`a2) Patent Application Publication (0) Pub. No.: US 2003/0065269 Al
`(43) Pub. Date: Apr. 3, 2003
`
`Vetter et al.
`
`US 20030065269A1
`
`METHOD AND DEVICE FOR PULSE RATE
`DETECTION
`
`(52) QESiGk: soonsnancsrenwesencscncniemns 600/503
`
`(54)
`
`(75)
`
`Inventors: Rolf Vetter, Yverdon (CH); Philippe
`Renevey, Lausanne (CH); Roland
`Gentsch, Hauterive (CH); Jens Krauss,
`Neuchatel (CH); Yves Depeursinge,
`Servion (CH)
`
`Correspondence Address:
`PARKHURST & WENDEL, L.L.P.
`1421 PRINCE STREET
`SUITE 210
`ALEXANDRIA, VA 22314-2805 (US)
`
`(73) Assignee: CSEM Centre Suisse d’ Electronique et
`de Microtechnique SA, Neuchatel (CH)
`
`(21)
`
`Appl. No.:
`
`10/255,068
`
`(22
`
`Filed:
`
`Sep. 26, 2002
`
`(30)
`
`Foreign Application Priority Data
`
`Sep..28; 2001
`
`CEP). .sserssssosssavonenesessenscasensers 01203686.9
`
`Publication Classification
`
`(57) WnteO)? sass AGIB 5/02
`
`(57)
`
`ABSTRACT
`
`There is described a device and a methodfor detecting the
`pulse rate. The measuring principle consists of emitting
`radiant energy at
`the surface of or through human body
`tissue (5) by means ofa light-emitting source (10), measur-
`ing the intensity of the radiant energy alter propagation
`through the human body tissue by meansof at leastfirst and
`secondlight detectors (21, 22, 23, 24) located at a deter-
`mined distance from the light-emitting source and providing
`first and secondinput signals (y,(t), y.(t)) representative of
`this propagation. Simultaneously, a motion detecting device
`(40), suchas a three dimensional accelerometers, provides a
`motion reference signal (a,(t), ay(t), a,(t)) representative of
`motion of the detecting device on and with respect to the
`human body tissue (5). The input signals are then processed
`in order to remove motion-related contributions due to
`motion of the detecting device (1) on and with respectto the
`human body tissue (5) and to produce first and second
`enhancedsignals. This processing basically comprises the
`elaboration of a model of the motion-related contributions
`based on the motion reference signal and the subtraction of
`this model from each of the input signals. Pulse rate is then
`extracted from the enhanced signals using for instance a
`maximum likelihood histogram clustering technique.
`
`A
`
`
`
`
`
`001
`
`U.S. Patent No.
`
`Apple Inc.
`APL1011
`8,923,941
`
`Apple Inc.
`APL1011
`U.S. Patent No. 8,923,941
`
`001
`
`
`
`Patent Application Publication
`
`Apr. 3,2003 Sheet 1 of 6
`
`US 2003/0065269 Al
`
`002
`
`002
`
`
`
`Patent Application Publication
`
`Apr. 3, 2003 Sheet 2 of 6
`
`US 2003/0065269 Al
`
`Fig.3
`
`MOTION
`REFERENCE
`SIGNAL
`ax(t),ay(t),az(t)
`
`INPUT SIGNALS
`yr(t),yatt)
`
`MOTION ARTEFACTS
`REMOVAL USING NON-
`LINEAR MODEL-BASED
`NOISE CANCELLING
`
`MEASUREMENT NOISE
`AND NON-MODELLED
`CONTRIBUTIONS
`REMOVAL USING
`NOISE REDUCTION
`ALGORITHM
`
`RELIABLE CANDIDATE
`
`ENHANCED
`SIGNALS
`
`PULSE RATE
`EXTRACTION
`
`SELECTION OF MOST
`
`003
`
`100
`
`110
`
`120
`
`130
`
`003
`
`
`
`Patent Application Publication
`
`Apr. 3, 2003 Sheet 3 of 6
`
`US 2003/0065269 Al
`
`027%
`
`jB9q-19}u|
`
`JEALOUI
`
`uoloexe
`
`jesodue}-oneds
`
`VOd
`
`Och
`
`OLL
`
`Ole
`
`OOL
`
`Soe
`
`wesHo}siy
`
`BHuaysnjo
`
`0€¢
`
`JBSUI|UON
`
`Buiddew
`
`Otl
`
` asind
`Ayseuoneys-uoN
`90 J0199}9p
`
`dl
`
`sjeuBis
`
`uonoy
`
`sjeubis
`
`BOUBIOJOY
`
`004
`
`004
`
`
`
`Patent Application Publication
`
`US 2003/0065269 Al
`
`Apr. 3, 2003 Sheet 4 of 6
`
`Fig. 5a
`
`005
`
`005
`
`
`
`Patent Application Publication
`
`Apr. 3, 2003 Sheet 5 of 6
`
`US 2003/0065269 Al
`
`ooo*
`
`woi©2.
`
`oO
`
`006
`
`006
`
`
`
`Patent Application Publication
`
`Apr. 3, 2003 Sheet 6 of 6
`
`US 2003/0065269 Al
`
`Fig. Sd
`
`
`
`
`
`
`Time[s]
`
`007
`
`007
`
`
`
`US 2003/0065269 Al
`
`Apr. 3, 2003
`
`METHOD AND DEVICE FOR PULSE RATE
`DETECTION
`
`[0001] This invention is in the field of signal processing
`and is more specifically directed to pulse rate detection.
`[0002] Portable heart rate monitoring devices are classi-
`cally composed ofa processing device and an external probe
`(e.g. electronic stethoscope, optical measure at ear lobe,
`chestbelt for electrocardiogram—ECG-based measurement,
`etc.). The use of an external probe is often considered as a
`reduction of the user’s comfort. ECG-based pulse rate
`detecting devices using external electrode probes are for
`instance disclosed in documents U.S. Pat. Nos. 4,108,166,
`6,018,677, 6,149,602 and WO 00/51680.
`
`[0003] Various pulse rate detection systems are known in
`the art. Pulse rate detection devices using pressure sensitive
`transducers such as piezoelectric elements are for instance
`disclosed in documents U.S. Pat. Nos. 3,838,684, 4,195,642,
`4,331,154, 5,807,267 and WO 80/00912.
`
`[0004] More recently, measuring techniques based on so-
`called photoplethysmography (or PPG) have been proposed.
`PPGis an electro-optic technique of measuring the cardio-
`vascular pulse wave found throughout the human body. This
`pulse wave is caused by the periodic pulsations of arterial
`blood volume and is measured by the changing optical
`absorption of radiant energy which this induces, The mea-
`surement system classically consists of a source of radiant
`energy (usually an infra-red light source), at
`least one
`detectorfor detecting the intensity of the radiant energy after
`propagation through the human body tissue and a data
`processing means for extracting bodily parameters such as
`pulse rate or oxygen concentration in the blood. Infra-red
`light
`is predominantly used since it
`is relatively well
`absorbed in blood and weakly absorbed in body tissue.
`Blood volume changes are therefore observed with a rea-
`sonable contrast. The principal advantage of PPG measure-
`ment resides in the fact that it is entirely non-invasive and
`can be applied to any blood bearingtissue, typically a finger,
`nail, ear lobe, nose and, in some instances, wrist.
`
`[0005] Since light is highly scattered in tissue, a detector
`positioned on the surface of the skin can measure reflections
`(or transmissions) from a range of depths and those reflec-
`lions (or transmissions) are variously absorbed depending on
`whether the light encounters weakly or highly absorbing
`tissue. Any change in blood volumewill be registered by the
`detector at
`the surface since increasing (or decreasing)
`volume will cause more (or less) absorption. The effect will
`be averaged over manyarteries and veins. In the absence of
`any blood volume changes, the signal level will be deter-
`mined by the tissue type, skin type, probe positioning, static
`blood volume content and of course the geometry and
`sensitivity of the sensor itself.
`[0006]
`PPG systems differentiate between light absorption
`due to blood volume and that of other fluid and tissue
`constituents by observation that arterial blood flow pulsates
`while tissue absorption remains static. As the illuminated
`blood flow pulsates,
`it alters the optical path length and
`therefore modulates the light absorption throughout
`the
`cardiac cycle. Non-pulsatingfluids andtissues do not modu-
`late the light but have a fixed level of absorption (assuming
`there is no movement).
`
`(0007] The result of this absorption is that any light
`reflected from (or transmitted through) the pulsating vascu-
`
`lar bed contains an AC component which is proportional to
`and synchronous with the patients plethysmographicsignal.
`It
`is this modulated component which is known as the
`photoplethysmographic signal. This PPG signal is superim-
`posed onto a DC level which represents the difference
`between incident radiant energy and the constant absorption
`ofthe tissue, blood and anything elsein the optical path with
`constant absorption.
`[0008]
`PPG measurement can be achieved by measure-
`ment of the intensity of radiant energy transmitted through
`(transmission mode systems) or reflected by (reflection
`mode systems) body tissue. A reflection mode system typi-
`cally has much poorer signal to noise ratio, resulting from
`the fact that a smaller proportion of the light which is not
`absorbed will be reflected than transmitted. That
`is the
`reason why most of the prior art devices and systems use a
`detecting arrangement that is placed on the user’s finger,
`nail, ear lobe, nose or part of the body through which light
`can easily be transmitted.
`[0009]
`PPG has widely been used for measuring arterial
`oxygen saturation known as pulse oximetry. The technique
`relies on the knowledge that haemoglobin and oxyhaemo-
`globin absorb light
`to varying degrees as a function of
`wavelength. In particular, the absorption characteristics of
`red and near infrared light are inverted for the two species.
`It is thus possible to derive the proportion of oxyhaemoglo-
`bin and therefore the arterial oxygen saturation from a
`knowledge of the absorption characteristics of the arterial
`blood at these two wavelengths. PPG-based oximetry sens-
`ing devices employing sensors whichare typically in contact
`with the user’s finger or nail are for instance disclosed in
`documents U.S. Pat. No. 5,237,994, 5,645,060, 5,662,106,
`5,934,277, 6,018,673, WO 99/52420, WO 99/62399 and
`WO 01/25802. PPG-based oximetry and heart rate detecting
`devices intended to be worn on or around other parts of the
`human body such as the wrist or ear, are also known, for
`instance from documents U.S. Pat. No. 5,807,267 and WO
`97/14357.
`
`[0010] One of the main problems of PPG measurement is
`corruption of the useful signal by ambient light and other
`electromagnetic radiations (so-called light artefacts) and by
`voluntary or
`involuntary subject movement
`(so-called
`motion artefacts). These artefacts lead to erroneous inter-
`pretation of PPG signals and degrade the accuracy and
`reliability of PPG-based algorithms for the estimation of
`cardiovascular parameters.
`[0011]
`Processing of ambient light artefacts is not critical
`because the influence of ambient
`light can be measured
`using multiplexing techniques and the PPG signal can be
`restored using subtractive-type techniques. Reference can
`here be made to the article “Effect of motion, ambient light,
`and hypoperfusion on pulse oximeter function”, Trivedi N.
`et al., Journal of Clinical Anaesthesia, vol 9, pp. 179-183,
`1997,
`for a description of these problems.
`In contrast,
`processing of motion artefacts is a tough task since its
`contribution often exceed that of the useful pulse-related
`signal by an order of magnitude. It is essentially caused by
`mechanical forces that induces changes in the optical cou-
`pling and the optical properties of the tissue. Motion arte-
`facts are a particularly critical problem for the design of a
`wrist-located pulse detecting device.
`[0012] Several methods have been proposed to reduce
`motion artefacts in PPG signals. Feature-based algorithms
`
`008
`
`008
`
`
`
`US 2003/0065269 Al
`
`Apr. 3, 2003
`
`have been proposed to discard the corrupted segments from
`the signals for instance in document WO 94/22360 (corre-
`sponding to U.S. Pat. No. 5,368,026). This kind of approach
`allows one to reduce the occurrence of false alarms in
`clinical environments, but it often degrades the signals with
`small motion artefacts contributions. This could lead to
`erroneous estimation of cardiovascular parameters.
`
`In order to circumvent this drawback, model-based
`(0013]
`noise cancelling techniques have been applied more recently
`for the enhancement of optical signals. Examples are for
`instance described in documents U.S. Pat. No. 5,490,505,
`WO 94/03102 and in articles “Simple photon diffusion
`analysis of the effects of multiple scattering on pulse oxim-
`etry’, Schmitt J., IEEE Transactions on Biomedical Engi-
`neering, vol. 38, pp. 1194-2002, December 1991, and
`“Noise-resistant oximetry using a synthetic reference sig-
`nal”, Coetzee F. M. et al., IEEE Transactions on Biomedical
`Engineering, vol. 47, pp. 1018-1026, August 2000. In such
`approaches a reference signal of motion is recorded and a
`parametric model
`is used subsequently to retrieve motion
`related influences
`in the optical signals. Nevertheless,
`motion references are classically obtained by piezo-sensors
`or optical measures and convey therefore only incomplete or
`local information of motion. This degrades the performance
`of model-based noise cancelling techniques since they
`require complete and low-noise motion reference signals.
`
`It is thus a principal object of the present invention
`[0014]
`to provide a device and method for accurately monitoring
`and detecting heart rate based on photoplethysmography,
`even under intense physical activity.
`
`[0015] More particularly, an object of the present inven-
`tion is to provide a solution that allows for adequate removal
`of ambient
`light and motion contributions in the optical
`signals.
`
`[0016] Another object of the invention is to provide a
`solution that
`is suitable for enabling measurement and
`detection to happen at the wrist level.
`
`[0017] Accordingly there is provided a portable pulse rate
`detecting device the features of which are recited in claim 1.
`
`Thereis also provided a method for detecting pulse
`[0018]
`rate the features of which are recited in claim 17.
`
`(0019] Other advantageous embodiments ofthe invention
`are the object of the dependent claims.
`
`invention, an accurate
`[0020] According to the present
`motion detecting device is used to provide a reliable motion
`reference signal. This motion detecting device is preferably
`a fully integrated three dimensional accelerometer which
`exhibits a high accuracy and very low noise.
`
`In order to achieve efficient removal of motion
`{0021]
`related artefacts in the optical signals, nonlinear model-
`based techniques are applied. This nonlinear modelling
`preferably consists in a polynomial expansion model using
`a moving average and an associated model selection based
`on the Minimum Description Length (MDL)criterion.
`
`Furthermore, in order to grasp the spatial diversity
`[0022]
`of the optical characteristics of the tissue, at least two optical
`sensors are used. This two-channel arrangement, associated
`with an adequate noise reduction algorithm (preferably an
`algorithm based on so-called spatio-temporal Principal
`
`Component Analysis or PCA), allows one to remove mea-
`surement noise and non-modelled stochastic signal contri-
`butions as well as reduce artefacts related to finger move-
`ments which are generally not recorded by the accelerometer
`and therefore not initially cancelled.
`[0023] Eventually,
`the heart rate is estimated from the
`enhanced signals using inter-beat extraction based on physi-
`ological properties of cardiac cells and maximum likelihood
`histogram clustering of the resulting time series.
`[0024] An assessment of the performance of the proposed
`solution according to the invention has shown its high
`robustness and accuracy. It has to be pointed out that the
`application of nonlinear
`instead of
`linear modelling
`decreases the standard deviation of the detected heart rate of
`about one to two percent. This is mainly dueto the inclusion
`of the parsimonious MDL-based model selection, which
`avoids an overfitting of the time series. Indeed,
`the full
`nonlinear model would retain pulse related components in
`the estimate of the motion artefacts. Since these components
`are subtracted from the optical signals, the quality of the
`enhanced signal and consequently the reliability of the
`estimated pulse are reduced. In contrast, MDL selects only
`movement related parameters in the model, which yields
`higher enhancement performance and a more accurate pulse
`estimation in adverse noisy environments.
`[0025] Other aspects,
`features and advantages of the
`present invention will be apparent upon reading the follow-
`ing detailed description of non-limiting examples and
`embodiments made with reference to the accompanying
`drawings, in which:
`[0026] FIG. 1 is a schematic view of the bottom side
`(intended to come into contact with the body tissue) of a
`portable pulse rate detecting device according to the inven-
`tion which is adapted to be worn on the wrist and comprising
`a light source and twopairs of light detectors arranged at the
`bottom side;
`
`{0027] FIG, 2 is a schematic side view of the device of
`FIG. 1 further illustrating the arrangement of the acceler-
`ometer;
`
`(0028] FIG. 3 is a flow chart of the preferred method for
`pulse rate detection according to the invention;
`[0029] FIG. 4 is a block diagram illustrating a dual
`channel pulse detection algorithm according to the present
`invention which is based on nonlinear model-based motion
`artefact cancelling, coherence-based reduction of measure-
`ment noise and stochastic signal contributions, and a pulse
`detection using maximum likelihood histogram clustering;
`and
`
`FIGS.5a to 5e are diagrams respectively illustrat-
`[0030]
`ing the evolution, as a functionof time,(a) ofoptical signals
`provided by two light detectors, (b) of acceleration signals
`detected by the accelerometer along three measurement
`axes, (c) of the two optical signals after removal of the
`motion artefacts, (d) of the two optical signals after mea-
`surement noise removal (using PCA) and (¢) a correspond-
`ing ECG electrocardiogram.
`[0031] FIGS. 1 and 2 schematically show a top view of
`the bottom side and a side view of a wrist-located pulse rate
`detecting device, indicated globally by reference numeral1,
`according to a preferred embodiment of the present inven-
`tion.
`
`009
`
`009
`
`
`
`US 2003/0065269 Al
`
`Apr. 3, 2003
`
`[0032] While the invention will be described hereinbelow
`with respect to a portable device which is adapted to be worn
`on the wrist and which is based on the measurement oflight
`reflected in the body tissue, it will be appreciated that the
`detecting device according to the present invention could be
`designed to be worn on other parts of the human body such
`as a patient’s finger, nail, ear lobe or any other suitable
`member or part of the human body. In addition, the same
`principles could be applied to a detecting device based on
`the measurement of light transmitted through the body tissue
`(such as those typically used in pulse oximetry) where the
`signal to noise ratio is higher. In addition, these principles
`could be applied for pulse oximetry on the red and IR
`signals.
`
`[0033] As shown in FIGS. 1 and 2, detecting device 1
`comprises a housing 2 and a strap 3 for attaching detecting
`device 1 on the patient’s wrist. Housing 2 comprises, located
`in a bottom side 2a of the device in contact with the skin, a
`light source 10 for emitting radiant energy at the surface of
`(or through) the humanbody tissue, designated by reference
`numeral 5, Light source 10 is preferably an infrared light
`emitting device (LED).
`
`{0034] According to the preferred embodiment, housing 2
`further includes two pairs oflight detectors 21, 22 and 23,
`24 for detecting the intensity of the radiant energy after
`propagation through the human body tissue. Such light
`detectors may conveniently be photodiodes. Preferably, the
`pairs 21, 22 and 23, 24 of light detectors are respectively
`disposedalongfirst and second axes,indicated by references
`A and B, which are substantially perpendicular and parallel
`to the longitudinal axis of the strap, respectively. More
`specifically,
`light source 10 is located in a substantially
`central part of bottom side 2a and light detectors 21 to 24 are
`disposed around and at a determined distance from light
`source 10. In this example, this distance is advantageously
`selected to be approximately equal to 10 mm.
`
`{0035] According to the invention, it will be appreciated
`that at least
`two light detectors are required for a proper
`detection ofthe heart rate. The detecting device of FIGS. 1
`and 2 could thus be designedto have only one pair, three or
`even more than four light detectors. The number and spatial
`arrangement of these light detectors should however be
`selected in an adequate mannerto provide sufficient spatial
`diversity for removing light-related artefacts and, as this will
`be seen hereinafter,
`to remove other contributions which
`cannot be detected by the accelerometer, such as reciprocal
`contributions due to finger movements. In that regard, the
`two-axes arrangement illustrated in FIGS, 1 and 2 has the
`advantage of allowing a good detection of such finger-
`related reciprocal contributions.
`
`[0036] Referring again to FIGS. 1 and 2, housing 2
`further comprises a motion detecting device 40 which is for
`example disposed in an upper part 2b of housing 2. This
`motion detecting device 40 is preferably a three dimensional
`accelerometer,
`that
`is,
`in effect, three accelerometers dis-
`posed along three orthogonal measurement axes and pro-
`viding three dimensional acceleration data representative of
`the acceleration to which the device is subjected. This
`accelerometer is preferably and advantageously an acceler-
`ometer of the type manufactured by the company Colibrys
`S. A. under reference MS 6100. It will however be appre-
`ciated that other types of accelerometers or motion detecting
`
`devices could be used provided they deliver a reliable
`measure of motion of the pulse rate detecting device on and
`with respect to the human body tissue.
`
`Processing of the signals can either be done by an
`[0037]
`external processing unit linked to the portable device (by
`means of a direct or wireless connection) or preferably by an
`adequately programmed digital signal processor or DSP
`(indicated schematically by reference numeral 50 in FIG, 2)
`housed within the device.
`
`[0038] Optionally, the portable pulse rate detecting device
`according to the invention may further comprise means for
`outputting an indication ofthe detected pulse rate in the form
`of an optical, audible signal, or other sensorial signal. Such
`means could be a display, a buzzer, a vibrating device or any
`other suitable device adapted for transmitting information
`representative of the pulse rate measurement to the user.
`Additionally, the detecting device may also comprise alarm
`means for generating an alarm when the detected pulse rate
`reaches a determined threshold, which could be either a low
`or high threshold or both.
`
`[0039] The basic principle of the invention resides in
`emilting an optical infrared (IR) signal at the surface of the
`human body tissue (or alternatively through the body tissue).
`This signal is then propagated through the tissue whereit is
`submitted to modifications due to reflection, refraction,
`scattering and absorption. ‘The resulting signal, after propa-
`gation through the tissue is grasped by the light detectors.
`Since variations of optical tissue characteristics are related
`to variations in the subcutaneous blood flow, the received
`signal can be used for the estimation of the heart rate.
`
`(0040] When light is transmitted through biological tissue,
`several mechanisms are involved in the interaction between
`
`the light and the tissue. These interactions are reflection,
`refraction, scattering and absorption. Reflection and refrac-
`tion occur at
`the interfaces between the probe and the
`subject. Scattering is due to the microscopic variations of the
`dielectric properties of the tissue. These variations are due to
`the cell membranes and the sub-cellular components(e.g.
`mitochondria and nuclei). For infra-redlight, the absorption
`is mainly due to chromophores such as haemoglobin, myo-
`globin, cytochrome, melanin, lipid, bilirubin, and water. The
`relative importance depends on the wavelength considered
`and their distribution in the tissue.
`
`[0041] Under ideal steady-state condition, the received IR
`light signal contains both a constant (DC) and a time varying
`(AC) component. The constant component
`is generally
`ascribed to baseline absorption ofblood and soft tissue, non
`expansive tissue such as bone, as well as reflectance loss.
`The time varying componentreflects the modification of the
`effective path length due to the expansion of the Ussues
`subject to the varying blood pressure.
`
`For the near IR wavelength, the light propagation
`[0042]
`into the tissue Is governed by scattering and absorption. The
`so-called Beer-Lambert equation is generally used to
`describe the phenomenon oflight absorption in biological
`lssue:
`
`010
`
`010
`
`
`
`US 2003/0065269 Al
`
`Apr. 3, 2003
`
`Li =60)ox) axenaso]
`
`jal
`
`ab
`
`[0048] where s,(t), s,(t) are pulse pressure related signal
`contributions,n,,,,(t), 0,,.(t) are artefacts due to motion and
`gravity, n,(t), o,(t) include measurement noise and non-
`modelled stochastic signal contributions and N, is the num-
`ber of observed samples.
`
`[0043] where 1,(t) and |,(t) are the input and output light
`intensity, ). is the wavelength oflight and c,(t), d,(t) and e,j
`represent, respectively, the concentrations, the spanning path
`length and the absorption coefficient of the different com-
`ponents. For further information about
`this subject, refer-
`ence can be made to the articles “Noise-resistant oximetry
`using a synthetic reference signal”, Coetzee F. M. et al.,
`IEEE Transactions on Biomedical Engineering, vol. 47, pp.
`1018-1026, August 2000, and “A review of the optical
`properties of biological tissues”, Cheong W. -F. et al., IEEE
`Journal of Quantum Electronic, vol. 26, pp. 2166-2185,
`1990,
`
`[0044] As briefly mentioned in the preamble part of the
`description, voluntary or involuntary movements corrupt the
`PPG signal and create motion-related artefacts. It is gener-
`ally accepted that motion artefacts are mainly due to modi-
`fication of the optical properties of the tissue (modification
`of blood pressure, modification of the optical path, etc.).
`These modifications affect the corresponding components of
`the Beer-Lambert equation. Therefore,
`in presence of
`motion artefacts, the received intensity can be rewritten in
`function of the major contributions
`
`(2)
`((O=E(OYYeissue'Ypatse(!)Ygraviry (O-Ymorion(t)
`[0045] where y,;,..,.
`18 the static attenuation due to the
`tissue, Y,,,1.e(t) is due to pulsatile absorption of the blood,
`¥, ssccigkt is due to change of position and y,,,otion(t) is due to
`dynamic changesof the tissue induced by the movement of
`the arm (assuming the device is worn on the wrist). It is
`obvious that the different contributions become additive if
`one takes the logarithm of expression (2) above.
`
`(0046] When the subjectis static, only the contributions of
`Yputse(t) changes with time and it is then straightforwardto
`remove the other contributions using a high-passfiltering.
`When the subject is moving, however, the contribution of
`the gravity and the modification of the interface between the
`detecting device and the body tissue are varying with time
`and they have to be removed from the signals in order to
`allow an accurate estimation of the heart rate. The contri-
`butions of the gravity are al
`low frequency and can be
`removed quite easily by an adaptation of the gain. The
`contributions of the motion is difficult to remove, especially
`if it is in the same frequency bandas the heart rate. Therefore
`techniques have to be developed in order to remove the
`motion artefacts to obtain an accurate estimation of the heart
`rate.
`
`It has been shown abovethat IR-signals recorded at
`[0047]
`the wrist are mainly affected by perturbations, such as tissue
`modifications, motion and gravity related artefacts. The
`main issue resides in the estimation of the mean heart rate
`from short time recordings of IR-signals (e.g. 10 seconds).
`It is assumedthat the tissue properties do not vary over the
`observed duration and for a dual channel approach,
`the
`log-corrected observed IR-signals (y,(t), v(t) given by
`expression (2) can be written as
`yi (O51 (OM (O40
`t-0,...,;N-1
` YAlN=satHryalf+n2(0)
`
`(3)
`
`In order to obtain a robust pulse detectionin a large
`[0049]
`variety of experimental conditions, namely non-stationary
`environments, the proposed method according to the present
`invention works on a frame-to-frame basis with a frame
`
`duration of e.g, 3 seconds and it consists of mainly a three
`step algorithm as shown in FIG,3.
`
`the observed optical signals
`In a first step 100,
`[0050]
`y,(t), y(t) are enhanced using nonlinear, model-based noise
`cancelling techniques (see for instance “Adaptive Filter
`Theory”, Haykin S., Prentice Hall, 1991). For this to be
`achieved, according to the present
`invention, an accurate
`motion reference signal (i.e. acceleration signals a,(t), a(t)
`and a,(t)) is provided by the accelerometer. The non-linear
`modelling essentially consists in a polynomial expansion
`model and an associated model selection based on the
`
`Minimum Description Length (MDL)criterion. Such tech-
`niques are already known as such by those skilled in the art.
`Reference can for instance be made to “Nonlinear Biomedi-
`cal Signal Processing” Celka P. et al., vol. 2, IEEE Press,
`2000, and to the PhD thesis of M. R. Vetter (co-inventor)
`entitled “Extraction of efficient and characteristic features of
`
`multidimensional time series”, EPFL Lausanne (Switzer-
`land) 1999, which are both incorporatedherein by reference.
`
`[0051] The use of the parsimonious MDLselectioncrite-
`rion avoids an overfitting of the time series and ensures in
`this way that no pulse pressure related signal contributions
`are cancelled.
`
`In a second step 110, measurement noise and
`[0052]
`non-modelled stochastic signal contributions in the two
`recorded channels are preferably removed. This is achieved,
`according to the preferred embodiment of the present inven-
`tion, by a noise reduction algorithm based on spatio-tem-
`poral Principal Component Analysis (PCA). For further
`information about
`this PCA algorithm, reference will be
`madeto the article “Blind source separation in highly noisy
`environments”, Vetter R. et al., in First International Work-
`shop on Independent Component Analysis and Signal Sepa-
`ration” (ICA’99), Aussois (France), pp. 491-496, 1999,
`which is also incorporated herein by reference. This step is
`not as such compulsory since a pulse rate measurement
`could be derived from the input signals after removal of the
`motion-related contributions.
`
`In addition to the removal of measurement noise
`[0053]
`and non-modelled signal contributions, spatio-temporal
`PCAallows one to reduce artefacts related to finger move-
`ments, which are generally not cancelled in step 100. Indeed,
`finger movements do not necessarily imply a global dis-
`placement of the forearm and are therefore not grasped by
`the accelerometer. Finger movements, often imply tny,
`reciprocal tendon related displacement of the forearm tissue,
`which yields reciprocal artefact contributions in the two
`channels. Due to the reciprocity of these signal contribu-
`tions, they can efficiently be cancelled by a spatio-temporal
`PCA.
`
`011
`
`011
`
`
`
`US 2003/0065269 Al
`
`Apr. 3, 2003
`
`Ina third step 120,the pulse rate is extracted from
`[0054]
`the enhanced IR-signals. This extraction essentially consists
`ofan inter-beat interval extraction achieved through a clas-
`sical maximum detection procedure, preferably with inhibi-
`tion of peak detection during the refractory periodof cardiac
`cells. In addition, a maximum likelihood histogram cluster-
`ing of the resulting inter-beat
`intervals is performed (cf.
`“Vector Quantization and Signal Compression”, Gersho A.
`et al., Kluwer Academic Publishers, 1992).
`
`[0055] Eventually, in a fourth step 130, the most reliable
`candidate can be selected. A robust and reliable estimate of
`the pulse rate can be obtained through a nonlinear mapping
`of the two candidate values in function of their reliability
`measures. This nonlinear mapping is advantageously
`achieved by Multiple Layer Perceptron (MLP), which has
`been trained on data of various experimental setups as
`described in “Neural Networks”, Haykin S., Macmillan
`College Publishing Company Inc., 1994.
`the preferred
`[0056] A more detailed description of
`embodiment of the present invention will now be described
`in reference to the diagrams of FIG. 4 and FIGS. 5ato Se.
`FIG, 4 showsa diagram illustrating the preferred algorithm
`according to the invention where block 200 refers to the
`nonlinear modelling based on the motion reference signal
`(a,(0), a,(0), a,()), block 210 refers to the measurementnoise
`and non-modelled contributions cancellation using PCA,
`block 220 refers to the inter-beat interval extraction on the
`two enhanced signals, block 225 refers to the maximum
`likelihood histogram clustering, block 226 refers to the
`detection of the non-stationary signal segments, and block
`230refers tothe final selection of the mostreliable candidate
`using a nonlinear mapping technique.
`[0057] One of the key elementin the proposed algorithm
`is the nonlinear model, which provides an estimation of the
`motion related contributions in the observed IR-signals
`(block 200 in FIG. 4). The relationship between time
`varying optical characteristics and its influence on IR-
`signals is globally described by the Beer-Lambert law here-
`inabove, Even though one can obtain linear characteristics
`of these variations of the optical characteristics by a loga-
`rithmic transformation, their relationship to a global motion
`reference signal, such as the one grasped by the accelerom-
`eters is complex and may be nonlinear. In order to take into
`account these potential nonlinear contributions, a third order
`polyno