EXHIBIT 2134
`
`EXHIBIT 2134
`
`IPR2017—003 17
`CONDITIONAL MOTION TO AMEND
`
`VALENCELL, INC.
`EXHIBIT 2134 — PAGE 1
`
`IPR2017-00317
`CONDITIONAL MOTION TO AMEND
`
`VALENCELL, INC.
`EXHIBIT 2134 - PAGE 1
`
`

`

`Proceedings of the 29th Annual International
`Conference of the IEEE EMBS
`Cité Internationale, Lyon, France
`August 23-26, 2007.
`
`ThD13.1
`
`1-4244-0788-5/07/$20.00 ©2007 IEEE
`
`1528
`
`
`
`
`A Comparative Evaluation of Adaptive Noise Cancellation
`Algorithms for Minimizing Motion Artifacts in a Forehead-Mounted
`Wearable Pulse Oximeter
`Gary Comtois, Member IEEE, Yitzhak Mendelson, Member IEEE, Piyush Ramuka
`
` Abstract— Wearable physiological monitoring using a pulse
`oximeter would enable field medics to monitor multiple injuries
`simultaneously, thereby prioritizing medical intervention when
`resources are limited. However, a primary factor limiting the
`accuracy of pulse oximetry is poor signal-to-noise ratio since
`photoplethysmographic (PPG) signals, from which arterial
`oxygen saturation (SpO2) and heart rate (HR) measurements
`are derived, are compromised by movement artifacts. This
`study was undertaken to quantify SpO2 and HR errors induced
`by certain motion artifacts utilizing accelerometry-based
`adaptive noise cancellation (ANC). Since the fingers are
`generally more vulnerable to motion artifacts, measurements
`were performed using a custom forehead-mounted wearable
`pulse oximeter developed for real-time remote physiological
`monitoring and triage applications. This study revealed that
`processing motion-corrupted PPG signals by least mean
`squares (LMS) and recursive least squares (RLS) algorithms
`can be effective to reduce SpO2 and HR errors during jogging,
`but the degree of improvement depends on filter order.
`Although both algorithms produced similar improvements,
`implementing the adaptive LMS algorithm is advantageous
`since it requires significantly less operations.
`
`T
`
`I. INTRODUCTION
`HE implementation of wearable diagnostic devices
`would enable real-time remote physiological assessment
`and triage of military combatants, firefighters, miners,
`mountaineers, and other individuals operating in dangerous
`and high-risk environments. This, in turn, would allow first
`responders and front-line medics working under stressful
`conditions to better prioritize medical intervention when
`resources are limited, thereby extending more effective care
`to casualties with the most urgent needs.
`Employing commercial off-the-shelf (COTS) solutions,
`for example finger pulse oximeters to monitor arterial blood
`oxygen saturation (SpO2) and heart rate (HR), or adhesive-
`type disposable electrodes for ECG monitoring, are
`
`
`
`Manuscript received April 2, 2007. This work is supported by the U.S.
`Army MRMC under Contract DAMD17-03-2-0006. The views, opinions
`and/or findings are those of the author and should not be construed as an
`official Department of the Army position, policy or decision unless so
`designated by other documentation.
`Y. Mendelson is a Professor in the Department of Biomedical
`Engineering, Worcester Polytechnic Institute, Worcester, MA 01609 USA
`(phone: 508-831-5103; fax: 508-831-5541; email: ym@wpi.edu).
`G. Comtois is a graduate student in the Department of Biomedical
`Engineering, Worcester Polytechnic Institute, Worcester, MA 01609 USA
`(comtoisg@wpi.edu).
`P. Ramuka is a graduate student in the Department of Biomedical
`Engineering, Worcester Polytechnic Institute, Worcester, MA 01609 USA
`(pramuka@wpi.edu).
`
`impractical for field applications because they limit mobility
`and can interfere with regular activity. Equally important,
`since these devices are designed for clinical settings where
`patient movements are relatively constrained, motion
`artifacts during field applications can drastically affect
`measurement accuracy while subjects remain active.
`Practically, the primary factor limiting the reliability of
`pulse oximetry is attributed to poor signal-to-noise ratio
`(SNR) due to motion artifacts. Since photoplethysmographic
`(PPG) signals, which are used to determine SpO2 and HR,
`are obscured during movements, the implementation of a
`robust pulse oximeter for field applications requires
`sophisticated noise
`rejection algorithms
`to eliminate
`erroneous readings and prevent false alarms.
`To minimize the effects of motion artifacts in wearable
`pulse oximeters, several groups proposed various algorithms
`to accomplish adaptive noise cancellation (ANC) utilizing a
`noise reference signal obtained from an accelerometer
`(ACC) that is incorporated into the sensor to represent body
`movements [1]-[3]. These groups demonstrated promising
`feasibility for movement artifact rejection in PPG signals
`acquired from the finger. However, they did not present
`quantifiable data showing whether accelerometry-based
`ANC resulted in more accurate determination of SpO2 and
`HR derived from PPG signals acquired from more motion-
`tolerant body locations that are more suitable for mobile
`applications.
`
`II. BACKGROUND
`Generally, linear filtering with a fixed cut-off frequency is
`not effective in removing in-band noise with spectral overlap
`and temporal similarity that is common between the signal
`and artifact. Thus, we utilized ANC techniques to filter noisy
`PPG waveforms acquired during field experiments. The
`performance of
`this signal processing approach was
`evaluated based on its potential to lower SpO2 and HR
`measurement errors.
`Among the most popular ANC algorithms are the least
`mean squares (LMS) and recursive least squares (RLS)
`algorithms. Briefly,
`to attenuate
`the
`in-band noise
`component in the desired signal, these algorithms assume
`that
`the reference noise received from
`the ACC
`is
`statistically correlated with the additive noise component in
`the corrupted PPG signal, whereas the additive noise is
`uncorrelated with the noise-free PPG signal. An error signal
`is used to adjust continuously the filter’s tap-weights in
`order to minimize the SNR of the noise-corrected PPG
`signal.
`
`IPR2017-00317
`CONDITIONAL MOTION TO AMEND
`
`VALENCELL, INC.
`EXHIBIT 2134 - PAGE 2
`
`

`

`1529
`
`The performance of ANC algorithms is highly dependent
`on various filter parameters, including filter order (M).
`Accordingly, careful consideration must be given to the
`selection of these parameters and the trade-off between
`algorithm complexity and its computation time.
`Although the basic principles of the LMS and RLS
`techniques share certain similarities, the LMS algorithm
`attempts to minimize only the current error value, whereas in
`the RLS algorithm, the error considered is the total error
`from the beginning to the current data point. Furthermore,
`the performance of each algorithm depends on different
`parameters. For example, the step size (µ) has a profound
`effect on the convergence behavior of the LMS algorithm.
`Similarly, the forgetting factor (λ) determines how the RLS
`algorithm treats past data inputs.
`Compared to the LMS algorithm, the RLS algorithm has
`generally a faster convergence rate and smaller error.
`However, this advantage comes at the expense of increasing
`complexity and longer computational time which increases
`rapidly and non-linearly with filter order.
`
`III. METHODS
`To simulate movement artifacts, we performed a series of
`outdoor and indoor experiments that were intended to
`determine the effectiveness of using the accelerometer-based
`ANC algorithms
`in processing motion-corrupted PPG
`signals acquired by a forehead pulse oximeter. The focus of
`this study was to compare the performance of each algorithm
`by quantifying the improvement in SpO2 and HR accuracy
`generated during typical activities that are expected to
`induce considerable motion artifacts in the field.
`Data were collected by a custom forehead-mounted pulse
`oximeter developed in our laboratory as a platform for real-
`time remote physiological monitoring and triage applications
`[4]-[6]. The prototype wearable system is comprised of three
`units: A battery-operated optical Sensor Module (SM)
`mounted on the forehead, a belt-mounted Receiver Module
`(RM) mounted on the subject’s waist, and a Personal Digital
`Assistant (PDA) carried by a remote observer. The red (R)
`and infrared (IR) PPG signals acquired by the small (φ =
`22mm) and lightweight (4.5g) SM are transmitted wirelessly
`via an RF link to the RM. The data processed by the RM can
`be transmitted wirelessly over a short range to the PDA or a
`PC, giving the observer the capability to periodically or
`continuously monitor the medical condition of multiple
`subjects. The system can be programmed to alert on alarm
`conditions, such as sudden trauma, or when physiological
`values are out of their normal range. Dedicated software was
`used to filter the reflected PPG signals and compute SpO2
`and HR based on the relative amplitude and frequency
`content of the PPG signals. A triaxial MEMS-type ACC
`embedded within the SM was used to get a quantitative
`measure of physical activity. The information obtained
`through the tilt sensing property of the ACC is also used to
`determine body posture. Posture and acceleration, combined
`with physiological measurements, are valuable indicators to
`assess the status of an injured person in the field.
`
`Body accelerations and PPG data were collected
`concurrently from 7 healthy volunteers during 32 jogging
`experiments. These jogging experiments comprised 16
`treadmill, 12
`indoor, and 4 outdoor exercises. Each
`experiment comprised a 1-minute free jogging at speeds
`corresponding to 3.75–6.5 mph, framed by 2-minute resting
`intervals. For validation, reference SpO2 and HR were
`acquired concurrently from the Masimo transmission pulse
`oximeter sensor attached to the subject’s fingertip which was
`kept in a relatively stationary position throughout the study.
`We chose the Masimo pulse oximeter because it employs
`unique signal extraction technology (SET®) designed to
`greatly extend its utility into high motion environments. A
`Polar™ ECG monitor, attached across the subject’s chest,
`provided reference HR data.
`The ACC provided reference noise inputs to the ANC
`algorithms. The X, Y, and Z axes of the triaxial ACC were
`oriented according to the anatomical planes as illustrated in
`Fig. 1. Accelerations generated during movement depend
`upon the types of activity performed. Generally, during
`jogging, acceleration is greatest in the vertical direction,
`although the accelerations in the other two orthogonal
`directions are not negligible. Therefore, the noise reference
`input applied to the ANC algorithms was obtained by
`summing all three orthogonal axes of the ACC. By
`combining signals from all three axes, measurements
`become insensitive to sensor positioning and inadvertent
`sensor misalignment that may occur during movements. To
`compensate for differences in response times, the SpO2 and
`HR measurements acquired
`from each device were
`processed using an 8-second weighted moving average.
`
`ACC axial
`orientations
`
`Finger
`Sensor
`
`SM + ACC
`
`Polar HR Monitor
`
`RM
`
`Masimo
`Pulse Oximeter
`
`Fig. 1: Experimental setup for data collection.
`
`The outputs of the MEMS ACC and raw PPG signals were
`acquired in real-time at a rate of 80 s/s using a custom
`written LabVIEW® program. Data were processed off-line
`using Matlab programming. The ANC algorithms were
`implemented in Matlab with parameters optimized for
`computational speed and measurement accuracy. The LMS
`algorithm was implemented using a constant µ of 0.016. The
`
`IPR2017-00317
`CONDITIONAL MOTION TO AMEND
`
`VALENCELL, INC.
`EXHIBIT 2134 - PAGE 3
`
`

`

`1530
`
`
`
`
`selected filter parameters for the RLS algorithm were
`λ = 0.99 and an inverse correlation matrix P = 0.1. These
`filter parameters were found to be optimal in preliminary
`experiments. For comparison, data were processed by each
`algorithm using variable order filters.
`
`IV. RESULTS
`SpO2 and HR data were derived from the R and IR PPG
`signals utilizing custom extraction algorithms. SpO2 root
`mean squared errors (RMSE) were quantified based on the
`differences between the readings measured by the custom
`and Masimo pulse oximeters, whereas HR errors were
`defined with respect
`to
`the Polar HR monitor. For
`comparison, RMSE were determined by processing the PPG
`signals off-line either with or without the ANC algorithms.
`Fig. 2 shows a representative tracing of SpO2 and HR
`measurements obtained from the custom pulse oximeter with
`and without ANC. Reference measurements obtained
`simultaneously from the Masimo pulse oximeter and Polar
`HR monitor during resting and outdoor jogging were also
`included for comparison.
`
`
`Masimo
`
`LMS
`
`RLS
`
`Without ANC
`
`Resting
`
`Jogging
`
`Resting
`
`Without ANC
`
`RLS
`
`Polar
`
`Masimo
`
`LMS
`
`frequency band ranging between 1.5–3.0 Hz. Further
`analysis of the data showed that in 8 out of the 32 jogging
`experiments (25%), the cardiac-synchronized frequencies
`and movement-induced acceleration frequencies shared a
`common band.
`The averaged errors observed from the series of 32
`experiments are summarized in Figures 3 and 4. Analysis of
`the data clearly revealed that utilizing either the LMS or
`RLS algorithm to process the noise-corrupted PPG signals
`can improve both SpO2 and HR accuracy during jogging.
`Although the degree of improvement varied, because
`different methods are employed to compute SpO2 and HR
`from
`the PPG signal,
`these figures show
`that
`the
`performance of both algorithms depends on filter order used
`to implement each algorithm.
`
`LMS
`RLS
`
`20
`
`15
`
`10
`
`05
`
`SpO2 RMSE (%)
`
`0
`
`4
`
`8
`
`12
`Filter Order (M)
`
`Fig. 3. Averaged SpO2 errors for varying filter orders. Error bars indicate
`±1SD. For comparison, M = 0 represents the error obtained without ANC.
`
`
`16
`
`20
`
`24
`
`LMS
`RLS
`
`25
`
`20
`
`15
`
`10
`
`05
`
`HR RMSE (%)
`
`
`
`0
`
`4
`
`8
`
`12
`Filter Order (M)
`
`Fig. 4. Averaged HR errors for varying filter orders. Error bars indicate
`±1SD. For comparison, M = 0 represents the error obtained without ANC.
`
`16
`
`20
`
`24
`
`V. DISCUSSION
`Pulse oximeters are used routinely in many clinical
`settings where patients are at rest. Their usage in other areas
`is limited because of motion artifacts which is the primary
`contributor to errors and high rates of false alarms. In order
`to design wearable cost-effective devices that are suitable for
`field deployment, it is important to ensure that the device is
`robust against motion induced disturbances. PPG signals
`recorded from the forehead are generally less prone to
`movement artifacts compared to PPG signals recorded from
`
`
`
`Resting
`
`Jogging
`
`Resting
`
`Fig. 2. Representative SpO2 (top) and HR (bottom) measurements obtained
`during outdoor jogging. Filter order M = 16.
`Spectral analysis of the data using FFT revealed that
`during jogging frequency components associated with body
`acceleration and the subject’s HR shared a relatively small
`
`IPR2017-00317
`CONDITIONAL MOTION TO AMEND
`
`VALENCELL, INC.
`EXHIBIT 2134 - PAGE 4
`
`

`

`1531
`
`essential in real-time applications. However, this comes at
`the expense of a longer computational time since the RLS
`algorithm requires M2 operations per iteration. Considering
`for example that an implementation based on a 24th-order
`filter would provide an acceptable error reduction, this
`implies that the LMS algorithm would require only 24
`operations compared to 576 operations that will be required
`to
`implement an adaptive RLS algorithm. Table 1
`summarizes the relative execution times of the LMS and
`RLS adaptive algorithms for processing one data point.
`
`Table 1. Execution times for LMS and RLS algorithms
`Filter Order
`LMS (ms)
`RLS (ms)
`2
`1.0
`6.5
`4
`1.8
`18.5
`8
`3.2
`63.0
`16
`6.2
`235.0
`
`VI. CONCLUSIONS
`This study was designed to investigate the performance of
`accelerometry-based ANC implemented using the LMS and
`RLS algorithms as an effective method to minimizing both
`and HR
`errors
`induced during movement.
`SpO2
`Measurements were performed using a custom, forehead-
`mounted wearable pulse oximeter that was developed in our
`laboratory to serve as a platform for real-time remote
`physiological monitoring and triage applications. The results
`obtained in this study revealed that processing motion-
`corrupted PPG signals by the LMS and RLS algorithm can
`reduce HR and SpO2 errors during jogging. Although both
`algorithms
`produced
`similar
`improvements,
`the
`implementation of the adaptive LMS algorithm is preferred
`since it requires significantly less operations.
`
`[1]
`
`REFERENCES
`J. Y. A. Foo, S. J. Wilson, “A computational system to optimize noise
`rejection in photoplethysmography signals during motion or poor
`perfusion states,” Med. Biol. Eng. Comput., 2006, pp. 140-145.
`[2] L. B. Wood, H. H. Asada, “Noise cancellation model validation for
`reduced motion artifact wearable PPG sensors using MEMS
`accelerometers,” in Proc. 28th IEEE/EMBS Ann. Int. Conf., New York,
`2006, pp. 3525-3528.
`[3] A.R. Relente, L.G. Sison, Characterization and adaptive filtering of
`motion artifacts in pulse oximetry using accelerometers, Proc. 2nd
`Joint EMBS/BMES Conf., Houston, TX, 2002, 1769-1770.
`[4] Y. Mendelson and V. Floroff, A PDA based ad-hoc mobile wireless
`pulse oximeter, in Proc. IASTED International Conference Telehealth
`2005, Banff, Canada, 2005, pp. 1-6.
`[5] Y. Mendelson, R. J. Duckworth, G. Comtois, “A wearable reflectance
`pulse oximeter for remote physiological monitoring,” in Proc. 28th
`IEEE EMBS Annual International Conf., New York, 2006, pp. 912-
`915.
`[6] G. Comtois, Y. Mendelson, “A noise reference input to an adaptive
`filter algorithm for signal processing in a wearable pulse oximeter,” in
`Proc. 33rd Annual Northeast Bioengineering Conf., New York, 2007,
`pp. 106-107.
`
`
`
` a
`
` finger. Nonetheless, morphological distortions of the
`underlying PPG waveforms, from which SpO2 and HR
`measurements are derived, could lead to measurement
`errors, false alarms, and frequent dropouts when subjects
`remain active. For example, as shown in Fig. 2, it is evident
`that the Masimo pulse oximeter, which employs advanced
`signal extraction technology designed to greatly extend its
`utility into high motion environments, was clearly unable to
`accurately track SpO2 and HR while the subject was jogging.
`Although
`to a
`lesser extent, we also noticed more
`pronounced fluctuations in SpO2 recorded by the wearable
`forehead pulse oximeter during jogging. These fluctuations
`are likely caused by PPG waveforms obscured by motion
`artifacts associated with heavier breathing.
`To address the need to improve the performance of a
`prototype reflectance pulse oximeter during jogging, we
`investigated the effectiveness of a MEMS ACC as a noise
`reference input to two popular ANC algorithms. We chose
`the LMS and RLS adaptive routines since other investigators
`showed the promising utility of these algorithms to reduce
`errors attributed to motion artifacts in pulse oximeters [1]-
`[3].
`Analysis of the data acquired during jogging experiments
`showed that ANC implemented using the LMS and RLS
`algorithms can help to improve considerably the accuracy of
`a pulse oximeter, as shown in Fig. 2. However, although the
`differences are not considered clinically significant, we
`found that processing the corrupted PPG signals by each
`algorithm produced slightly different improvements. These
`differences are anticipated since different computational
`principles are employed by a pulse oximeter.
`Since ANC-based filtering implements an adaptive notch
`filter with a notch frequency corresponding to the dominant
`frequency of the measured ACC signal, we expected that an
`overlap of the HR and movement-induced ACC frequencies
`would attenuate
`the fundamental cardiac-synchronized
`frequency of the PPG signals and, therefore significantly
`affecting SpO2 and HR measurements. However, separate
`analysis of
`the data from experiments where body
`accelerations and cardiac rhythms were found
`to be
`synchronized confirmed that applying either the LMS or
`RLS algorithm did not adversely impact the ability to obtain
`accurate SpO2 and HR readings while subjects remain
`active.
`As shown in Fig. 3 and Fig. 4, we found that the degree of
`improvement depends on the filter order (M) used to
`implement each adaptive algorithm, however filters order
`greater
`than 24 produced diminished
`improvements.
`Furthermore, we also found that the LMS algorithm was
`slightly more effective in reducing HR errors compared to
`the RLS implementation.
`Given similar performances, it is important to take into
`consideration
`the complexity of
`the LMS and RLS
`algorithms and the trade-off between algorithmic complexity
`and computation
`time. These principal
`tradeoffs are
`important since our goal is to implement ANC to improve
`the performance of a wearable pulse oximeter during
`motion. For example, compared to the LMS algorithm, the
`RLS algorithm has a faster convergence rate which is
`
`IPR2017-00317
`CONDITIONAL MOTION TO AMEND
`
`VALENCELL, INC.
`EXHIBIT 2134 - PAGE 5
`
`