Vertical jump height prediction using EMG characteristics and neural networks

Journal of Cognitive Systems Research 1 (2000) 135–141www.elsevier.com/ locate /cogsys

Vertical jump height prediction using EMG characteristics andneural networks

Action editor: Lee Gilesa , b*Brijesh Verma , Chris Lane

aSchool of Information Technology, Griffith University-Gold Coast, Campus PMB 50, Gold Coast Mail Center QLD 9726, AustraliabSchool of Exercise Science, Griffith University-Gold Coast, Campus PMB 50, Gold Coast Mail Center QLD 9726, Australia

Accepted 13 November 1999

Abstract

The purpose of this study was to investigate the Artificial Neural Networks (ANNs) in conjunction with EMGcharacteristics to predict vertical jump height. This paper investigates three key areas for the development of a real-timebiomechanical biofeedback system. The first area of investigation is the EMG data processing and characteristics that bestsuit the training of ANNs for prediction. The second area of interest is the best ANN topology, and the final one, whichANN topology would best predict muscle coordination for biofeedback to the coach or athlete. 2000 Elsevier ScienceB.V. All rights reserved.

Keywords: Neural networks, Electromyography, Prediction, Muscle coordination, Vertical jump

1. Introduction athlete uses not only his or her success but also thecoach’s feedback as a guide to improvement. The

The use of biomechanical analysis techniques and coach often is unable to identify the areas in whichANNs in the constructive analysis of biomechanics is the athlete did not perform correctly. The use ofrelatively new due to recent advances in technology biomechanical analysis and ANNs to give the coach(Lapham & Bartlett, 1995; McLean & Lafortune, and athlete biofeedback of the success of muscle1988). The combination of human evaluation and coordination may improve sporting performance.ANN fields has been applied to many medical The central nervous system is the major con-applications but the use of ANNs in biomechanical tributor to the ability of an individual to haveanalysis techniques is only in its infancy. muscular coordination or control. The nervous sys-

The acquisition of a sporting skill takes many tem stores information of the mechanical characteris-hours of training (McLean & Lafortune, 1988). The tics of the muscle (Chapman & Sanderson, 1990). If

this is not the case, the nervous system may havesome type of feedback to the mechanical characteris-*Corresponding author.

E-mail address: [email protected] (B. Verma) tic of the muscle activation. In most cases, individual

1389-0417/00/$ – see front matter 2000 Elsevier Science B.V. All rights reserved.PI I : S1389-0417( 00 )00005-X

136 B. Verma, C. Lane / Journal of Cognitive Systems Research 1 (2000) 135 –141

talent begins with the ability of an individual to raphy diagnosis of muscular disorders. Four differentperform a task from a given set of initial conditions. training methods were investigated using different

Muscular coordination can have two distinct pat- ANN models. It was found that a feedforward two-terns. The first can be a temporal sequence in which layer ANN with a modified back propagation algo-the coordination is a task and is ideal to repeat on rithm for training yielded the same if not highereach occasion. An example of this task would be the performance in diagnosis.vertical jump whereby the task is completed identi- This paper investigates three key areas for thecally every time. The second technique of muscular development of a real-time biomechanical biofeed-coordination is when variations of the basic temporal back system. The first area of investigation is thesequence are applied due to certain conditions. An EMG data processing and characteristics that bestexample of this technique is cycling on a flat or suit the training of ANNs for prediction. The secondhorizontal plane compared to a hill or vertical climb. area of interest is the best ANN topology, andAlthough there may be variations of either technique finally, which ANN topology would best predictthe overall conclusion is that an ideal sequence of muscle coordination for biofeedback to the coach ormuscular coordination will allow the individual to athlete.perform a task to maximal performance. The ability The remainder of this paper is organised asto analyse the muscular coordination of the indi- follows: in Section 2, we describe the proposedvidual may allow the coach or athlete to get biofeed- technique. Some experimental results are shown inback which in turn will aid the individual’s nervous Section 3. An analysis and discussion of the resultssystem in the control of the muscle action. A number are presented in Section 4. Finally, Section 5 con-of techniques can be used to analyse muscular cludes the paper.coordination (Basmajian et al., 1975). One suchmethod is the direct measurement of muscularactivation through electromyography (Chapman & 2. Proposed techniqueSanderson, 1990; Clarys & Cabri, 1993).

As a whole, the area of EMG analysis (Basmajian 2.1. Input data& De Luca, 1985; Basmajian, 1973; Clarys, 1987)examines three main aspects of both clinical and 2.1.1. Testing equipmentkinesiological fields. The ability of the muscle to The tests were conducted using a combination ofswitch on/off at a particular time, the amplitude of three separate systems (Fig. 1). The first system wasthe contraction and the fatigue of the muscle under the EMG telemetry system, which obtained EMGcontraction. In the analysis of muscle coordination signals and communicated these signals to computeron/off timing and extent of muscle contraction number one. The second system consisted of a kistlerbecome the significant independent variables for force plate and amplifiers connected to a peakperformance (Enoka, 1988). motion analysis system using computer two. The

The amplitude and pattern of muscle activity is final system needed the development of a Lab Viewcritical in all areas of muscle movements, but it is of data acquisition program. The software program usedforemost importance in complex musculoskeletal a National Instruments data acquisition card tomovements. Research has shown that it is possible sample the EMG signal and synchronise this with theusing ANN analysis to distinguish between subjects peak motion analysis system. The software programwith chronic back pain and control subjects with a sampled the EMG signal at 1000 Hz. The EMGsuccess rate in the order of 75% (Oliver & Atsma,1996). It was also discovered that ANNs coulddistinguish two patterns of muscular activation toperform the same task (Nussbaum & Chaffin, 1997;Nussbaum et al., 1997).

Abel et al. (1996) examined the use of differentANN types and training methods in electromyog- Fig. 1. Testing equipment.

B. Verma, C. Lane / Journal of Cognitive Systems Research 1 (2000) 135 –141 137

signal was acquired after a trigger was set when aforce of 0 N and 260 N was applied to the forceplate. A pre-trigger sample of 400 was recorded witha post-trigger of 600 samples.

Fig. 3. The software program structure.2.1.2. Testing protocolThe assessment of the best paradigms and training

algorithms required the need to process 60 squat amplitude of each muscle. The data processing andvertical jumps of a subject. A male subject between EMG parameter extraction are shown in Fig. 3.the ages of 20 and 30 performed a warming-up of anumber of squat vertical jumps. The subject was 2.1.4. Desired outputinstructed to keep his arms akimbo and to perform The reaction forces of the vertical jump were10 maximum squat jumps to obtain maximum vol- calculated using software from a peak motion analy-untary contraction (MVC) data value for normalisa- sis system. The data were exported to a text file to betion of each muscle measured. The EMG data used by another program. A Lab View program wascollection and processing required the analysis of developed to calculate the impulse of each jumpeight major muscle groups of the lower limb of a leg allowing for calculation of vertical jump height inshown in Fig. 2. meters.

2.1.3. Data processing 2.2. Artificial neural networkThe 1000 EMG data points acquired with the

software program which removed any offset or drift The biological brain is a multitude of biologicalthat occurred (Oliver & Atsma, 1996; Winter, 1991). neurons, similarly an ANN as shown in Fig. 4 is aThe EMG data were rectified and passed through a combination of numerous artificial neurons, eachbutterworth low pass digital filter with a cutoff having a small amount of local memory. The artifi-frequency of 5 Hz (Oliver, 1995). The filtered EMG cial neurons are connected by unidirectional com-data were then normalised between the range of zero munication channel ‘connections,’ which carryto one with MVC representing a value of one numeric as opposed to symbolic data. The units(Winter, 1991; Yang & Winter, 1984). operate only on their local data and on the inputs

The maximum amplitude of each muscle contrac- they receive via the connections (Anderson & Rosen-tion was extracted and placed into two separate data feld, 1990; Wasserman, 1989).files. The time of muscle activation was calculated All ANNs in this study were multi-layered feedwhen each muscle reaches 20% of MVC. In the first forward neural networks. It can be said that an ANNsituation the trigger impulse force of 260 N was 0.4 has three distinct components. The input layer re-of a second into the jump. It was then said that event ceives the stimuli from the external environment.time 0 seconds occurred 0.4 seconds prior to the The output layer which interfaces the ANN back toimpulse force. In the second situation the data the external environment and as shown in Fig. 4 thecollected were re-synchronised so that 0 N was 0.1 hidden layer that can consist of a multitude of layersof a second into the jump. It was then calculated that (Zahedi, 1993). The hidden layer can be consideredevent time 0 seconds occurred 0.1 seconds prior to the computation area of the ANN (Wasserman,the impulse force. Each time value was appended to 1989). The amount of hidden layers that consist inthe separate data files that contained the maximum

Fig. 2. Eight major muscles of the lower limb. Fig. 4. The ANN structure.


the ANN is measured without the inclusion of the plitude ANNs. If the timing or amplitude ANNsinput layer. This is because the input layer plays no were able to classify the vertical jump height, thensignificant role in the computation and acts only as we could suggest through biofeedback to the coachan interface between the environment and the ANN. or athlete that the timing and amplitude characteris-

The timing and amplitude of the EMG signal of tics of that vertical jump were correct for theeach muscle was fed into the input layer, which had intensity of the jump.7, 8, 14 and 16 nodes. The input nodes weredependent on which muscles, and whether timing oramplitude or both, were applied. The sigmoidal 3. Experimental resultsactivation function was used in all networks. The‘hidden layer’ consisted of 1, 2, 4, 7, 8, 14 or 16 The proposed approach has been implemented innodes and the output layer consisted of 1 node. The C/C 1 1 on the SP2 Supercomputer. Once the dataANN topology can be represented as 16:4:1, indicat- sets were preprocessed they were placed in a traininging 16 input layer nodes, 4 hidden layer nodes and 1 matrix and transferred to the SP2 Supercomputer atoutput layer node. All network topologies were Griffith University in Brisbane. Some preliminarytrained using the back propagation algorithm. The experiments and promising results are shown below.training process consisted of set parameters oflearning-rate 0.4, momentum 0.5, rms-error cutoff 3.1. Timing and amplitude topology0.0005 and iterations of 200,000. The number oftraining and testing pairs varied between 50–10, Two different data sets (training and testing pairs)40–20, 45–10 and 35–20. After each test, the rms were used to train and test five different ANNerror in training was recorded and if an error cutoff topologies. The first data set was created from eightwas reached, the training was stopped and iterations muscles with amplitude and timing inputs. Each datawere recorded. Also, test result accuracy was calcu- set was synchronised to 260 N of impulse force onlated and the total accuracy on each test group the force plate. The data set comprised the adjust-recorded. If each test was above 90% accurate, it was ment of the training pairs to 50 and the testing pairsrecorded as a positive classification. A set of parame- to 10 (see Table 1).ters of less than 0.3 meter, greater than 0.35 meter The second data set comprised the adjustment ofwas set to examine good and bad jump classification. the training pairs to 40 and the testing set to 20 (seeIf the jump was between these parameters, then to Table 2). All other parameters remained the same asdistinguish a correct classification, the jump predic- the previous experiment.tion must lie within 0.02 meter. The third data set comprised the adjustment of the

Finally, a third type of ANN was implemented to training pairs to 50 and the testing pairs to 10 (seegive the coach and athlete biofeedback to the per- Table 3). In this experiment the tibalis anteriorformance of the vertical jump. The structure of the muscle amplitude and timing values and five corruptANN was a combination of three ANNs as shown in tests were removed. The synchronisation of verticalFig. 5. The two input ANNs — one each for timing jumps was then set to zero force production on theand amplitude — were trained using the same set of force plate.parameters as above. The number of hidden unitswas restricted to one. The predicted vertical height of Table 1

Training 50 pairs / testing 10 pairs synchronised to 260 N of forcethe jump was also read from the timing and am-

Inputs Hidden units RMS error Accuracy Prediction(%) (correct / total)

16 1 0.0385 79.4 6/1016 2 0.0164 82.7 5/1016 4 0.0057 81.9 4/1016 8 0.0062 82.7 6/1016 16 0.0006 75.0 7/10

Fig. 5. ANN for biofeedback.


Table 2 Table 5Training 40 pairs / testing 20 pairs synchronised to 260 N of force Amplitude ANN synchronised to 260 N of force

Inputs Hidden units RMS error Accuracy Prediction Inputs Hidden units RMS error Accuracy Prediction(%) (correct / total) (%) (correct / total)

16 1 0.0357 87.7 11/20 16 1 0.0521 89.2 8/1016 2 0.0179 79.6 12/20 16 2 0.0380 81.7 6/1016 4 0.0032 82.7 12/20 16 4 0.0241 58.8 5/1016 8 0.0006 88.6 13/20 16 8 0.0088 77.8 4/1016 16 0.0011 87.0 17/20 16 16 0.0037 77.3 4/10

Table 6Table 3Timing ANN synchronised to 260 N of forceTraining 50 pairs / testing 10 pairs synchronised to 0 N of force

Inputs Hidden units RMS error Accuracy PredictionInputs Hidden units RMS error Accuracy Prediction(%) (correct / total)(%) (correct / total)

16 1 0.0458 89.9 6/1014 1 0.0264 92.6 7/1016 2 0.0430 89.9 6/1014 2 0.0188 77.2 5/1016 4 0.0355 86.7 6/1014 4 0.0090 86.5 7/1016 8 0.0317 86.6 6/1014 7 0.0023 84.9 8/1016 16 0.0324 86.9 6/1014 14 0.0011 78.7 6/10

The fourth data set comprised the adjustment of muscles with seven timing inputs. All other parame-the training pairs to 40 and the testing set to 20 (see ters remained the same as in the previous test.Table 4). All other parameters remained the same as The fourth data test comprised the adjustment ofin the previous experiment. the training pairs to 45 and the testing set to 10 (see

Tables 7 and 8). In this test, the tibalis anterior3.2. Timing or amplitude topology muscle amplitude values and five corrupt tests were

removed. The synchronisation of vertical jumps wasTwo different data sets (timing or amplitude) were then set to zero force production on the force plate.

used to train and test four different ANN topologiesto examine the effect of network conformation on

Table 7rate and likelihood of convergence. The first data set Amplitude ANN synchronised to 0 N of forcecomprised a set of 50 training and 10 testing data

Inputs Hidden units RMS error Accuracy Predictionconsisting of eight muscles with amplitude inputs. (%) (correct / total)Each test was synchronised to 260 N of positive

14 1 0.0471 88.8 7/10force on the force plate (see Table 5).14 2 0.0355 91.6 4/10

The third data test comprised the adjustment of the 14 4 0.0292 85.0 4/10training pairs to 50 and the testing set to 10 (see 14 7 0.0102 79.9 6/10

14 14 0.0060 72.1 5/10Table 6). The training and testing data used eight

Table 4 Table 8Training 40 pairs / testing 20 pairs synchronised to 0 N of force Timing ANN synchronised to 0 N of force

Inputs Hidden units RMS error Accuracy Prediction Inputs Hidden units RMS error Accuracy Prediction(%) (correct / total) (%) (correct / total)

14 1 0.0249 88.8 15/20 14 1 0.0360 89.0 6/1014 2 0.0179 85.0 14/20 14 2 0.0325 89.3 5/1014 4 0.0032 87.5 13/20 14 4 0.0214 87.0 4/1014 7 0.0006 89.5 15/20 14 7 0.0213 87.1 4/1014 14 0.0011 79.8 11/20 14 14 0.0216 87.4 4/10


Table 9 biofeedback to the coach or athlete. It was possibleFinal ANN trained for classification to train two ANNs, one for timing and the other forANNs Hidden unit RMS error Accuracy Prediction amplitude, to predict vertical jump height. The final

(correct / total) ANN was then trained successfully to classify theFinal ANN 1 0.0361 90.8% 8/10 vertical jump. In this structure, it was possible to

feed off the outputs from both the timing andamplitude ANN. This made it possible to compare

3.3. Combination topology the actual vertical jump height with the predictedvalues. If the vertical jump height was classified by

The timing and amplitude ANNs were trained with the timing ANN then it was considered that the35 training pairs and an extra 10 pairs added to these vertical jump muscle timing patterns were charac-for testing (see Table 9). The outputs of the test set teristic of that jump height. This procedure alsowere then used to train the third ANN for classifica- applies to the amplitude ANN.tion of good or bad jumps. The ANNs were then In the future more training data will be used totested using a final data set of 10 pairs. improve the prediction rate. The other bootstrap

neural network techniques (Efron, 1979; Efron &Tibshirani, 1991, 1993) will also be employed which

4. Analysis and discussion of the results might improve the results. One limitation of theproposed technique is that the vertical jump height

In Tables 1–4, the results for timing and am- must be known to compare the predicted values ofplitude topology using only one neural network are the timing and amplitude ANNs.presented. When we used different hidden units andtraining and testing pairs, it was found that theprediction rate was increased when number of hidden

5. Conclusionunits were increased; however with 0 N force andremoved corrupt data set, over 92% prediction rate

We have investigated and presented a neuralwas achieved using only one hidden unit. In Tables

network system to predict vertical jump height. The5–8, the results for timing or amplitude topology are

results obtained are very encouraging and the pro-presented. As shown in Tables 5–8, a single hidden

posed system can be used in real-world jump predic-layer neural network with only timing or only

tion applications. It was possible to train an ANN toamplitude produces the worst results compared to

predict vertical jump height using certain EMGtiming and amplitude.

characteristics. It was also possible to implement aAs it is shown above that ANNs can be trained to

combination of ANNs to give biofeedback to thepredict the vertical jump height reasonably well. It

coach or athlete of muscle timing and amplitude. Thewas found that the combination of timing and

influence of biofeedback to the coach and athlete isamplitude vertical jump characteristics provided the

one area for further investigation. Also the trainingbest result of 92.6% with single hidden unit. The

of the ANN for multiple individuals and how theresults also reflected which EMG characteristics

ANN predicts the vertical jump height of an in-provided the best results. These EMG characteristics

dividual that it has not been trained with is anotherare as follows: seven muscles not including the

area for future research. The bootstrap neural net-tibalis anterior muscle, a low pass filter of 5 Hz, a

work techniques will also be employed in the future,synchronisation or trigger of 0 N and removal of any

which might improve the results.corrupt test recordings in training of the ANN.

In Table 9, the results using combined topology asshown in Fig. 5 are presented. The combined topolo-gy produced over 80% prediction rate, which is quite Referencesgood on such small training data. This experimentprovided an insight into the ability of ANNs to give Abel, E. W., Zacharia, P. C., Forster, A., & Farrow, T. L. (1996).


Neural network analysis of the EMG interference pattern. Lapham, A. C., & Bartlett, R. M. (1995). The use of artificialMedical Engineering & Physics 18(1), 12–17. intelligence in the analysis of sports performance: a review of

Anderson, J. A., & Rosenfeld, E. (1990). Neurocomputing: applications in human gait analysis and future directions forfoundations of research, MIT Press, Cambridge. sports biomechanics. Journal of Sports Science 13, 229–237.

Basmajian, J. V. (1973). Electrodes and electrode connectors. In: McLean, B, & Lafortune, M. (1988). Improving pedaling tech-Desmedt, J. (Ed.), New developments in EMG and clinical nique with ‘real time’ biomechanical feedback. Excel 5(1),neurophysiology, Karger, Basel, pp. 498–501. 15–18.

Basmajian, J. V., & De Luca, C. J. (1985). Muscles alive: their Nussbaum, M. A., & Chaffin, D. B. (1997). Pattern classificationfunctions revealed by electromyography, Butterworths, Balti- reveals inter-subject group differences in lumbar muscle recruit-more. ment during static loading. Clinical Biomechanics 12(2), 97–

Basmajian, J. V., Clifford, H. C., McLeod, W. D., & Nunnally, H. 106.N. (1975). Computers in electromyography, Butterworths, Nussbaum, M. A., Martin, B. J., & Chaffin, D. B. (1997). ABaltimore. neural network model for simulation of torso muscle coordina-

Chapman, A. E., & Sanderson, D. J. (1990). Muscular coordination. Journal of Biomechanics 30(3), 251–258.tion in sporting skills. Multiple muscle systems, Springer, New Oliver, C. W. (1995). Artificial intelligence in the detection of lowYork. back pain. Journal of Orthopaedic Rheumatology 8, 207–210.

Clarys, J. P. (1987). Application of EMG for the evaluation of Oliver, C. W., & Atsma, W. J. (1996). Artificial intelligenceperformance in different sports. In: Marconnet, P., & Komi, P.V. analysis of paraspinal power spectra. Cinical Biomechanics(Eds.), Muscular function in exercise and training, Karger, 11(7), 422–424.Basel, pp. 200–233. Wasserman, P. D. (1989). Neural computing: theory and practice,

Clarys, J. P., & Cabri, J. (1993). Electromyography and the study Van Nostrand Reinhold, New York.of sports movements: a review. Journal of Sports Science 11, Winter, D. A. (1991). Electromyogram recording, processing, and379–448. normalization: procedures and considerations. Journal of

Efron, B. (1979). Bootstrap methods: another look at the jack- Human Muscle Performance 1(2), 5–15.knife. Annals of Statistics 7, 1–26. Yang, J., & Winter, D. A. (1984). Electromyographic amplitude

Efron, B., & Tibshirani, R. (1991). Statistical data analysis in the normalization methods: improving their sensitivity as diagnosticcomputer age. Science 253, 390–395. tools in gait analysis. Archives of Physical Medicine and

Efron, B., & Tibshirani, R. (1993). An introduction to the Rehabilitation 65, 517–521.bootstrap, Chapman & Hall, New York. Zahedi, F. (1993). Intelligent systems for business, expert systems

Enoka, R. M. (1988). Neuromechanical basis of kinesiology, with neural networks, Wadsworth, Boston.Human Kinetics Books, Champaign, IL.

Vertical jump height prediction using EMG characteristics and neural networks

Documents

Transcript of Vertical jump height prediction using EMG characteristics and neural networks