566
`
`IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 3, MARCH 2006
`
`REFERENCES
`[1] S. K. Moore, “Making chips to probe genes,” IEEE SpectrumMag., vol.
`38, no. 3, pp. 54–60, Mar. 2001.
`[2] R. Jörnsten, W. Wang, B. Yu, and K. Ramchandran, “Microarray
`image compression: SLOCO and the effect of information loss,” Signal
`Process., vol. 83, pp. 859–869, 2003.
`[3] Information Technology—JPEG 2000 Image Coding System, ISO/IEC
`Standard 15 444-1, 2000.
`[4] D. S. Taubman and M. W. Marcellin, JPEG 2000: Image Compression
`Fundamentals, Standards and Practice. Norwell, MA: Kluwer Aca-
`demic, 2002.
`[5] Information Technology—Coded Representation of Picture and Audio
`Information—Progressive Bi-Level Image Compression, ISO/IEC Stan-
`dard 11 544, Mar. 1993.
`[6] Information Technology—Lossless and Near-Lossless Compression of
`Continuous-Tone Still Images, ISO/IEC Standard 14 495-1, 1999.
`[7] J. Hua, Z. Xiong, Q. Wu, and K. Castleman, “Fast segmentation and
`lossy-to-lossless compression of DNA microarray images,” presented
`at the Workshop Genomic Signal Processing and Statistics, GENSIPS,
`Raleigh, NC, Oct. 2002.
`[8] N. Faramarzpour, S. Shirani, and J. Bondy, “Lossless DNA microarray
`image compression,” in Proc. 37th Asilomar Conf. Signals, Systems, and
`Computers, vol. 2, Nov. 2003, pp. 1501–1504.
`[9] S. Lonardi and Y. Luo, “Gridding and compression of microarray im-
`ages,” presented at the IEEE Computer Society Bioinformatics Conf.,
`(CSB-2004), Stanford, CA, Aug. 2004.
`
`Motion Artifact Reduction in Photoplethysmography Using
`Independent Component Analysis
`
`Byung S. Kim and Sun K. Yoo*
`
`Abstract—Removing the motion artifacts from measured photoplethys-
`mography (PPG) signals is one of the important issues to be tackled for the
`accurate measurement of arterial oxygen saturation during movement. In
`this paper, the motion artifacts were reduced by exploiting the quasi-peri-
`odicity of the PPG signal and the independence between the PPG and the
`motion artifact signals. The combination of independent component anal-
`ysis and block interleaving with low-pass filtering can reduce the motion
`artifacts under the condition of general dual-wavelength measurement. Ex-
`periments with synthetic and real data were performed to demonstrate the
`efficacy of the proposed algorithm.
`
`Index Terms—Block interleaving, ICA, motion artifact, photoplethys-
`mography.
`
`I. INTRODUCTION
`
`Photoplethysmography (PPG) is an electro-optic technique to mea-
`sure the pulse wave of blood vessels. In pulse oximeter, the measuring
`apparatus for PPG [1], motion artifacts can limit the accuracy of the
`measured PPG signal during movement. Particularly, the motion arti-
`facts cannot be easily managed because of the frequency overlapping
`
`Manuscript received November 12, 2003; revised July 1, 2005. This work
`was supported by the Korea Health 21 R & D Project, Ministry of Health &
`Welfare, Republic of Korea under Grant 02-PJ3-PG6-EV08-0001. Asterisk in-
`dicates corresponding author.
`B. S. Kim is with Graduate School of Biomedical Engineering, Yonsei uni-
`versity, Seoul 120–752, Korea.
`*S. K. Yoo is with the Department of Medical Engineering, Center for
`Emergency Medical
`Informatics, Human Identification Research Center,
`Yonsei University College of Medicine, Seoul 120–752, Korea (e-mail:
`sunkyoo@yumc.yonsei.ac.kr).
`Digital Object Identifier 10.1109/TBME.2005.869784
`
`Fig. 1.
`
`ICA model for motion artifact separation.
`
`between PPG and the motion artifact signals [2]. Since general fre-
`quency domain filtering methods can be unsuccessful, some methods
`have been researched to manage the motion artifacts from measured
`PPG signals [1], [2]. However, further research is still required to im-
`prove the performance of motion artifact rejection.
`In this paper, the new motion artifact reduction method was proposed
`under the constraint of dual-wavelength measurement. We combined
`independent component analysis (ICA) and a signal enhancement pre-
`processor to separate the PPG signal from the motion artifact-contami-
`nated measured signals. Experiments with synthetic and real data were
`performed to demonstrate the efficacy of the proposed algorithm.
`
`II. MOTION ARTIFACT REDUCTION
`
`The motion artifact reduction method, consisting of the preprocessor
`and the ICA, is newly designed based on the quasi-periodicity of PPG
`signal and the independence between the PPG and the motion artifact
`signals. The preprocessor enhances the PPG component from mea-
`sured signal and then the ICA separates the PPG signal from prepro-
`cessed signal. The preprocessor consists of period detection, block in-
`terleaving, low-pass filtering, and block de-interleaving. In particular,
`the ICA model with two independent sources is considered to comple-
`ment the popular dual-wavelength optical probe.
`
`A. ICA Model for Motion Artifact Separation
`
`The PPG and motion artifact signal sources can be assumed to be
`independent of each other, since the heart pulsation for the PPG signal
`has little correlation with the physical movement for the motion artifact
`signal. As shown in Fig. 1, two measured signals (X), can be modeled
`as the linear mixture of motion artifact and PPG signal sources (S) with
`an unknown mixing matrix (A), if they are independent
`
`X = AS:
`
`(1)
`
`The unknown A and the unknown S can be estimated from the mea-
`sured X (motion artifact contaminated signals) by ICA. The separated
`sources U (= S), the PPG signal and the motion artifact signal, can
`be obtained by estimated W (= A1). The W can be estimated by a
`fast ICA algorithm [3], [4]. In other words, the PPG source separation
`achieves the motion artifact reduction.
`However, the actual number of independent sources contained in the
`measured X can be more than two. The motion artifact signal is postu-
`lated as the complex combination of multiple sources [2]. In addition to
`the motion artifacts, other noise can be added to X [1]. In order to sep-
`arate PPG from multiple sources using the ICA model for two indepen-
`dent sources, the preprocessor should be employed to suppress noise in
`measured X, which in turn enhances the PPG signal comparing with
`other noise sources, before applying the ICA model.
`
`B. Preprocessor for PPG Signal Enhancement
`
`In order to remove noise without the deterioration of the PPG signal,
`we exploited the quasi-periodicity of the PPG signal. The PPG signal
`
`Authorized licensed use limited to: The George Washington University. Downloaded on September 14,2021 at 17:19:45 UTC from IEEE Xplore. Restrictions apply.
`
`0018-9294/$20.00 © 2006 IEEE
`
`AliveCor Ex. 2002 - Page 1
`
`

`

`IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 3, MARCH 2006
`
`567
`
`Fig. 2. Block interleaving.
`
`associated with heart pulsation is inherently quasi-periodic [5]. How-
`ever, in measured X, most noise is not periodic, and even periodic noise
`may not be synchronized with heart pulsation. Hence, if the period of
`the PPG signal is appropriately estimated, the noise can be suppressed
`without affecting the PPG signal.
`The preprocessor consists of period detection, block interleaving,
`low-pass filtering, and block de-interleaving. The period of the PPG
`signal can be estimated by auto-correlation [6]. Then, the block
`interleaving shuffles the measured signal samples into rearranged
`sample sequence, which has a different order based on the estimated
`period of the PPG signal. As shown in Fig. 2, a block consists
`of M periodic waves with N samples. The block interleaver is
`implemented by selecting the starting point from the continuous
`input samples, writing the bits into a matrix (N  M) row by row,
`and then reading them column by column. The block interleaving
`process can group input samples into frequency-related samples.
`As shown in Fig. 2, M interleaved sample points can comprise the
`DC line (low-frequency) if waves are periodic, and samples are
`synchronized with the period. Hence, the low frequency components
`and the high frequency components in block interleaved samples
`are associated with the synchronized samples (periodic PPG signal)
`and the nonsynchronized samples (noise), respectively. Because of
`the frequency rearrangement property of the block interleaving, the
`noise associated with high frequency components can be simply
`reduced by low-pass filtering without deterioration of the PPG signal.
`The low-pass filtering is conducted by three sample moving average
`filter [6]. Indirectly, the optimum number of M can be determined
`by means of numerical experimentaion. When we adjusted the M
`of the known signal (typical PPG signal obtained from normal
`human) with the added noise of 3-dB signal–to-noise ratio (SNR),
`it was observed that the quality of the separated signal converged
`after M increased to more than 10. The M of 10 can be chosen
`as one of the possible tradeoff values between the quality and
`the delay. Finally,
`the block de-interleaving, which reverses the
`interleaving operation, recovers the interleaved sampled order for
`the application of ICA.
`
`III. RESULTS
`
`As shown in Fig. 3, the performance comparison between the pro-
`posed ICA with the preprocessing method (PICA) and the ICA only
`method (ICA) was evaluated by the use of two synthesized reference
`signals, 5 sin(2n=300) and 2 sin(2n=300), with different levels of
`synthesized motion artifact signals (0–13.5 dB SNR). The period of
`300 (1.66 Hz) of the reference signals approximately matched the heart
`rate of 100 beats/min associated with a 500-Hz sampling frequency.
`Three different sinc functions (rectangular pulses in the time domain)
`with different levels and center frequencies were inter-mixed in the
`frequency domain to synthesize three different motion artifact signals
`having different accentuated frequencies. Subscripts 1, 2, and 3 corre-
`spond to accentuated frequencies of 0, 2.5, and 5 Hz, respectively. In
`
`Fig. 3. The performance comparison between the proposed ICA with
`preprocessing (PICA) method and the ICA-only method (ICA) using a
`synthesized reference with a period of 1.66 Hz and three motion artifact
`signals. Subscripts 1, 2, and 3 correspond to accentuated frequencies of 0, 2.5,
`and 5 Hz for the synthesized motion artifact signals, respectively.
`
`Fig. 4. MSE with respect to the amount of period mismatch.
`
`both the ICA and PICA methods, the mean-square error ( MSE)—be-
`tween the known reference signal and the separated PPG signal—in-
`creases as the power of the motion artifact increases, and as the amount
`of frequency overlapping between the reference and the motion artifact
`signals increases (the MSE performances for an accentuated frequency
`of 5 Hz, ICA3 and PICA3, are significantly better than those for 0 Hz
`and 2.5 Hz). However, PICA shows a markedly smaller MSE than the
`ICA, as the SNR decreases (the level of motion artifacts increases).
`Moreover, PICA is less sensitive to the frequency overlapping than is
`ICA for all three accentuated frequencies.
`The performance degradation due to inaccurate estimation of the
`periodicity was tested as shown in Fig. 4. When we adjusted the
`amount of period mismatch of the synthesized reference signals with
`known period artificially, significant quality degradation of the sepa-
`rated PPG signal was observed after the amount of period mismatch
`reached 30 ms.
`Experiments using real data were conducted using dual-wavelength
`illumination (890 nm for infrared illumination, and 660 nm for red illu-
`mination) with a sampling frequency of 500 Hz. Four different types of
`motion artifacts were produced by the following artificial movements:
`1) pressurizing the probe clip; 2) bending the finger; 3) waving the
`
`Authorized licensed use limited to: The George Washington University. Downloaded on September 14,2021 at 17:19:45 UTC from IEEE Xplore. Restrictions apply.
`
`AliveCor Ex. 2002 - Page 2
`
`

`

`568
`
`IEEE TRANSACTIONS ON BIOMEDICAL ENGINEERING, VOL. 53, NO. 3, MARCH 2006
`
`Fig. 5. The proposed method (PICA) was visually compared with the ICA-only method (ICA) with respect to four different types of motion artifacts.
`
`hand; 4) the combined motions of 1), 2), and 3). The PICA was visu-
`ally compared with ICA, as shown in Fig. 5. The PICA can separate
`the motion-artifact reduced PPG signals (s[n]) and the motion-artifact
`signals (N[n]) better than ICA for all four test cases.
`
`IV. DISCUSSION
`
`Some factors should be carefully considered for the successful ap-
`plication of the proposed method. First, the number of independent
`sources in the ICA model is important, since it can affect the process of
`the pursuit of statistical independence from multivariate statistical data
`[4]. Nevertheless, the number of independent sources can be restricted
`by the number of input channels available in measuring apparatus. One
`of the possible alternatives that would increase the number of indepen-
`dent sources is to adopt triple-wavelength illumination rather than the
`dual-wavelength illumination. However, the performance gain associ-
`ated with the use of an additional independent source is not significant,
`since the preprocessing (interleaving with low-pass filtering) can effec-
`tively reduce multiple noise sources associated with the number of in-
`dependent sources. In addition, the use of dual-wavelength illumination
`is of practical significance due to its popularity [1]. Secondly, inaccu-
`rate estimation of the periodicity due to the imperfection of the period
`estimator can directly influence the block interleaving (refer to Fig. 4).
`The slight amount of inaccurate period estimation of PPG signal and of
`heart rate variability during continuous measurement can be tolerable,
`
`but specific procedures should be considered to cope with the undesir-
`able situation where the periodicity estimator has failed. Thirdly, the
`number of periodic waves (M) is associated with the processing delay
`and the tolerability to period variation. As M increases, the tolerability
`to period variation increases positively due to the accumulated effect
`of multiple waves, but the delay increases negatively due to the block
`interleaving. Finally, the proposed method is less sensitive to the SNR
`of the measured signal (refer to Fig. 3) than is the existing ICA, and it
`is also less sensitive to the selection of starting point, because the fre-
`quency rearrangement property associated with the block interleaving
`is irrespective of the initial selection of starting point.
`
`REFERENCES
`[1] J. G. Webster, Design of Pulse Oximeters. Bristol, U.K.: IOP, 1997.
`[2] M. J. Hayes and P. R. Smith, “A new method for pulse oximetry pos-
`sessing inherent insensitivity to artifact,” IEEE Trans. Biomed. Eng., vol.
`48, no. 4, pp. 452–461, Apr. 2001.
`[3] A. Hyvrinen, “Fast and robust fixed-point algorithms for independent
`component analysis,” IEEE Trans. Neural Netw., vol. 10, no. 3, pp.
`626–634, Mar. 1999.
`[4] A. Hyvrinen, J. Karhunen, and E. Oja, Independent Component Anal-
`ysis. New York: Wiley, 2001.
`[5] P. E. McSharry, G. D. Clifford, L. Tarassenko, and L. A. Smith, “A
`dynamical model for generating synthetic electrocardiogram signals,”
`IEEE Trans. Biomed. Eng., vol. 50, no. 3, pp. 289–294, Mar. 2003.
`[6] D. F. Elliot, Handbook of Digital Signal Processing Engineering Appli-
`cations. New York: Academic, 1987.
`
`Authorized licensed use limited to: The George Washington University. Downloaded on September 14,2021 at 17:19:45 UTC from IEEE Xplore. Restrictions apply.
`
`AliveCor Ex. 2002 - Page 3
`
`