Towards Simplifying Assessment of Athletes Physical Fitness:

Evaluation of the Total Physical Performance by Means of Machine

Learning

Vladimir Kublanov

, Anton Dolganov

, Viktoriya Badtieva

2,3

and David Akopyan

Research Medical and Biological Engineering Centre of High Technologies, Ural Federal University,

Mira 19, 620002, Yekaterinburg, Russian Federation

Moscow Research Center of Medical Rehabilitation and Sports Medicine,

Zemlyanoi val 53, 105120, Moscow, Russian Federation

Sechenov University, Trubeczkaya 8, 119991, Moscow, Russian Federation

Keywords: Physical Fitness, Machine Learning, Stabilography, Heart Rate Variability, Genetic Programming.

Abstract: The paper describes the methodology for the evaluation of the total physical performance of athletes on the

basis of simultaneously recorded signals of stabilography and heart rate variability. An objective assessment

of the level of physical performance was carried out using testing on the bicycle ergometer. The use of

genetic programming and linear discriminant analysis allowed obtaining the set of diagnostically significant

features. The set of diagnostically significant features is able to determine the level of physical fitness using

only data from stabilographic studies and heart rate variability with an accuracy of at least 97%. Strength

and weaknesses of the proposed approach are discussed.

1 INTRODUCTION

Currently, in various areas of human activity,

machine learning is increasingly being used in the

most varied application of this term: classification,

clustering, regression, etc (Pedregosa et al., 2011).

Applied tasks are also found in various areas, and

their number is constantly growing. For example,

medical diagnostics, where patients act as objects of

classification. From various patient data, a number

of features characterizing this patient are formed. By

analyzing these features, it is possible to solve the

following tasks: to classify the type of disease, to

determine the most appropriate method of treatment,

to find syndromes, and so on (Pombo et al., 2015,

López-Martínez et al., 2018). Specially selected

functional tests are used to assess the functional state

of the cardiovascular, nervous and neuromuscular

systems of a person, as a result allowing one to study

the patient's condition in detail.

It is known that the autonomic nervous system

(ANS) plays an exceptional role in the organization

of the organism functioning (Cardinali 2017).

Therefore, the application of the study of heart rate

variability (HRV) – an indirect method of ANS

evaluation – is promising when solving the above-

mentioned tasks. The HRV features are good

indicators of not only changes in the state of the

ANS, but also the ability of a person to adapt to

environmental changes (Aubert et al., 2012).

The use of stabilometric studies is less common.

However, this is not a consequence of the

insufficient information on this method: the reason is

the insufficient knowledge of the mechanisms of

control of the vertical posture of a person (McArdle

et al., 2010).

Regular exercise and physical activity develop

motor skills, increase strength, endurance, agility,

etc. Depending on the type of sport a person

competes, its body and skeleton are formed and a

certain level of movement is established. The

functional development of the motor apparatus and

its regulatory centers are directly related to the

equilibrium system, the reliability of which

predetermines the success of the training and sports

performance (Tucker and Collins, 2012).

In this paper, a description of the methodology

for studying the physical performance of volunteers

by means of simultaneously recorded signals of

stabilography and HRV is given. The possibility of

classifying volunteers with different levels of

physical fitness from these data is considered.

Kublanov, V., Dolganov, A., Badtieva, V. and Akopyan, D.

Towards Simplifying Assessment of Athletes Physical Fitness: Evaluation of the Total Physical Performance by Means of Machine Learning.

DOI: 10.5220/0007699105390544

In Proceedings of the 12th International Joint Conference on Biomedical Engineering Systems and Technologies (BIOSTEC 2019), pages 539-544

ISBN: 978-989-758-353-7

539

2 MATERIALS AND METHODS

2.1 Testing on a Bicycle Ergometer

with Maximum, Stepwise

Increasing Load and Gas Analysis

To determine physical performance, direct methods

were used, which are usually applied in the

examination of highly trained people, and based on

direct determination of the amount of work that the

subject can perform under various loads (Noreen et

al., 2010). With all these loads, the subject reaches

the maximum level of oxygen consumption. The

obtained indicators are objective criteria for physical

fitness of a person in those activities, the results of

which are closely related to aerobic performance

(cyclic endurance sports, etc.).

In this study used the maximum load test with

gas analysis (ergospirometry) and registration of the

electrical activity of the heart in real time (including

the first 5 minutes of recovery).

The test in the “step load increase” test was

carried out using an Oxycon ergospiometric

installation from Jaeger (Germany). Before testing,

gas analyzers were calibrated using a gas mixture

with standard O2 and CO2 concentrations, and

volumetric calibration of the instrument used was

carried out. The load was carried out on the Monark

Ergomedik 839E bicycle.

The main parameter used to determine the level

of overall physical performance was - Wmax Wt/kg

- the maximum power of the performed load, in

terms of a kilogram of body weight.

Conversion of power to the level of total physical

performance was determined in accordance with

Table 1

Table 1: Total physical performance.

Wmax, Wt/kg Male Female

1,0-1,5 - Low

1,5-2,0 Low Below average

2,0-2,5 Below average Average

2,5-3,0 Average Above average

3,0-3,5 Above average High

3,5-4,0 High -

2.2 Athletes Data Description

The study was conducted in Moscow Research

Center of Medical Rehabilitation and Sports

Medicine.

The study involved 38 athletes involved in

various sports, including team sports, wrestling, and

athletics. Of these, 28 athletes were male, 10 -

female. Of the 38 athletes, 3 had the title of master

of sports (equates to international champion), 10

were candidates for the master of sport (equates to

nationally ranked player), 6 people had the 1st junior

category (equates to regional champion), and 1

person had 2nd junior category (equates to state

champion). The average age was 20.34±3.52 years.

According to the results of testing on a bicycle

ergometer, the initial sample of athletes was divided

as follows:

• 11 with a high level of total physical

performance;

• 13 with a total level of physical performance

above average;

• 11 with an average total level of physical

performance;

• 3 with a total level of physical performance

below average.

2.3 Stabilometric Studies

In addition to the ergometer application in the study

the stabilometric studies were applied, which were

effective in classification of physical training in

earlier pilot studies (Dolganov et al., 2017). A

stabilizer analyzer computerized platform with

biological feedback "Stabilan-01-2" (Taganrog) was

used.

The sequence diagram of the research includes

three stages, in each of which the subject stood on

the platform of a stabilo-analyzer:

• functional rest (Stage 1);

• research using the “Target” load test

implemented by hardware and software of the

Stabilan-01-2 stabilo-analyzer (Stage 2);

• aftereffect (Stage 3).

The time of each stage was 5 minutes. After carrying

out the three stages of the study, the stabilography

signals and the HRV signal were saved and exported

in text format.

Athletes underwent stabilometric studies in one

day with testing on a bicycle ergometer, no later

than half hour after the testing.

2.4 Biomedical Signals

2.4.1 Stabilometric Features

Common stabilometric studies involve two signals

in the frontal (X) and sagittal (Y) planes, which

tracks position for the subject center of mass (CM).

In addition, stabilographic data in the polar

coordinate system were considered. The amplitude R

RAIDERS 2019 - Special Session on Real-world Assessment of Individuals During Everyday Routines

540

of the stabilographic signal in this case was

calculated as:



√









where X is the value of the position of the CM on the

frontal and Y is on the sagittal, planes.

The evaluation of X, Y and R signals was

performed using statistical, spectral (Fourier and

wavelet), as well as non-linear features. Overall, for

each signal 38 features were evaluated. In addition,

X and Y data was used for evaluation of 15 joint

features – to be referred as XY data (Dolganov et al.

2017).

2.4.2 Heart Rate Variability Features

That list of 64 HRV features in this study, was used

previously in prediction of the arterial hypertension

(Kublanov et al., 2017). It included time-domain and

frequency-domain features established by the

European Society of Cardiology (Malik, 1996,

Tarvainen et al., 2014), list of significant non-linear

features (Sivanantham and Devi, 2014) as well as in-

house wavelet transform features (Egorova et al.,

2014).

2.5 Machine Learning

At this stage, it was decided to solve the problem of

multi-class classification. The total level of physical

performance was used as class labels. For even

distribution between the classes, groups of athletes

with “average” and “below average” were

combined. As a method of machine learning, linear

discriminant analysis was used, which is robust in

calculation and relatively simple in interpretation

(Cacoullos, 2014).

The formation of diagnostically significant

features sets occurred in accordance with the

previously proposed genetic algorithm (Kublanov et

al., 2017), which proved itself in creating the

decision support system for a physician in the

diagnosis of arterial hypertension (Dolganov and

Kublanov, 2018).

The main points to determine when applying

genetic algorithms are the encoding, the initial

population, the selection criterion and the evolution

strategy.

In the evaluations, we encode features set by a

binary encoding. Each of features set "chromosome"

consists of 579 genes (3*38+15+64 features in 3

different stages). In particular, “1” in the

chromosome means that the specific feature is

“included” in the set, “0” means that the specific

feature is not included in the set.

As the initial population, it was decided to

choose 100 randomly created sets of three features.

The selection criterion is the accuracy of the

classification obtained using the leave-one-out cross-

validation (Zhang and Yang, 2015).

As a rule, the strategy of evolution is determined

by the ratio of the three main genetic operations -

copying, crossing-over and mutation. In this case, 10

best representatives of population are directly copied

to the next generation; 30 are created by randomly

crossing chromosomes of best representatives.

Finally, the 60 representatives in the next generation

are created through mutation – each of 10 best

representatives goes through six independent

mutation iterations. In this case each gene in the

chromosome can be flipped with a 5% chance.

Previous experience of applying the genetic

algorithms has shown that an optimal number of

generations is 20 for a classification tasks. For a

greater accounting of different probabilities, the

Genetic algorithm was applied 50 times total.

3 RESULTS

The Figure 1 shows the color map of the genetic

algorithms application results for all 50 evolutions.

Figure 1: Total classification accuracy, %.

Data in figure 1 shows that in 21 cases the

genetic algorithms failed to achieve good results –

total classification accuracy was less than 80%. In 6

cases the classification accuracy was over 90%.

Finally, there was a single evolution which resulted

in best (97%) total classification accuracy scores.

The set of diagnostically significant features,

which achieved highest classification score, is

presented in table 2. The set is composed of features

from all three stages and all signals.

In table 2, F1, F2, F3 correspond to the spectral

components of the stabilometric signals.

89 82 79 82 89 82 84 84 92 76

95 89 71 97 76 92 76 76 84 84

82 89 71 87 79 79 79 89 84 84

87 87 74 92 92 76 76 71 79 87

82 76 84 79 82 82 76 74 76 71

Towards Simplifying Assessment of Athletes Physical Fitness: Evaluation of the Total Physical Performance by Means of Machine Learning

541

Table 2: Diagnostically significant features set.

Stage Signal Feature

1 Y F3n

1 Y f(F2max)

1 R M

1 R En

1 XY L

1 HRV SDHF

1 HRV SD1/SD2

2 Y EnF3

2 R F3max

2 XY alpha

3 X M

3 Y f(F3max)

3 Y F3n

3 R CV

3 R f(F1max)

3 R f(F3max)

3 XY α

3 HRV AM0

3 HRV IAS

F1 - spectral power of stabilogram in the first

zone. The first zone is the zone of high-frequency

fluctuations (6 - 2) Hz and characterizes the

oscillations of the subject's CM, associated with

physiological processes, tremor, etc;

F2 is the spectral power of the stabilogram in the

second zone. The second zone is the zone of low-

frequency fluctuations, (2 - 0.2) Hz, and

characterizes the oscillations of the subject's CM

associated with the regulation of posture;

F3 - stabilogram spectral power in the third zone.

The third zone is a zone of fluctuations of a very low

frequency, (0.2 - 0.003) Hz, and characterizes

fluctuations of the subject's CM associated with

slow, often uncontrolled, postural control processes.

F3n, refers to the ratio, between spectral power

in F3 zone to the total spectral power of all three

zones. FXmax is a maximal value of the spectrum in

the zone X, while f(FXmax) is the corresponding

frequency.

M , is the mean value of the time-series. CV is a

Coefficient of Variation, which is defined as a ratio

between standard deviation and mean value.

En is a Shannon entropy of the corresponding

time-series (Shannon et al. 1993). EnF3 is a

Shannon entropy, evaluated for a wavelet-based

time-series in the F3 spectral zone.

L – is the total length of the stabilogramm,

defined by the formula

L

∑







X













Y











where N, is the number of points in the stabilometric

signals.

α– is an average direction of the oscillations in the

stabilogramm:

α3













tan



Cov



X,Y







D









D







tan



Cov



X,Y







D









D





where Cov



X,Y



– is a covariation of the X and Y;





– is a dispersion of the corresponding

component.

The HRV features in table 2 are presented by:

 SDHF, is a Standard Deviation, evaluated for a

wavelet-based time-series in the High Frequency

range (from 0.4 to 0.15 Hz);

 SD1/SD2 is a ratio taken from the Poincare plot;

 AM0, is the Mode Amplitude, number of

occurrences for a most frequent value in a time-

series.

 IAS, is an index defined by the ratio of Low

Frequency (from 0.15 to 0.04 Hz) and Very Low

Frequency (from 0.04 to 0.003 Hz) HRV spectral

components.

4 DISCUSSIONS AND

CONCLUSIONS

In this paper, a description of the methodology for

the evaluation of the total physical performance of

athletes on the basis of simultaneously recorded

signals of stabilography and HRV has been

presented. The possibility of classifying athletes

with different levels of physical fitness from these

data was considered. An objective assessment of the

level of physical performance was carried out using

testing on the bicycle ergometer.

The use of genetic programming and linear

discriminant analysis allowed to obtain the single set

of diagnostically significant features, which in trun

allowed to determine the level of physical fitness

using only data from stabilographic studies and heart

rate variability with an accuracy of at least 97%.

RAIDERS 2019 - Special Session on Real-world Assessment of Individuals During Everyday Routines

542

Among the strengths of proposed approach are

the following: instead of a heavy and expensive load

(bicycle ergometer) and rather complex equipment

(gas analysis), an alternative method has been

proposed, which is non-intrusive and requires less

effort. This opens a possibility for creating a simple,

efficient and cheap methodology.

However, at the moment the approach is not

without the weaknesses: in the present work a set of

diagnostically significant features was obtained,

which consists of 19 parameters registered in 3

functional states (background-“target”-aftereffect).

This makes it necessary to conduct a study lasting 15

minutes and requires video-feedback equipment. In

addition it takes features of two signals.

It is advisable to further analyze the data in order

to search for set of diagnostically significant features

that would contain a smaller number of parameters,

and at the same time - a smaller number of different

stages. This can be done by changing the selection

criterion, to value less features in sets. This can

results in application of proposed approach during

typical behaviors in real-world environments,

instead of controlled laboratory conditions.

Among other perspective tasks that our scientific

team is determined to solve are: search for way to

reduce signal record time, analyzing the contribution

of separate features to the decision made by the

classifier, and conduction of additional studies to

confirm the results.

ACKNOWLEDGEMENTS

The work was supported by Act 211 Government of

the Russian Federation, contract № 02.A03.21.0006.

REFERENCES

Aubert, A.E., Seps, B., and Beckers, F., 2012. Heart Rate

Variability in Athletes. Sports Medicine, 33 (12), 889–

919.

Cacoullos, T., 2014. Discriminant analysis and

applications. Academic Press.

Cardinali, D.P., 2017. Autonomic Nervous System: Basic

and Clinical Aspects. Springer.

Dolganov, A. and Kublanov, V., 2018. Towards a

Decision Support System for Disorders of the

Cardiovascular System - Diagnosing and Evaluation

of the Treatment Efficiency. In: Proceedings of the

11th International Joint Conference on Biomedical

Engineering Systems and Technologies - Volume 5:

AI4Health (BIOSTEC 2018). Presented at the

International Workshop on Artificial Intelligence for

Health, 727–733.

Dolganov, A.Y., Kublanov, V.S., Yamaliev, D.R., and

Goncharova, E.A., 2017. Classification of the physical

training level by heart rate variability and

stabilography data. In: 2017 Siberian Symposium on

Data Science and Engineering (SSDSE). Novosibirsk

Akademgorodok, Russia: IEEE, 49–54.

Egorova, D.D., Kazakov, Y.E., and Kublanov, V.S., 2014.

Principal Components Method for Heart Rate

Variability Analysis. Biomedical Engineering, 48 (1),

37–41.

Kublanov, V., Dolganov, A., and Gamboa, H., 2017.

Genetic programming application for features

selection in task of arterial hypertension classification.

In: Proceedings. Presented at the 2017 International

Multi-Conference on Engineering, Computer and

Information Sciences (SIBIRCON), Novosibirsk

Akademgorodok, Russia: IEEE, 561–565.

Kublanov, V., Dolganov, A., and Kazakov, Y., 2017.

Diagnostics of the Arterial Hypertension by Means of

the Discriminant Analysis Analysis of the Heart Rate

Variability Signals Features Combinations.

Proceedings of the 10th International Joint

Conference on Biomedical Engineering Systems and

Technologies, Vol 4: Biosignals, 291–298.

López-Martínez, F., Schwarcz, A., Núñez-Valdez, E.R.,

and García-Díaz, V., 2018. Machine Learning

Classification Analysis for a Hypertensive Population

as a Function of Several Risk Factors. Expert Systems

with Applications.

Malik, M., 1996. Heart rate variability: Standards of

measurement, physiological interpretation, and clinical

use. Circulation, 93 (5), 1043–1065.

McArdle, W.D., Katch, F.I., and Katch, V.L., 2010.

Exercise physiology: nutrition, energy, and human

performance. Lippincott Williams & Wilkins.

Noreen, E.E., Yamamoto, K., and Clair, K., 2010. The

reliability of a simulated uphill time trial using the

Velotron electronic bicycle ergometer. European

journal of applied physiology, 110 (3), 499–506.

Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V.,

Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P.,

Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,

Cournapeau, D., Brucher, M., Perrot, M., and

Duchesnay, É., 2011. Scikit-learn: Machine Learning

in Python. Journal of Machine Learning Research, 12,

2825−2830.

Pombo, N., Garcia, N., Bousson, K., and Felizardo, V.,

2015. Machine learning approaches to automated

medical decision support systems. In: Handbook of

research on artificial intelligence techniques and

algorithms. IGI Global, 183–203.

Shannon, C.E., Wyner, A.D., and Sloane, N.J., 1993.

Claude E. Shannon: Collected Papers. John Wiley &

Sons.

Sivanantham, A. and Shenbaga Devi, S., 2014. Cardiac

arrhythmia detection using linear and non-linear

features of HRV signal. In: 2014 International

Towards Simplifying Assessment of Athletes Physical Fitness: Evaluation of the Total Physical Performance by Means of Machine Learning

543

Conference on Advanced Communication Control and

Computing Technologies (ICACCCT). 795–799.

Tarvainen, M.P., Niskanen, J.-P., Lipponen, J.A., Ranta-

Aho, P.O., and Karjalainen, P.A., 2014. Kubios HRV–

heart rate variability analysis software. Computer

methods and programs in biomedicine, 113 (1), 210–

220.

Tucker, R. and Collins, M., 2012. What makes

champions? A review of the relative contribution of

genes and training to sporting success. Br J Sports

Med, 46 (8), 555–561.

Zhang, Y. and Yang, Y., 2015. Cross-validation for

selecting a model selection procedure. Journal of

Econometrics, 187 (1), 95–112.

RAIDERS 2019 - Special Session on Real-world Assessment of Individuals During Everyday Routines

544