-based fea-
`tures, i.e., heart rate, respiration rate, and QRS detection, are
`provided by accelerateIQTM. Both heart rate and respiration
`
`rate are average rates of a 1 minute ECG segment and will
`be treated as ground truth of each corresponding 1 minute
`PPG segment. The QRS detection, providing the location of
`each QRS complex [21] in ECG, will be used to derive the
`instantaneous heart rate for the bandpass filter design. To
`measure the quality of a raw PPG segment, we generate the
`corresponding pseudo clean PPG signal and assess whether
`it contains sufficient vital information for health monitoring.
`The proposed pseudo clean PPG signal generation consists
`of a two step process. First, based on the instantaneous heart
`rate, we design a band-pass filter to apply to the raw PPG.
`Then, we use Principal Component Analysis (PCA) to further
`improve the signal quality. Once we obtain the pseudo clean
`PPG signal, we can evaluate its quality by calculating the
`corresponding PPG-based heart rate and respiration rate and
`compare them with the respective accelerateIQTM features. In
`this section, we first illustrate the signal acquisition process
`in Section II-A. We then describe the pseudo clean PPG
`generation in Section II-B and Section II-C. Finally, we
`discuss the process of extracting cardiovascular features from
`the pseudo clean PPG in Section II-D.
`
`A. Signal Acquisition and Data Preparation
`All subjects were asked to wear a Samsung’s Galaxy
`WatchTM on his/her non-dominant hand and VitalConnect’s
`VitalpatchTM on his/her chest, simultaneously over a 7 day
`period. In total, 8 subjects (5 male and 3 female) participated.
`Participants ranged in age from 27 to 60, with a mean
`age of 42.875. Using the accelerateIQTM platform, QRS
`detection and average one minute heart rate and respiratory
`rate were extracted from the chest worn device. As the chest-
`based extracted vital signs are FDA-approved, both chest
`heart rate and respiration rate served as our ground truth
`in this work. The QRS detection records the time-points
`of each QRS complex [21] in ECG. The PPG signal was
`sampled at 25 Hz and all data were segmented into 1 minute
`intervals for analysis. The experimental procedures involving
`human subjects described in this paper were approved by the
`Institutional Review Board.
`
`B. Band-Pass Filter Design
`The chest patch sensor collects vital signs in a much
`robust manner compare with the wrist devices. Based on
`the detected QRS complex in chest ECG, the pauses be-
`tween heartbeats (R-R interval) is recorded and used for
`instantaneous heart rate estimation. We select the 1 minute
`PPG signal and find the corresponding instantaneous heart
`rate range derived from the chest device. The passband of
`bandpass filter is selected based on the range of instantaneous
`heart rate so that all heartbeat activity could be preserved.
`Therefore, the filter should remove the MA out of the cardiac
`activity frequency band.
`To avoid the ripple response of the digital filter, we
`implemented a fifth order Butterworth filter, which has a
`flat frequency response in the passband [22]. Based on the
`instantaneous heart rate derived from the ECG signal, a low-
`pass and a high-pass digital filter are designed and cascaded
`
`AliveCor Ex. 2003 - Page 2
`
`

`

`Fig. 2: Overview of our proposed reliable PPG selection system.
`
`for narrow passband filtering. The filters process the input
`PPG signal in both forward and reverse directions to preserve
`time-domain alignment.
`
`C. Principal Component Analysis
`To further improve the quality of filtered PPG signal,
`we implemented Principal Component Analysis (PCA) to
`remove additional noisy components. As shown in Figure 3,
`we use a fixed n length sliding window in stride of t on
`the 1 minute filtered PPG signal, sampled m overlapped
`PPG observations Xi, and stacked them together as our
`input matrix Xinput. Based on Singular Value Decomposition
`(SVD) [23], the Xinput can at most be decomposed into
`min(m,n) orthogonal basis. The goal of PCA is to further
`reduce the MAs by selecting a subset of the basis. We select
`p principal components and reproject them back into the
`input space for reconstruction. The reconstructed matrix Y
`is also m× n that contains m overlapped PPG observations
`corresponding to Xinput. For overlapped regions in the re-
`constructed signal, we take the average and apply a Gaussian
`filter for waveform smoothing. The final output is a flattened
`1 minute signal for quality assessment.
`
`D. PPG Assessment Using Heartbeat and Respiratory Ac-
`tivities Information
`In order to evaluate the cardio-respiratory features em-
`bedded in the pseudo clean PPG signal, we derive the
`average heart rate and respiration rate of the PPG segment
`and compare to the ground truth chest-derived features. A
`peak detection method based on Sobel filter [24] is adopted
`to obtain the cycle of heartbeat in pseudo clean PPG and
`find the corresponding heart rate value. We calculate the
`Mean Absolute Error (MAE) between the estimation and
`the reference rate from the chest features, which can be
`fomulated as:
`i=1|yi − xi|
`∑n
`n
`is reference heart/respiration rate and xi
`where yi
`is our
`estimated heart/respiration rate. The n represents the total
`number of segments we evaluated. The respiratory signal is
`usually calculated by the temporal information in the PPG.
`Three kinds of Respiratory Modulation are commonly used:
`
`MAE =
`
`(1)
`
`Fig. 3: An example of Principal Component Analysis (PCA)
`on single-channel PPG.
`
`baseline wander (BW), amplitude modulation (AM), and
`frequency modulation (FM) [25]. In our implementation, we
`use the AM and FM for respiration rate estimation since the
`bandpass filter removes the baseline wander.
`AM of the PPG is caused by reduced stroke volume
`during inhalation due to changes in intrathoracic pressure,
`reducing pulse amplitude [26]. FM, on the other hand, is
`the manifestation of the spontaneous increase in heart rate
`during inspiration and decreases during exhalation, known
`as Respiratory Sinus Arrhythmia (RSA) [27]. Therefore, the
`accuracy of respiration rate estimation based on AM and FM
`strongly relies on the temporal information in the pseudo
`clean PPG. For AM, based on the peak detection result,
`we extract the maximum intensity of the PPG pulses. This
`intensity trend sequence is resampled into an even 4-HZ
`sequence for spectrum analysis based on the Fast Fourier
`
`AliveCor Ex. 2003 - Page 3
`
`

`

`Transform [28]. The maximum frequency content within the
`Respiration Rate (RR) frequency range is selected as the
`RR. For FM, the peak detection for heart rate estimation
`provides the timepoints of each heartbeat and can be used
`to determine the beat interval and convert to tachograms.
`We resample the tachograms into an even 4-Hz grid and
`find the maximum power frequency within the RR frequency
`range in the spectrum. The normal RR range is between
`12-30 Breath/Min. (BrPM) [29]. We set the frequencies of
`interest to 10-50 BrPM to include all possible rates. We take
`the mean value to fuse the RR estimations from these two
`modulations.
`
`III. EXPERIMENTAL RESULTS
`The PPG signal is collected from the Samsung Galaxy
`WatchTM. We also simultaneously collect chest ECG signal
`from the VitalPatchTM. Three ECG-based features, i.e., heart
`rate, respiration rate, and QRS detection, are provided by
`accelerateIQTM. The accelerateIQTM platform also removes
`the chest ECG signal during wrist device charging period
`and synchronized wrist PPG and chest ECG signal. A total of
`55,093 samples were collected from 8 subjects. Each sample
`contains 1 minute segment of wrist PPG, the corresponding
`QRS complex timepoints, reference 1 minute average heart
`rate, and reference 1 minute average respiration rate. The
`raw PPG goes through the bandpass filter, whose passband
`is dynamically determined by the instantaneous heart rate
`derived from QRS complex timepoints and then fed to the
`PCA module. In the PCA module, the 1 minute 25 Hz
`PPG signal is a 1,500-length sequence, and we use a 400-
`length sliding window in stride of 5 to segment 220 16-
`s PPG observations and stack them together to perform
`PCA. We empirically select the first 30 components, which
`provides the best correlation coefficient between pseudo
`clean PPG heart rate and reference heart rate for our dataset
`and construct the pseudo clean PPG signal.
`
`A. Pseudo Clean PPG Evaluation
`As shown in Figure 4, our method could generate a pseudo
`clean PPG corresponding to the noisy wrist ADL PPG signal.
`For better visualization, we normalize both 1 minute signals
`and plot the 30-s interval signals. We first evaluate the esti-
`mated average heart rate of all the pseudo clean PPG signals
`and compare with the reference rates from the chest sensor.
`As shown in Figure 5, although there are some unreliable
`PPG segments, the Pearson correlation coefficient of 0.76
`indicates that there is a strong linear relationship between
`the reference heart rates and our estimated pulse rates. We
`also generate the Bland-Altman plot in Figure 6. We observe
`that the majority of the data points (∼ 82.49%) are within
`an general acceptable error of 10 Beats/Min. (BPM) [30],
`[31]. Both plots show that accurate heart rate can be derived
`from most wrist PPG segments successfully. However, the
`data points with large error in plots also indicate that the
`wrist PPG collection contains the unreliable PPG segments.
`To further analyze the distribution of the quality of the
`collected PPG, we split the PPG signals into 4 different
`
`Fig. 4: Comparison between raw PPG collected from a
`subject doing Activities of Daily Living (ADL) and the
`corresponding pseudo clean PPG generated by our method.
`
`quality groups based on the heart rate estimation. If the
`absolute error of the 1 minute average heart rate estimation is
`within 1 BPM, it is level-1 high quality PPG segments. The
`level-2 high quality PPG portion corresponds to heart rate
`estimation error in the range of (1,3] BPM. The error range
`of (3,5] BPM indicates level-3 high quality PPG segments.
`We also consider heart rate estimation error larger than 5
`BPM as low quality data. As shown in Table I, for each group
`we calculate the MAE for both heart rate and respiration
`rate, in BPM and BrPM, based on the methods described in
`Section II-D for each group and also include the portion (%)
`of PPG signals in each group to analyze the distribution.
`Our results show that the pseudo clean PPG segments that
`yield more accurate heart rate estimation also contain higher
`quality temporal information to generate better respiration
`rate estimation. If we set the heart rate error less than 5 BPM
`as the threshold to collect reliable PPG segments, combining
`the first three group results 70.99% PPG to be considered
`as high quality segments. In this case, the average MAE
`of heart rate and respiration rate are 1.46 BPM and 3.93
`BrPM, respectively. For the low quality group, although the
`heart rate MAE 19.25 BPM indicates that the pulse rate
`information of many PPG segments are highly corrupted, the
`MAE of respiration rate is still within 6 BrPM, which means
`the low quality group also contains PPG segments with valid
`respiration rate information. If we use the respiration rate
`MAE as criterion for reliable PPG selection, many PPG
`segments in the low quality group could also be used.
`For each quality group, we also include the whisker plot
`to show the ground truth heart rate and respiration rate
`distribution. As shown in Figure 7a and Figure 8a, the box
`extends from the lower quartile (25%) to upper quartile
`(75%) values of the rates, the red lines indicate the median
`values. Two caps show the minimum and maximum values
`of the heart/respiration rate of each group. We also plot the
`blue scatters to visualize the heart rate and respiration rate
`distributions. Figure 7b and Figure 8b show the zoom-in
`portions of the whisker plots to better visualize the quartile
`
`AliveCor Ex. 2003 - Page 4
`
`

`

`TABLE I: Heart Rate (HR) and Respiration Rate (RR) Esti-
`mation Analysis and Distributions of Different Quality PPG
`Segments (BPM:Beats/Min., BrPM:Breaths/Min.): Level-1
`High Quality: HR error ≤ 1 BPM; Level-2 High Quality:
`1 < HR error ≤ 3 BPM; Level-3 High Quality: 3 < HR
`error ≤ 5 BPM; Low Quality: HR error > 5 BPM.
`
`Quality Group
`
`Level-1 High Quality
`Level-2 High Quality
`Level-3 High Quality
`Low Quality
`
`HR MAE
`(BPM)
`0.46
`1.80
`3.89
`19.25
`
`RR MAE
`(BrPM)
`3.80
`3.95
`4.32
`5.45
`
`Portion
`%
`34.58%
`25.75%
`10.66%
`29.01%
`
`(a) Heart rate distribution of different
`(b) Zoom-in version of quartile val-
`quality groups
`ues
`Fig. 7: Whisker plots depict the distributions of ground truth
`heart rate of each PPG group. The boxes indicate the first
`quartile (25%) and third quartile (75%) values of the heart
`rate and the horizontal red lines indicate the median values.
`
`treat the heart rate error within 10 BPM as high quality
`PPG. In this implementation, we select heart rate error
`less than 5 BPM as high quality PPG and still get 71%
`of collected PPG segments qualified. For certain clinical
`studies, a stricter standard may be applied which requires
`us to consider the temporal details, e.g., heart rate variability
`(HRV) and respiration rate information, of the pseudo clean
`PPG to determine the selection criteria. We will explore this
`in our future work. In addition, our pseudo clean PPG and
`raw noisy PPG are aligned in the time domain because of
`(1) the band-pass filter processes the input PPG signal in
`both forward and reverse directions to eliminate the phase
`distortion, and (2) the PCA module is a linear transformation
`which does not influence the phase response. Therefore, the
`pseudo clean PPG selected from reliable raw PPG could be
`used in training-based methods for subsequent processing
`and analysis.
`
`IV. CONCLUSION
`In this paper, we present a novel framework to select
`reliable wrist ADL PPG based on the cardio-respiratory
`features derived from chest ECG. The pseudo clean PPG is
`generated for each 1 minute raw PPG segment and embeds
`vital information to assess the quality of original raw PPG.
`We designed bandpass filters that are guided by ECG-based
`features with additional PCA refinement. Our experimental
`results show that our pseudo clean PPG not only can be
`used for heart rate estimation but also contains rich temporal
`
`Fig. 5: Pearson correlation plot of heart rate estimated from
`pseudo clean PPG and the chest reference rate in Beats/Min.
`
`Fig. 6: Bland–Altman plot of Wrist Estimated (WE) heart
`rate and the Chest Reference (CR) heart rate in Beats/Min.
`WE heart rate is derived from the pseudo clean PPG.
`
`values. Although the overall distributions of both heart rate
`and respiration rate slightly increase as the quality decrease,
`both figures show that none of the quality group biases
`towards a specific range of heart rate or respiration rate.
`
`B. Discussion
`Collecting reliable PPG in naturalistic setting, e.g., during
`ADLs, is valuable for health monitoring. Therefore, how to
`remove unreliable PPG is an important factor to consider
`during PPG signal selection. In this work, we propose a
`bandpass filter design and additional PCA refinement for
`pseudo clean PPG generation. The passband, as determined
`by instantaneous heart rate, provides the narrowband and
`removes MAs in other frequency range. The PCA module
`further removes small unrelated components of the filtered
`signal so that the most valuable physiological information is
`preserved in the pseudo clean PPG. Due to the flat frequency
`response of our IIR filter design, temporal details such as
`the amplitude of peaks are also well preserved and can be
`used for respiration rate estimation by the AM and FM
`based methods. Our generated pseudo clean PPG can adapt
`to different study purposes for high quality PPG selection.
`For example, commercial purpose health monitoring would
`
`AliveCor Ex. 2003 - Page 5
`
`

`

`[11] R. Yousefi, M. Nourani, S. Ostadabbas, and I. Panahi, “A motion-
`tolerant adaptive algorithm for wearable photoplethysmographic
`biosensors,” IEEE journal of biomedical and health informatics,
`vol. 18, no. 2, pp. 670–681, 2013.
`[12] S. M. L.-S. R. Giannetti, M. L. Dotor, J. P. Silveira, D. Golmayo,
`F. Miguel-Tobal, A. Bilbao, M. Galindo, and P. Mart´ın-Escudero,
`“Heuristic algorithm for photoplethysmographic heart rate tracking
`during maximal exercise test,” Journal of Medical and Biological
`Engineering, vol. 32, no. 3, pp. 181–188, 2012.
`[13] M.-Z. Poh, N. C. Swenson, and R. W. Picard, “Motion-tolerant
`magnetic earring sensor and wireless earpiece for wearable photo-
`plethysmography,” IEEE Transactions on Information Technology in
`Biomedicine, vol. 14, no. 3, pp. 786–794, 2010.
`[14] Z. Zhang, Z. Pi, and B. Liu, “Troika: A general framework for
`heart rate monitoring using wrist-type photoplethysmographic signals
`during intensive physical exercise,” IEEE Transactions on Biomedical
`Engineering, vol. 62, no. 2, pp. 522–531, 2015.
`[15] X. Sun, P. Yang, and Y.-T. Zhang, “Assessment of photoplethys-
`mogram signal quality using morphology integrated with temporal
`information approach,” 2012 Annual International Conference of the
`IEEE Engineering in Medicine and Biology Society, pp. 3456–3459,
`2012.
`[16] K. Li, S. Warren, and B. Natarajan, “Onboard tagging for real-time
`quality assessment of photoplethysmograms acquired by a wireless
`reflectance pulse oximeter,” IEEE Transactions on Biomedical Circuits
`and Systems, vol. 6, no. 1, pp. 54–63, 2011.
`[17] Q. Li and G. D. Clifford, “Dynamic time warping and machine learn-
`ing for signal quality assessment of pulsatile signals,” Physiological
`measurement, vol. 33, no. 9, p. 1491, 2012.
`[18] E. K. Naeini, I. Azimi, A. M. Rahmani, P. Liljeberg, and N. Dutt, “A
`real-time ppg quality assessment approach for healthcare internet-of-
`things,” Procedia Computer Science, vol. 151, pp. 551–558, 2019.
`[19] M. S. Roy, R. Gupta, J. K. Chandra, K. D. Sharma, and A. Taluk-
`dar, “Improving photoplethysmographic measurements under motion
`artifacts using artificial neural network for personal healthcare,” IEEE
`Transactions on Instrumentation and Measurement, vol. 67, no. 12,
`pp. 2820–2829, 2018.
`[20] J. Lee, S. Sun, S. M. Yang, J. J. Sohn, J. Park, S. Lee, and H. C. Kim,
`“Bidirectional recurrent auto-encoder for photoplethysmogram denois-
`ing,” IEEE journal of biomedical and health informatics, vol. 23, no. 6,
`pp. 2375–2385, 2018.
`[21] Z. Abedin and R. Conner, ECG Interpretation: The Self-Assessment
`Approach.
`John Wiley & Sons, 2008.
`[22] M. S. Mahmud, H. Fang, and H. Wang, “An integrated wearable
`sensor for unobtrusive continuous measurement of autonomic nervous
`system,” IEEE Internet of Things Journal, vol. 6, no. 1, pp. 1104–
`1113, 2018.
`[23] G. H. Golub and C. Reinsch, “Singular value decomposition and least
`squares solutions,” in Linear algebra. Springer, 1971, pp. 134–151.
`[24] N. Kanopoulos, N. Vasanthavada, and R. L. Baker, “Design of an
`image edge detection filter using the sobel operator,” IEEE Journal of
`solid-state circuits, vol. 23, no. 2, pp. 358–367, 1988.
`[25] P. H. Charlton, T. Bonnici, L. Tarassenko, D. A. Clifton, R. Beale,
`and P. J. Watkinson, “An assessment of algorithms to estimate res-
`piratory rate from the electrocardiogram and photoplethysmogram,”
`Physiological measurement, vol. 37, no. 4, p. 610, 2016.
`[26] A. A. Alian and K. H. Shelley, “Respiratory physiology and the
`impact of different modes of ventilation on the photoplethysmographic
`waveform,” Sensors, vol. 12, no. 2, pp. 2236–2254, 2012.
`[27] A. Ben-Tal, S. Shamailov, and J. Paton, “Evaluating the physio-
`logical significance of respiratory sinus arrhythmia: looking beyond
`ventilation–perfusion efficiency,” The Journal of physiology, vol. 590,
`no. 8, pp. 1989–2008, 2012.
`[28] E. O. Brigham, The fast Fourier transform and its applications.
`Prentice-Hall, Inc., 1988.
`[29] J. A. Noah, C. Boliek, T. Lam, and J. F. Yang, “Breathing frequency
`the onset of stepping in human infants,” Journal of
`changes at
`neurophysiology, vol. 99, no. 3, pp. 1224–1234, 2008.
`[30] K. Nakajima, T. Tamura, and H. Miike, “Monitoring of heart and
`respiratory rates by photoplethysmography using a digital filtering
`technique,” Medical engineering & physics, vol. 18, no. 5, pp. 365–
`372, 1996.
`[31] J. Townshend, B. J. Taylor, B. Galland, and S. Williams, “Compari-
`son of new generation motion-resistant pulse oximeters,” Journal of
`paediatrics and child health, vol. 42, no. 6, pp. 359–365, 2006.
`
`(a) Respiration rate distribution of
`(b) Zoom-in version of quartile val-
`different quality groups
`ues
`Fig. 8: Whisker plots depict the distributions of ground truth
`respiration rate of each PPG group. The boxes indicate the
`first quartile (25%) and third quartile (75%) values of the
`respiration rate and the horizontal red lines indicate the
`median values.
`
`information for respiration rate prediction. Based on the vital
`information quality of pseudo clean PPG and different appli-
`cation requirements, we can acquire reliable PPG segments
`for ADL. Since the PPG is extracted from subjects in the
`real-life settings, the natural MA recorded in raw PPG is
`highly correlated to daily motions and valuable for investi-
`gating practical PPG denoising and analysis applications for
`daily health monitoring.
`
`REFERENCES
`[1] T. Tamura, Y. Maeda, M. Sekine, and M. Yoshida, “Wearable pho-
`toplethysmographic sensors—past and present,” Electronics, vol. 3,
`no. 2, pp. 282–302, 2014.
`[2] J. Allen, “Photoplethysmography and its application in clinical phys-
`iological measurement,” Physiological measurement, vol. 28, no. 3,
`p. R1, 2007.
`[3] Y. Zhang, S. Song, R. Vullings, D. Biswas, N. Sim˜oes-Capela,
`N. Van Helleputte, C. Van Hoof, and W. Groenendaal, “Motion
`artifact reduction for wrist-worn photoplethysmograph sensors based
`on different wavelengths,” Sensors, vol. 19, no. 3, p. 673, 2019.
`[4] A. E. Johnson, T. J. Pollard, L. Shen, H. L. Li-Wei, M. Feng,
`M. Ghassemi, B. Moody, P. Szolovits, L. A. Celi, and R. G. Mark,
`“Mimic-iii, a freely accessible critical care database,” Scientific data,
`vol. 3, no. 1, pp. 1–9, 2016.
`[5] M. A. Pimentel, A. E. Johnson, P. H. Charlton, D. Birrenkott, P. J.
`Watkinson, L. Tarassenko, and D. A. Clifton, “Toward a robust esti-
`mation of respiratory rate from pulse oximeters,” IEEE Transactions
`on Biomedical Engineering, vol. 64, no. 8, pp. 1914–1923, 2016.
`[6] H. Han, M.-J. Kim, and J. Kim, “Development of real-time motion
`artifact reduction algorithm for a wearable photoplethysmography,”
`2007 29th Annual international conference of the IEEE engineering
`in medicine and biology society, pp. 1538–1541, 2007.
`[7] M. T. Islam, S. Tanvir Ahmed, I. Zabir, C. Shahnaz, and S. A.
`Fattah, “Cascade and parallel combination (cpc) of adaptive filters
`for estimating heart rate during intensive physical exercise from
`photoplethysmographic signal,” Healthcare technology letters, vol. 5,
`no. 1, pp. 18–24, 2018.
`“Combining
`and M. Huang,
`J. Yang,
`[8] C. Wan, D. Chen,
`parallel adaptive filtering and wavelet
`threshold denoising for
`photoplethysmography-based pulse rate monitoring during intensive
`physical exercise,” IEICE Transactions on Information and Systems,
`vol. 103, no. 3, pp. 612–620, 2020.
`[9] G. Joseph, A. Joseph, G. Titus, R. M. Thomas, and D. Jose,
`“Photoplethysmogram (ppg) signal analysis and wavelet de-noising,”
`2014 annual international conference on emerging research areas:
`Magnetics, machines and drives (AICERA/iCMMD), pp. 1–5, 2014.
`[10] K. A. Reddy, B. George, and V. J. Kumar, “Use of fourier series
`analysis for motion artifact reduction and data compression of photo-
`plethysmographic signals,” IEEE transactions on instrumentation and
`measurement, vol. 58, no. 5, pp. 1706–1711, 2008.
`
`AliveCor Ex. 2003 - Page 6
`
`