Determining Cardiopulmonary Resuscitation Parameters with

Differential Evolution Optimization of Sinusoidal Curves

Christian Lins

, Andreas Klausen

, Sebastian Fudickar

, Sandra Hellmers

, Myriam Lipprandt

Rainer R

ohrig

and Andreas Hein

OFFIS – Institute for Information Technology, Oldenburg, Germany

Carl von Ossietzky University, Oldenburg, Germany

Keywords:

CPR Training, Curve Fitting, Evolutionary Algorithm, Cardiac Massage.

Abstract:

In this paper, we present a robust sinusoidal curve ﬁtting method based on the Differential Evolution (DE)

algorithm for determining cardiopulmonary resuscitation (CPR) parameters – naming chest compression fre-

quency and depth – from skeletal motion data. Our implementation uses skeletal data from the RGB-D (RGB

+ Depth) Kinect v2 sensor and works without putting non-sensor related constraints such as speciﬁc view an-

gles or distance to the system. Our approach is intended to be part of a robust and easy-to-use feedback system

for CPR training, allowing its unsupervised training. We compare the sensitivity of our DE implementation

with data recorded by a Laerdal Resusci Anne mannequin. Results show that the frequency of the DE-based

CPR is recognized with a variance of ±4.4 bpm (4.1%) in comparison to the reference of the Resusci Anne

mannequin.

1 INTRODUCTION

Cardiac arrests represent one of the most prominent

diseases and can signiﬁcantly affect the independent

living if medical treatment is not available within 3-5

minutes. Thus, the proper training of medical person-

nel is as essential as the training of non-specialists,

which can offer resuscitation support much faster.

Since Advanced Life Support (ALS) resuscitation

training, due to the high material and training costs,

is only used for medical professionals, technological

training systems might represent a well-suited alter-

native to train both professionals and non-specialists.

The functioning of the human body depends on

a continuous support of oxygen and glucose due to

the following fundamental biological processes, cla-

rifying the urgency of support. The organs of the hu-

man body are composed of various specialized cells,

e.g., the nerve cells in the brain. These cells need

energy, i.e., in the form of glucose and oxygen to keep

structure and function. With the respiration, the air

gets into the lungs. In the lungs, the oxygen of the

air gets into the blood. The blood carries oxygen and

glucose - fragmented carbohydrates from the food -

to the cells. The heart pumps the blood through the

circulation.

Thus, in case of a cardiac arrest, the transport of

oxygen and glucose to the cells stops immediately.

First of all the reserves of glucose and oxygen in the

cells are consumed and the functionality remains. If

the cardiac arrest persists, the function of the cells

will be reduced, at ﬁrst without permanent damage. If

there is no blood circulation for a long time the struc-

ture of the cells will be irreparably damaged. For the

cells of the nervous system including the brain, that

means that the functionality reduces after 10 seconds

(i.e., loss of consciousness). The death of the cells

begins after about 3 minutes. (Schmidt et al., 2011)

With medical treatment including ventilation, me-

dication, and deﬁbrillation (Advanced Life Support,

ALS) (Soar et al., 2015) there is a high probability of

maintaining sufﬁcient circulation. In the case of a car-

diac arrest, the most critical measure is to do a cardiac

massage ideally in combination with rescue breathing

(Basic Life Support, BLS) (Perkins et al., 2015). By

doing this, a minimal circulation to carry oxygen to

the nerve cells is sustained. During a cardiac mas-

sage, the heart is compressed by orthogonal pressure

onto the breastbone. The depth of this compression is

ideally 5 cm. The blood in the heart is ejected. A com-

plete decompression is crucial to ﬁll the heart with

blood again. The frequency of the cardiac massage is

100 to 120 beats per minute (bpm). If it is possible

to do a sufﬁcient rescue breathing the heart pressure

Lins C., Klausen A., Fudickar S., Hellmers S., Lipprandt M., RÃ˝uhrig R. and Hein A.

Determining Cardiopulmonary Resuscitation Parameters with Differential Evolution Optimization of Sinusoidal Curves.

DOI: 10.5220/0006732806650670

In Proceedings of the 11th International Joint Conference on Biomedical Engineering Systems and Technologies (HEALTHINF 2018), pages 665-670

ISBN: 978-989-758-281-3

massage will be stopped for a short time and will be

continued after two accelerated breathings.

In summary, the following points are important for

the BLS:

• the position of the hands in the middle of the bre-

astbone between the nipples,

• orthogonal pressure onto the breastbone,

• depth of pressure of 5 cm,

• frequency of compressions at 100 to 120 beats,

per minute,

• complete decompressions of the thorax,

• and when doing a rescue breathing:

– short interruptions of the heart pressure mes-

sage after every 30 compressions,

– 2 rapid ventilations,

– fast continuation of the heart pressure message

Learning the right technique is necessary, to per-

form high-quality resuscitation. The method of resus-

citation can be trained with simulation mannequins.

Most of these mannequins provide a real-time feed-

back of the quality of cardiac massage and breathing.

Although the simulation mannequins provide imme-

diate and precise feedback, they are costly systems.

An approach with a low-cost RGB-D camera such as

the Microsoft Kinect (or even an RGB camera with

software skeleton tracking) would provide such trai-

ning to a larger audience thus improving the quality of

the CPR. Additionally, a technique only relying on ex-

ternal (ambient) sensors would retrieve feedback data

even from real resuscitations.

Two crucial parameters of the CPR are the fre-

quency at which the compressions are performed and

the compression depth of the chest. Typically, the fre-

quency is stated as beats (compressions) per minute

(bpm) and the compression depth in cm or mm. In

this paper, we focus on these two parameters and pro-

pose a method to derive these parameters from motion

data coming from an RGB-D-based skeleton tracking

(Microsoft Kinect v2). We show how the Differen-

tial Evolution optimization algorithm can be used to

ﬁt a sinusoidal curve which robustly provides the two

wanted parameters. The outline of the paper is as fol-

lows: in section 2, we discuss related work. In section

3, we introduce the details and speciﬁcs of the Diffe-

rential Evolution algorithm. In section 4, we intro-

duce our concept and a ﬁrst implementation of our

system. Afterwards, we evaluate the sensitivity as a

pilot-study in section 5. We conclude with some lear-

ned lessons.

2 RELATED WORK

There are a few approaches to support CPR training

with RGB-D sensors and resulting feedback.

Tian et al. use Kinect data to model a virtual envi-

ronment with patient and performer and drive a haptic

device in the real world (Tian et al., 2014). In the vir-

tual environment, the performer can see the CPR on a

virtual avatar while responding to the sensation of the

haptic device. The authors focus more on the basics

of cardiac massage than speciﬁc optimization of the

training.

Semeraror et al. and Loconsole et al. present re-

sults from their system called RELIVE which is si-

milar to our approach (Semeraro et al., 2017; Locon-

sole et al., 2016). RELIVE uses data from the Ki-

nect v1 and extract depth and frequency parameters

from the motion data with the intention to improve

the quality of CPR (training). In contrast to our ap-

proach they use the raw RGB pixel and depth image

data of the Kinect to identify hands, arms and the trai-

ning body. We compare our results of the frequency

detection in the last section of this paper. The prede-

cessor to RELIVE is probably the Mini-VREM tool,

which is a Kinect with a software-based audio- and

video-feedback-system (Semeraro et al., 2013) but re-

quires a colored marker at the subject’s hands.

Wang et al. have utilized the Kinect v1 sensor

to create a real-time feedback system for CPR trai-

ning (Wang et al., 2017). The system shows the cur-

rent compression depth and frequency on a computer

screen so that the trainee can adapt her or his actions

accordingly. The authors have evaluated their system

with 100 health care professionals and conclude that

the system can signiﬁcantly improve the CPR quality

at least for trainees with a body weight < 71kg.

Higashi et al. developed and evaluated an aug-

mented reality system that enables the user to cor-

rect her or his posture while performing cardiac mas-

sage compression (Higashi et al., 2017). The focus of

this work lays on the correct posture of the performer

primarily to differentiate between extended position

compression and bent position compression. Unfor-

tunately, they do not provide a quantitative evaluation

of their system.

3 CONCEPT

3.1 CPR Parameters Determination

with Kinect

With our approach it is possible to derive CPR quality

parameters such as the compression (CC) frequency

Skeletal

Motion Data

Feature

Selection

Distance to

Floor Plane

● Hands

● Elbows

● Shoulders

Sinusoidal

Curve Fitting

CPR

Parameter

Feedback

Improvement

Cardiac

massage

Figure 1: CPR training with Kinect.

and the compression depth with one (or more) exter-

nal RGB-D sensors. It does not require a simulation

mannequin and is therefor suitable for CPR trainings

or in-situ observations, e.g. in a shock room. Figure

1 shows the process of our approach within CPR trai-

ning.

3.2 Differential Evolution Optimization

of Sinusoidal Curves

The Differential Evolution (DE) algorithm (Storn and

Price, 1997) is a generic evolutionary optimisation al-

gorithm, which is used here to solve four parameters

of a sinusoidal curve. Our approach utilizes the perio-

dic nature of the CPR to ﬁt the time series of distances

(upper limbs to ground) to a sine curve.

The generic parameters sine function can be writ-

ten as follows:

y(t) = A · sin(2 f t + ρ) + D

Parameter A and f are of primary interest here;

A is the amplitude, f the ordinary frequency. When

we assume that the arms of a person performing CPR

are orthogonal (and rigid) on the patient’s chest, then

the relative distances of his/her arms are equal to the

chest compression depth. Moreover, additionally, the

frequency of low to high to low compression depth re-

presents one compression cycle. To ﬁnd a solid CPR

frequency and compression depth, we ﬁt the time se-

ries data of upper limb distances to a sinusoidal curve.

On a ﬁtted function, parameter f is the CPR fre-

quency and A is the compression depth.

To ﬁt the function y(t), we minimize the least-

squares distances using an evolutionary approach. A

common evolutionary algorithm is Differential Evo-

lution that works particularly well with nonlinear i.e.

sinusoidal cost functions. DE searches and evaluates

a parameter space in parallel and ﬁnds multiple near-

optimal but distinct solutions to a problem.

As all evolutionary algorithms, DE is population-

based and optimizes the population throughout se-

veral generations:

i,G

with i = 1..NP, G = 1..G

max

i,G

is a 4-dimensional vector of individual i for ge-

neration G. So in every generation NP individuals are

optimized up to G max generations.

The optimization is done between a transition

from one generation G to another generation G + 1.

Most evolutionary algorithms - as does DE - comprise

the steps mutation, crossover and selection.

3.2.1 Mutation

For every generation a mutation step is performed for

every individual x

i,G

. We used the step from (Storn

and Price, 1997) with a ﬁxed ampliﬁcation factor:

i,G+1

= x

+ 0.8 · (x

2,G

− x

3,G

)

with v the mutated individual and r

, r

∈

{1, 2, . . . , NP}, r

6= r

6= i randomly chosen.

3.2.2 Crossover

The crossover step decides which of the four parame-

ters of one individual are preserved in the next genera-

tion. For every parameter a uniform random number

r ∈ [0, 1] is chosen. If r ≤ CR = 0.9 then the parame-

ter from the mutant is chosen, otherwise the one from

the original individual.

3.2.3 Selection

The selection step decides which individual is pas-

sed to next generation by evaluating it against the cost

function. In our approach the squared errors are sum-

med up for every solution candidate x

i,G

∑

t=0

(s(t) − y

i,G

(t))

with s being a T -length vector of samples (joint-

ﬂoor-distances) and y the parameterized sinusoid

function of individual x

i,G

4 IMPLEMENTATION

We implemented the Differential Evolution algorithm

and the visualization software using C# on a Win-

dows system. Figure 2 shows a typical screenshot

of the application while processing motion data from

the Kinect sensor. The RGB image is shown on the

Figure 2: Our software processing Kinect motion data from a CPR training session. The graph at the bottom shows the

distances between the ﬂoor and the limbs (red) and the ﬁtted sinusoid curve (blue).

screen together with a green line overlay represen-

ting the arms. Additionally detected skeleton joints

are marked as small yellow dots.

The Kinect sensor API notiﬁes our application

when a new data frame is available from the sensor

and provides the application with a Skeleton frame

containing the joint positions and the ﬂoor plane es-

timation. The application the calculates the distance

between the ﬂoor plane and the upper limb joints (i.e.,

wrists, elbows, or shoulders) and pass the result to

the sample set. On every ten new samples the last

150 samples are being evaluated against the current

ﬁtted sine curve, and the squared errors are calcu-

lated. As soon as the error reaches a threshold, the

Differential Evolution algorithm is started with a new

updated sample set using NP = 500 individuals and

G max = 300 generations. The application continu-

ously draws the current ﬁtting in the graph on the

bottom (see Figure 2).

5 EVALUATION

5.1 Methods

We use the Laerdal Resusci Anne Simulator man-

nequin as the reference for our system. Electro-

mechanical sensors measure the depth of thorax com-

pression and decompression, the frequency of the

compressions and the volume of ventilation. On a ta-

blet, you can see all values in real-time and a sum-

mary of the training.

A Resusci Anne mannequin was placed on the

ﬂoor. The Kinect v2 sensor was placed at a distance

of 1,5m at the height of 1m. The person performing

the CPR was placed on the other side of the manne-

quin facing the Kinect camera. The subject was as-

ked to perform CPR compressions on the mannequin

with standard CPR frequency and depth for about 90

seconds. The Kinect, as well as the Resusci Anne,

were collecting data which was synchronized manu-

ally after the recording using the RGB color image of

the Kinect. The DE optimization algorithm requires a

few compressions to adapt, so the ﬁrst compressions

were omitted until the algorithm’s error drops below

a reasonable threshold (about 1% of the initial error).

5.2 Evaluation Results

5.2.1 Frequency Prediction

We compare the compressions per minute variable

of the Resusci Anne mannequin with the prediction

of our system. The Resusci Anne provides a comp-

MeanFrequency variable for every recorded compres-

sion (compEvent). The results for all four trials can

be found in Table 1 and graphically in Figure 3 for

Trial 1.

5.2.2 Depth Prediction

For every recorded compression the Resusci Anne re-

cords a maximum compression depth. We compare

these maximum compression depth value with the

doubled amplitude (2 · A) of the corresponding ﬁtted

sinusoidal curve. The results for all four trials can

be found in Table 2 and graphically in Figure 4 for

Trial 4.

Table 1: Absolute and relative mean frequency variation (in bpm) of our method compared to Resusci Anne mannequin.

Feature Hands Feature Elbows Feature Shoulders

Trial 1 / Subject 1 9.1 (*) bpm / 7.7% 3.3 bpm / 2.8% 5.1 bpm / 4.3%

Trial 2 / Subject 2 14.1 (*) bpm / 12.3% 3.4 bpm / 3.0% 6.0 bpm / 5.1%

Trial 3 / Subject 3 5.1 bpm / 4.6% 4.9 bpm / 4.5% 4.9 bpm / 4.5%

Trial 4 / Subject 3 6.4 bpm / 6.2% 6.1 bpm / 5.9% 5.7 bpm / 5.5%

Ø 8.7 bpm / 7.7% 4.4 bpm / 4.1% 5.4 bpm / 4.9%

Table 2: Absolute and relative mean compression depth variance (in mm) of our method compared to Resusci Anne Manne-

quin.

Feature Hands Feature Elbows Feature Shoulders

Trial 1 / Subject 1 25mm / 44% 10mm / 18% 18mm / 31%

Trial 2 / Subject 2 19mm / 33% 17mm / 31% 19mm / 34%

Trial 3 / Subject 3 5mm / 15% 39mm / 110% 28mm / 80%

Trial 4 / Subject 3 15mm / 36% 16mm / 41% 6mm / 15%

Ø 16mm / 32% 21mm / 50% 18mm / 40%

112

116

120

124

105000 110000 115000 120000 125000 130000

Kinect Timestamp

Compressions per minute (bpm)

colour

Kinect v2

Resusci Anne

Figure 3: Comparison of detected compression rates of our system and Resusci Anne mannequin (Subject 1, Trial 1).

500 1000 1500

Timestamp

Compression depth (in mm)

colour

Kinect v2

Resusci Anne

Figure 4: Comparison of compression depths (in mm) between our system and Resusci Anne (Subject 3, Trial 4).

6 LESSON LEARNED

We presented an interactive software system that uses

motion data from an RGB-D (Kinect) sensor and the

Differential Evolution optimization algorithm to dyn-

amically ﬁt sinusoidal curves to derive frequency and

depth parameters for cardiopulmonary resuscitation

training. We evaluated the system with three diffe-

rent subjects and tested the data of three different limb

regions (hands, elbows, and shoulders) for their suit-

ability to derive the parameters. Using the elbow fe-

atures of the Kinect skeleton, the results for the fre-

quency determination show a mean variance of 4.1%

(4.4 bpm) compared to the results of a Laerdal Re-

susci Mannequin (acting as the gold standard). This

is superior to the shoulder features with 4.9%, which

probably suffer from small dampening effects of the

elbow joint. The hand features are often occluded

even under near-optimal view situations and show a

higher variance of 7.7% variance. We will investigate

if the combined consideration of all three limb regions

will enhance the sensitivity furthermore.

While the results determining the CPR frequency

are already promising, the results for the compression

depth are not yet conclusive. The discrepancy to the

Resusci Anne results ranged between 5-39mm (15-

115%) with the hands features working best (mean

16mm). The high variance between the results made

it impossible to ﬁnd a stable constant offset or factor

that may improve the accuracy. Here, more investi-

gations are needed, to determine why only singular

results are so far promising (see Figure 4).

Thus, our work represents an initial step towards a

more complete and precise modeling of the ERC CPR

using ambient sensors.

ACKNOWLEDGEMENTS

This work was supported by the funding initiative

Nieders

achsisches Vorab of the Volkswagen Founda-

tion and the Ministry of Science and Culture of Lo-

wer Saxony as a part of the Interdisciplinary Research

Centre on Critical Systems Engineering for Socio-

Technical Systems II.

The authors would like to thank the anonymous

reviewers for their helpful comments.

REFERENCES

Higashi, E., Fukagawa, K., Kasimura, R., Kanamori, Y.,

Minazuki, A., and Hayashi, H. (2017). Development

and evaluation of a corrective feedback system using

augmented reality for the high-quality cardiopulmo-

nary resuscitation training.

Loconsole, C., Frisoli, A., Semeraro, F., Stroppa, F., Ma-

stronicola, N., Filippeschi, A., and Marchetti, L.

(2016). Relive: a markerless assistant for cpr trai-

ning. IEEE Transactions on Human-Machine Sys-

tems, 46(5):755–760.

Perkins, G., Handley, A., Koster, R., Castren, M., Smyth,

M., Olasveengen, T., Monsieurs, K., Raffay, V., Gras-

ner, J., Wenzel, V., et al. (2015). European resuscita-

tion council guidelines for resuscitation 2015. Resus-

citation, 95:81–99.

Schmidt, R. F., Lang, F., and Heckmann, F. (2011). Physio-

logie des Menschen. Springer Verlag.

Semeraro, F., Frisoli, A., Loconsole, C., Bann

o, F., Tam-

maro, G., Imbriaco, G., Marchetti, L., and Cerchiari,

E. L. (2013). Motion detection technology as a tool

for cardiopulmonary resuscitation (CPR) quality trai-

ning: A randomised crossover mannequin pilot study.

Resuscitation, 84(4):501–507.

Semeraro, F., Frisoli, A., Loconsole, C., Mastronicola, N.,

Stroppa, F., Ristagno, G., Scapigliati, A., Marchetti,

L., and Cerchiari, E. (2017). Kids (learn how to) save

lives in the school with the serious game Relive. Re-

suscitation, 116:27–32.

Soar, J., Nolan, J. P., B

ottiger, B. W., Perkins, G. D., Lott,

C., Carli, P., Pellis, T., Sandroni, C., Skrifvars, M. B.,

Smith, G. B., et al. (2015). European resuscitation

council guidelines for resuscitation 2015. Resuscita-

tion, 95:100–147.

Storn, R. and Price, K. (1997). Differential Evolution A

Simple and Efﬁcient Heuristic for global Optimization

over Continuous Spaces. Journal of Global Optimiza-

tion, 11(4):341–359.

Tian, Y., Raghuraman, S., Yang, Y., Guo, X., and Prabha-

karan, B. (2014). 3d immersive cardiopulmonary re-

suscitation (cpr) trainer. In Proceedings of the 22nd

ACM international conference on Multimedia, pages

749–750. ACM.

Wang, J.-C., Tsai, S.-H., Chen, Y.-H., Chen, Y.-L., Chu,

S.-J., and Liao, W.-I. (2017). Kinect-based real-time

audiovisual feedback device improves cardiopulmo-

nary resuscitation quality of lower-body-weight res-

cuers. The American Journal of Emergency Medicine.