POSTER 2017, PRAGUE MAY 23 1

Spatial Analysis of Cardiorespiratory Thorax Movementwith Motion Capture and Deconvolution

Christoph HOOG ANTINK1, David HEJJ1, Bernhard PENZLIN1

1Philips Chair for Medical Information Technology, Helmholtz-Institute for Biomedical EngineeringRWTH Aachen University, Pauwelsstr. 20, 52074 Aachen, Germany

hoog.antink@hia.rwth-aachen.de, david.hejj@rwth-aachen.de, penzlin@hia.rwth-aachen.de

Abstract. Unobtrusive sensing is a growing aspect inthe field of biomedical engineering. While many modal-ities exist, a large fraction of methods ultimately re-lies on the analysis of thoracic movement. To quantifycardiorespiratory induced thorax movement with spatialresolution, an approach using high-performance motioncapture, electrocardiography and deconvolution is pre-sented. In three healthy adults, motion amplitudes areestimated that correspond to values reported in the lit-erature. Moreover, two-dimensional mappings are cre-ated that exhibit physiological meaningful relationships.Finally, the analysis of waveform data obtained via de-convolution shows plausible pulse transit behavior.

KeywordsCardiorespiratory Movement, Motion Capture,Deconvolution, Biosignalprocessing, UnobtrusiveSensing

1. IntroductionAmbient and unobtrusive cardiorespiratory sens-

ing techniques form an increasing sub-field withinbiomedical engineering [1]. Sensor principles range fromcamera-based methods [2] over ultrasound [3], radar [4]and laser [5] to force sensors [2] and high-frequency os-cillator circuits [6]. For a comprehensive overview, theinterested reader is referred to [1]. While the sensorsspan a wide range of devices and physical principles,many methods that allow for unobtrusive monitoringof heart and lung are ultimately based on the analysisof thoracic movement. Most of the time, unobtrusivesensing modalities are evaluated in terms of their abil-ity to detect respiratory rate, respiratory patterns, heartrate or beat-to-beat intervals. Thus, they are commonlycompared to a medical gold standard, such as the elec-trocardiogram (ECG) or a an air flow sensor. However,spatially resolved quantification of the actual thoracicmotion is seldom performed. For this, specialized mea-surement equipment is necessary. Moreover, if cardiac

induced motion is to be quantified, sub-millimeter ac-curacy is required. At the same time, spatially resolvedinformation could help to optimize regions of interestfor existing and future sensing modalities.

Thus, the aim of this proof-of-concept study is todevelop a method for the spatially resolved analysisand mapping of cardiorespiratory induced thorax move-ment. For this, healthy volunteers sitting in an armchairare monitored with a high-performance motion capture(MoCap) system and an ECG. To extract and quantifythe spatial impulse responses of cardiac and respiratoryactivity, deconvolution [7] is used.

2. Materials and MethodsA Schematic of the overall system setup is given

in Fig. 1. Three of the seven cameras used for MoCapwhich are placed around the subject in a semicircle arevisualized.

DAQ

PC

ECG

Motion CaptureSystem

PassiveMarkers

yz

Fig. 1.Schematic overview of the system setup.

2 C. HOOG ANTINK, ANALYZING CARDIORESPIRATORY THORAX MOVEMENT WITH MOCAP AND DECONVOLUTION

2.1. HardwareFor motion capture, the “Oqus” system from Qual-

isys AB, Göteborg, Sweden configured with seven “Opus500+” infrared (IR) tracking cameras and ten passivereflective markers was used. Each camera has a resolu-tion of 4 megapixels at fs,max = 180Hz and is equippedwith a C-mount lens with a focal length of 13mm. Theaperture was set to f/4.0 and manual focusing was per-formed. Illumination was provided by rings of IR LEDsintegrated into each camera unit. Marker positions weretracked at fs,M = 100Hz. For ECG acquisition, the“MP30” patient monitor, manufactured by Philips, Am-sterdam, The Netherlands was used as analog front end.Digitalization was performed at fs,A = 800Hz with aDAQ system which was synchronized with the cameras.

2.2. Trial SetupThree healthy volunteers participated in the trial

which was conducted as self-experimentation. Aschematic of the marker positions is given in Fig. 2 A),which are further described in Tab. 2.2. Fig. 2 B) showsa photograph of the marker positions for one subjectand in Fig. 2 C), the coordinate system is given, x be-ing the right/left axis. Since the subjects are sitting ina chair with approximately 20◦ recline, y and z only ap-proximate the dorsoventral and the craniocaudal axis,respectively.

1

23

45

6 7 8

9 10

ECG1 ECG2

ECGref

B) C)A)

x

y

z

Fig. 2.A) schematic of marker positions, B) photogra-phy of actual marker positions, C) coordinate sys-tem.

# Position x Position z1 Central axis Center of forehead2 Right nipple 1 cm above right nipple3 Central axis Sternum, superior part4 Left nipple 1 cm above left nipple5 Central axis Sternum, inferior part6 Right nipple 1 cm below costal margin7 Central axis As 6 and 88 Left nipple 1 cm below costal margin9 Centered between 6 and 7 Navel

10 Centered between 7 and 8 Navel

Tab. 1. Position of reflective markers for motion capture.

2.3. Data AnalysisThe proposed model is shown in Fig. 3. The mo-

tion of each marker i is described as yi and is influencedby a cardiac and a respiratory component. For eachmarker, these components are generated by filtering avirtual train of impulses originating from the heart andthe lung with marker-specific transfer functions acard,iand aresp,i, respectively. If yi, acard,i and aresp,i areknown, the filter coefficients can be estimated via de-convolution [7].

heart

lung

virtual source

signal scard

filter a respi,

filter ai,card

+thoracic

motion signals yi

virtual source

signal sresp

yi,card

yi,resp

Fig. 3.Proposed source-filter structure of thoracic move-ment influenced by a superposition of cardiac andrespiratory component. For each marker i andaxis x, y, z, individual filters a exist.

In the proposed approach deconvolution is per-formed separately for the cardiac and the respiratorycomponent. For either component, the convolution isdescribed by

yi[n] =Q∑τ=0

ai[τ ]s[n− τ ] + z[n], (1)

where n is the discrete time, Q the filter order and zis additive noise. The subscripts ‘card’ and ‘resp’ areomitted for reasons of brevity. In the Fourier domain,this can be expressed via

yi[k] =

Q∑τ=0

ai[τ ]s[k]e−j2πkτ/N + zi[k]. (2)

Here xi[k], s[k] and zi[k] are the Fourier transforms oftheir time-domain equivalents. The discrete frequencyis marked by k and N is the number of samples. Thefilter coefficients are not transformed but the delay isexpressed explicitly by the term e−j2πkτ/(N−1). Thiscan be expressed in matrix notation as

yi[k] = ai · e[k] · s[k] + zi[k], (3)

with the Fourier-domain delay-vector

e[k] =[1, e−j2πk/N , ..., e−j2πkQ/N

]T(4)

and the vector of filter coefficients

ai = [ai[0], ai[1], ..., ai[Q]] . (5)

POSTER 2017, PRAGUE MAY 23 3

Let [·]T be the transpose, [·]H the Hermitian transposeand [·]∗ the conjugation operator. If the source signalis known, the optimal filter coefficients in terms of aminimal quadratic error can be determined with

ai,est =[y∗i

· EST + yi· ESH

]· (6)[

ES∗ · EST + ES · ESH]−1

with the matrix

ES = [e[0] · s[0], e[1] · s[1], . . . , e[N − 1] · s[N − 1]] (7)

and the vector

yi

=[yi[0], y

i[1], . . . , y

i[N − 1]

]. (8)

For a more detailed derivation, the interested reader isreferred to [8], where a modified version of the algorithmis applied to to multichannel blind deconvolution. Inthis scenario, filter and source are unknown and inferredfrom multichannel observations.

Note that the calculations are carried out sepa-rately for all i markers, for the respiratory componentand for the cardiac component. Moreover, a separatefilter is estimated for the each dimension x, y, and z. Toobtain the cardiac related motion yi,card, the marker po-sition is filtered with a second-order Butterworth band-pass with 0.8 to 30Hz passband. The respiratory com-ponent yi,resp is obtained via filtering with a passbandof 0.01 to 0.3Hz. The cardiac impulse signal scard isobtained from R-peaks of the ECG, while sresp is ob-tained via peak-detection on the geometric mean of yand z component of the central marker #5. To quantifythe cardiac and respiratory induced motion, the rangeof ai,est, i.e. the maximum minus the minimum of therespective impulse response, is calculated. For visual-ization, the range is color-coded and interpolated, themapping is shown in Fig. 4.

1

23

45

6 7 8

9 10

ECG1 ECG2

ECGref

Fig. 4.Mapping of the color-coded ranges, see Fig. 6.Note that the mapping is slightly distorted asthe subjects were monitored in a sitting positionwith a recline of approximately 20◦.

3. Results and DiscussionIn Fig. 6, the color-coded ranges are visualized.

Several observations can be made. First, neither car-diac nor respiratory activity create a significant x (i.e.left-right) motion in any subject, which is to be ex-pected. Next, the strongest respiratory motion is ob-served in the y (dorsoventral) direction and occurs inthe belly area, whereas the strongest z (craniocaudal)component is measured in the upper part of the thorax.This is consistent with normal physiological breathingin rest, where the ribcage is lifted only slightly and thestrong downward motion of the diaphragm displaces or-gans which extend the belly. In terms of amplitude, themaximum respiratory induced motion is in the rangeof 5 to 7mm, while the maximum cardiac induced mo-tion ranges from 0.25 to 0.3mm, which is consisted withvalues reported in the literature [9, 10]. In conclusion,more than an order of magnitude lies between the car-diac and the respiratory effect. Similar to the obser-vations made with respect to respiration, the strongestz component can be measured in the upper part of thethorax close to the heart, whereas the strongest y move-ment of the thorax due to cardiac activity occurs in thebelly region. At the same time, a relatively strong y-motion of approximately 0.25mm can be observed onthe foreheads of all subjects. Compared to the thoracicregion, the influence of respiratory induced motion onthe forehead is relatively low.

In Fig. 5, the impulse response of the z-componentof marker 4 and 9 are shown for subject 1. On can see

0 200 400 600 800

-0.15

-0.1

-0.05

0

0.05

0.1

0.15

Fig. 5. Impulse response of the z component of cardiacinduced thoracic movement for subject 1, mark-ers 4 and 9. Each maximum value is marked withan x.

that the largest peak occurs at 110ms in the timecourseof marker 4, whereas for marker 9, the largest peak isfound at 310ms. Thus, the marker closest to the heartexhibits a delay of 110ms with respect to the R-peak,i.e. the electrical activity, whereas the pulse transit timebetween both markers is 200ms.

4 C. HOOG ANTINK, ANALYZING CARDIORESPIRATORY THORAX MOVEMENT WITH MOCAP AND DECONVOLUTION

1

2

3

4

5

6

1

2

3

4

5

6

7

1

2

3

4

5

0.1

0.15

0.2

0.25

0.1

0.15

0.2

0.25

0.3

0.05

0.1

0.15

0.2

0.25

0.3

Fig.6. Mapping of respiratory (left) and cardiac (right) induced movement, see also Fig. 4. For each subject, anindividual color code was used for respiratory and cardiac movement, which was constant for x, y, and zcomponent. Values are given in mm.

POSTER 2017, PRAGUE MAY 23 5

4. Conclusions and OutlookIn this paper, the use of motion capture and de-

convolution to quantify cardiorespiratory thorax move-ment was demonstrated. Using a high-performance Mo-Cap system and an ECG, movement amplitudes fromthree healthy subjects were obtained that are consis-tent with values reported in the literature. From thesevalues, spatially resolved maps were created that ex-hibit physiological plausible distributions. Moreover,the phenomenon of pulse transit could be observed andquantified. In the future, a larger study with a statis-tical analysis needs to be performed. It is hoped thatour findings can be used in the future for the optimiza-tion of regions of interest for unobtrusive sensing, forexample by focusing on the head or belly region.

AcknowledgementsResearch described in the paper was supervised by

Univ.-Prof. Dr.-Ing. Dr. med. Steffen Leonhardt.The authors would like to thank Jakob Orschulik forparticipating in the measurements. CHA gratefully ac-knowledges financial support provided by the GermanResearch Foundation (DFG), grant no. LE 817/26-1.

References[1] BRÜSER, C., HOOG ANTINK, C., WARTZEK, T., WAL-

TER, M., LEONHARDT, S., Ambient and UnobtrusiveCardiorespiratory Monitoring Techniques, IEEE Reviews inBiomedical Engineering, 2015, vol. 8, pp. 30–43.

[2] HOOG ANTINK, C., GAO, H., BRÜSER, C., LEON-HARDT, S., Beat-to-beat heart rate estimation fusing mul-timodal video and sensor data, Biomedical Optics Express,2015, vol. 6, no. 8, pp. 2895–2907.

[3] YAMANA, Y., TSUKAMOTO, S., MUKAI, K., MAKI, H.,OGAWA, H., YONEZAWA, Y., A sensor for monitoringpulse rate, respiration rhythm, and body movement in bed,33rd Annual Internation Conference of the IEEE EMBS(EMBC), Boston, MA, USA, vol. 2011, 2011 pp. 5323–6.

[4] MASSAGRAM, W., LUBECKE, V., HOST-MADSEN, A.,BORIC-LUBECKE, O., Assessment of Heart Rate Variabil-ity and Respiratory Sinus Arrhythmia via Doppler Radar,IEEE Transactions on Microwave Theory and Techniques,2009, vol. 57, no. 10, pp. 2542–2549.

[5] HONG, H., FOX, M.D., Noninvasive detection of cardiovas-cular pulsations by optical Doppler techniques, Journal ofBiomedical Optics, 1997, vol. 2, no. 4, pp. 382–390.

[6] OUM, J.J., LEE, S.S., KIM, D.W., HONG, S., Non-contactheartbeat and respiration detector using capacitive sensorwith Colpitts oscillator, Electronics Letters, 2008, vol. 44,no. 2, p. 87.

[7] ZHAO-SHUI HE, SHENG-LI XIE, YU-LI FU, A new blinddeconvolution algorithm for SIMO channel based on neuralnetwork, 2005 International Conference on Machine Learn-ing and Cybernetics, IEEE, 2005 pp. 3602–3616 Vol. 6.

[8] HOOG ANTINK, C., BRUSER, C., LEONHARDT, S., Mul-timodal Sensor Fusion of Cardiac Signals via Blind Deconvo-lution: A Source-Filter Approach, Computing in Cardiology,2014, vol. 41, pp. 805–808.

[9] DROITCOUR, A.D., Non-contact measurement of heartand respiration rates with single-chip microwave Dopplerradar, Diss. Stanford University, 2006.

[10] SUZUKI, S., MATSUI, T., SUGAWARA, K., ASAO, T.,KOTANI, K., An Approach to Remote Monitoring of HeartRate Variability (HRV) Using Microwave Radar duringa Calculation Task, Journal of PHYSIOLOGICAL AN-THROPOLOGY, 2011, vol. 30, no. 6, pp. 241–249.

About the AuthorsChristoph HOOG ANTINK

was born in Lohne (Oldenburg), Ger-many in 1985 where he finished hisAbitur in 2005 and then began hisstudies in Aachen. As a visiting stu-dent on a Fulbright Travel Grant heobtained a M.S. degree in mechani-cal engineering from the University atBuffalo, Buffalo, New York, USA in

2011. In 2012 he finished his diploma in electrical en-gineering at the RWTH Aachen University where he iscurrently working towards a Ph.D. degree in electricalengineering at the Philips Chair for Medical Informa-tion Technology. His research interests include medicalsensor fusion and imaging technologies.

David HEJJ was born in Bu-dapest, Hungary in 1991. He fin-ished his A-levels in 2010 and pro-ceeded with his Bachelor studies inMechatronics at the Technical Uni-versity of Budapest. He special-ized on Biomechatronics and spent aSemester abroad in Karlsruhe, Ger-many with a DAAD grant. He re-

ceived his Bachelor of Science degree (with honors) in2014. He got the scholarship of the Mummert Founda-tion and began the Biomedical Engineering M.Sc. inAachen, Germany, at RWTH Aachen University. Heis currently working on his thesis at the Philips Chairfor Medical Information Technology and will obtain theMasters degree in 2017. His interests include sensortechnologies, signal analysis, and complex prostheses.

Bernhard PENZLIN wasborn in Bad Soden am Taunus,Germany, and received the Dipl.-Ing.degree from OvGU University,Magdeburg, Germany, in 2012. Hehad been a scientific employee at thechair for mechatronics, Institute forMobile Systems at OvGU Magdeburg

University from end of 2012 to September 2014. Cur-rently he is a scientific employee and PhD student atthe Philips Chair for Medical Information Technology,Helmholtz-Institute for Biomedical Engineering atRWTH Aachen University. His research interestsinclude rehabilitation robotics, especially exoskeletons,design of active orthosis and trajectory control.