Short communication Classification accuracy of a single tri-axial accelerometer for training background and experience level in runners Dylan Kobsar n , Sean T. Osis, Blayne A. Hettinga, Reed Ferber Faculty of Kinesiology, University of Calgary, Canada article info Article history: Accepted 11 April 2014 Keywords: Running Principal component analysis Accelerometer Gait abstract Accelerometers are increasingly used tools for gait analysis, but there remains a lack of research on their application to running and their ability to classify running patterns. The purpose of this study was to conduct an exploratory examination into the capability of a tri-axial accelerometer to classify runners of different training backgrounds and experience levels, according to their 3-dimensional (3D) acceler- ometer data patterns. Training background was examined with 14 competitive soccer players and 12 experienced marathon runners, and experience level was examined with 16 first-time and the same 12 experienced marathon runners. Discrete variables were extracted from 3D accelerations during a short run using root mean square, wavelet transformation, and autocorrelation procedures. A principal component analysis (PCA) was conducted on all variables, including gait speed to account for covariance. Eight PCs were retained, explaining 88% of the variance in the data. A stepwise discriminant analysis of PCs was used to determine the binary classification accuracy for training background and experience level, with and without the PC of Speed. With Speed, the accelerometer correctly classified 96% of runners for both training background and experience level. Without Speed, the accelerometer correctly classified 85% of runners based on training background, but only 68% based on experience level. These findings suggest that the accelerometer is effective in classifying athletes of different training back- grounds, but is less effective for classifying runners of different experience levels where gait speed is the primary discriminator. & 2014 Elsevier Ltd. All rights reserved. 1. Introduction Recently, there has been an exponential increase in the availability of cost-effective accelerometers for human movement analysis. These devices collect three-dimensional trunk or limb accelerations without the space, time, and monetary requirements of conventional gait analysis systems and the improved sensitivity and resolution of modern-day accelerometers has promoted their use in activity classi- fication and gait analysis (Chen et al., 2012; Tao et al., 2012). Previous studies have demonstrated the ability to use an accelerometer to classify different activities such as walking, running, jumping, and stair climbing with 70–99% accuracy (Altun et al., 2010; Godfrey et al., 2011; Preece et al., 2009). Additionally, a single accelerometer has shown sufficient sensitivity to measure age-related changes in gait (Kobsar et al., 2014), as well as discriminate pathological gait patterns such as Complex Regional Pain Syndrome (Yang et al., 2012) and Huntington's disease patients (Dalton et al., 2013) with 99–100% accuracy. These findings suggest that not only can accelerometers be highly effective tools for activity classification, but they also have the ability to classify individuals based on their gait patterns. Applying accelerometry technology to the movement patterns of runners to examine experience level or training background could provide signi ficant benefit to the training and performance of runners. For example, it is thought that runners should minimize movement at the centre of mass in order to maximize efficiency (Saunders et al., 2004). McGregor et al. (2009) supported this concept and reported that based on experience levels, trained runners exhibited lower magni- tudes of accelerations at the centre of mass as compared to untrained runners. In addition to experience level, these differences may exist across training backgrounds where maximizing efficiency is not the primary objective (e.g., soccer players vs. marathon runners). A single accelerometer placed on the lower back provides a simple and effective method to examine this movement near the centre of mass. This information could be used to model or classify a runner's movement patterns across experience levels or athletic backgrounds. Unfortu- nately, we know of no research that has attempted to classify runners based on training background or experience using accelerometry. Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/jbiomech www.JBiomech.com Journal of Biomechanics http://dx.doi.org/10.1016/j.jbiomech.2014.04.017 0021-9290/& 2014 Elsevier Ltd. All rights reserved. n Corresponding author at: Faculty of Kinesiology, University of Calgary, Calgary, AB, T2N 1N4, Canada. Tel.: þ1 306 681 7110. E-mail address: [email protected](D. Kobsar). Please cite this article as: Kobsar, D., et al., Classification accuracy of a single tri-axial accelerometer for training background and experience level in runners. Journal of Biomechanics (2014), http://dx.doi.org/10.1016/j.jbiomech.2014.04.017i Journal of Biomechanics ∎ (∎∎∎∎) ∎∎∎–∎∎∎
Transcript
Short communication
Classification accuracy of a single tri-axial accelerometer for trainingbackground and experience level in runners
Dylan Kobsar n, Sean T. Osis, Blayne A. Hettinga, Reed FerberFaculty of Kinesiology, University of Calgary, Canada
Accelerometers are increasingly used tools for gait analysis, but there remains a lack of research on theirapplication to running and their ability to classify running patterns. The purpose of this study was toconduct an exploratory examination into the capability of a tri-axial accelerometer to classify runners ofdifferent training backgrounds and experience levels, according to their 3-dimensional (3D) acceler-ometer data patterns. Training background was examined with 14 competitive soccer players and12 experienced marathon runners, and experience level was examined with 16 first-time and the same12 experienced marathon runners. Discrete variables were extracted from 3D accelerations during ashort run using root mean square, wavelet transformation, and autocorrelation procedures. A principalcomponent analysis (PCA) was conducted on all variables, including gait speed to account for covariance.Eight PCs were retained, explaining 88% of the variance in the data. A stepwise discriminant analysis ofPCs was used to determine the binary classification accuracy for training background and experiencelevel, with and without the PC of Speed. With Speed, the accelerometer correctly classified 96% ofrunners for both training background and experience level. Without Speed, the accelerometer correctlyclassified 85% of runners based on training background, but only 68% based on experience level. Thesefindings suggest that the accelerometer is effective in classifying athletes of different training back-grounds, but is less effective for classifying runners of different experience levels where gait speed is theprimary discriminator.
& 2014 Elsevier Ltd. All rights reserved.
1. Introduction
Recently, there has been an exponential increase in the availabilityof cost-effective accelerometers for human movement analysis. Thesedevices collect three-dimensional trunk or limb accelerations withoutthe space, time, and monetary requirements of conventional gaitanalysis systems and the improved sensitivity and resolution ofmodern-day accelerometers has promoted their use in activity classi-fication and gait analysis (Chen et al., 2012; Tao et al., 2012). Previousstudies have demonstrated the ability to use an accelerometer toclassify different activities such as walking, running, jumping, and stairclimbing with 70–99% accuracy (Altun et al., 2010; Godfrey et al., 2011;Preece et al., 2009). Additionally, a single accelerometer has shownsufficient sensitivity to measure age-related changes in gait (Kobsar etal., 2014), as well as discriminate pathological gait patterns such asComplex Regional Pain Syndrome (Yang et al., 2012) and Huntington's
disease patients (Dalton et al., 2013) with 99–100% accuracy. Thesefindings suggest that not only can accelerometers be highly effectivetools for activity classification, but they also have the ability to classifyindividuals based on their gait patterns.
Applying accelerometry technology to the movement patterns ofrunners to examine experience level or training background couldprovide significant benefit to the training and performance of runners.For example, it is thought that runners should minimize movement atthe centre of mass in order to maximize efficiency (Saunders et al.,2004). McGregor et al. (2009) supported this concept and reported thatbased on experience levels, trained runners exhibited lower magni-tudes of accelerations at the centre of mass as compared to untrainedrunners. In addition to experience level, these differences may existacross training backgrounds where maximizing efficiency is not theprimary objective (e.g., soccer players vs. marathon runners). A singleaccelerometer placed on the lower back provides a simple and effectivemethod to examine this movement near the centre of mass. Thisinformation could be used to model or classify a runner's movementpatterns across experience levels or athletic backgrounds. Unfortu-nately, we know of no research that has attempted to classify runnersbased on training background or experience using accelerometry.
Please cite this article as: Kobsar, D., et al., Classification accuracy of a single tri-axial accelerometer for training background andexperience level in runners. Journal of Biomechanics (2014), http://dx.doi.org/10.1016/j.jbiomech.2014.04.017i
Therefore, the purpose of this exploratory study was to utilize a singletri-axial accelerometer and determine whether we can classify runnersof different training backgrounds and experience levels according totheir patterns of motion.
2. Methods
2.1. Subjects
Fourteen competitive soccer players (age: 1971 yr; mass: 60.478.8 kg), 16first-time marathon runners (age: 4778 yr; mass: 70.6712.9 kg), and 12 experi-enced marathon runners (age: 30710 yr; mass: 61.678.4 kg) were selected for thestudy. First-time marathon runners (no marathon times) were recruited from abeginner marathon programme, while experienced marathon runners (averagemarathon race times: 242732 min) were recruited from local running clubs andhad completed one or more races in the past year. These groups were chosen toallow for binary classification of training background (soccer players vs. experi-enced marathon runners) and experience level (experienced marathon runners vs.first-time marathon runners). All participants were female and free of injury for atleast three months prior to testing. This research was approved by the university'sethics review and informed consent was obtained prior to testing.
2.2. Procedure
A single tri-axial accelerometer (G-Link Wireless Accelerometer Node710 g,9 mg resolution at 10 g; Microstrain Inc., VT) sampling at 617 Hz (MicroStrain, Inc.,2011) was securely mounted with semi-rigid elastic straps to the subjects' lowerback (L3 vertebra) to approximate the centre of mass (Moe-Nilssen and Helbostad,1998). Subjects ran on an instrumented treadmill (Bertec, Columbus, OH) for 2 minto acclimatize, and then 15 s of accelerometer data were collected equating toapproximately 20–25 footfalls. Given the various training backgrounds and experi-ence levels, a self-selected, rather than a pre-defined speed, was used and variancesin gait speed were factored into the analysis.
2.3. Data analysis
A discriminant analysis of principal components (PCs) was used to classifyrunners. A total of 44 variables, 43 determined from attitude-corrected 3Daccelerations (McGregor et al., 2009; Moe-Nilssen and Helbostad, 1998, 2004;Preece et al., 2009) and one, gait speed, determined from the treadmill, were usedto describe various aspects of running gait. The potential redundancy in thesevariables was reduced using a principal component analysis and a discriminantanalysis to classify runners based on training background or experience level. Alldata analysis was performed in MATLAB 7.10 (The MathWorks Inc., Natick, MA).
The root mean square (RMS; g's) of accelerations in the vertical (V), medio-lateral (ML), anteroposterior (AP), and resultant (R) direction was calculated as ameasure of the overall amplitude of accelerations for each axis. The economy ofaccelerations (ECON; g/m/s) measures how efficient a runner's movement is fortheir self-selected speed and was determined as the RMS in each axis divided bythe gait speed (McGregor et al., 2009). Therefore, a high value for ECON representsinefficient movement as a large amount of accelerations occur for a given speed.The ratio of accelerations (RATIO; unitless) is used to assess the relative
contribution of each axis during running and was computed as the RMS in eachaxis (V, ML, AP) divided by the RMSR (McGregor et al., 2009). A large RATIO value for agiven axis represents a large contribution towards the overall movement, while asmall value represents little contribution towards the overall movement.
A discrete wavelet transformation procedure was used to examine the time-frequency domain of the 3D accelerations and measure the various high and lowfrequencies that occur during the repetitive impacts of running. This procedureinvolved the decomposition of accelerometer data in each axis by iterativelyremoving high frequency detail components using a bandpass filter with apassband of fmax/2, fmax (Preece et al., 2009). Specifically, a 5-level waveletdecomposition using a Daubechies 5-mother wavelet was used to obtain five levelsof time-frequency detail bands (WAVE1-5) from the original signal, with WAVE5being of the highest frequency and WAVE1 being of the lowest frequency. Theamplitude of each of the five levels was quantified as the RMS in each axis.
The regularity and symmetry of the accelerometer waveform in each axis wasdefined using an unbiased autocorrelation procedure (Moe-Nilssen and Helbostad,2004). This procedure computes the correlation within the acceleration waveformat a phase shift equal to the average step time (step regularity: REG1) and theaverage stride time (stride regularity: REG2). The percent difference between theregularity of steps and strides was used to define the symmetry (SYM) of theacceleration waveform in each axis, with score of 0 indicating perfect symmetry.
Given both the large number of dependent variables and the potential forredundancy of data, a principal component analysis was used as a data reductiontechnique. A principal component analysis creates a set of orthogonal variables thatdescribe the majority of variance in the original dataset (Jolliffe, 2002). Prior to theprincipal component analysis, outliers were adjusted using aWinsorizing technique andall variables were standardized to a mean of 0 and a standard deviation of 1 (Kettaneh etal., 2005). The number of PCs retained was based on analysis of the scree plot and factorloadings, and a Varimax rotation was applied to the retained components to maximizethe factor loadings and increase interpretability (Worthington and Whittaker, 2006).
Two discriminant analyses were used to find the multivariate combination ofthe rotated PCs that would achieve the best binary discrimination on trainingbackground (soccer players vs. experienced marathon runners) and experiencelevel (experienced marathon runners vs. first-time marathon runners). Thediscriminant analysis of PCs has been shown to be an effective method to retainthe important aspects of the data, while eliminating inherent limitations of thediscriminant analysis (Jombart et al., 2010). The PC scores were entered in astepwise fashion (entry¼0.05 and exit¼0.1) to compute the most effectivediscriminant function for group separation. Two models of this stepwise discrimi-nant analysis were computed for each binary classification problem; Model1 containing all PCs (i.e., with gait speed) and Model 2 containing only PCs onwhich gait speed had no loading (i.e., without gait speed). The classificationaccuracy of the discriminant analysis for each model in each binary classificationproblem was determined using a leave-one-out cross-validation method.
3. Results
A total of eight PCs were retained, explaining 88.2% of thevariance in the data (all eigenvalues41.0). Based on the factorloadings in Table 1, these PCs (cumulative variance) were labelledas WaveletLevels2–5 (44.0%), Mediolateral (55.1%), Anteroposterior(65.7%), Vertical (72.6%), Symmetry (77.5%), WaveletLevel1 (81.9%),Regularity (85.6%), and Speed (88.2%). The PC labelled Speed was
Table 1List of principal components (PC) and associated variables (loading) following varimax rotation. Cross-loading not shown.
PC 1 PC 2 PC 3 PC 4 PC 5 PC 6 PC 7 PC 8Wavelet2-5 Mediolateral Anteroposterior Vertical Symmetry Wavelet1 Regularity Speed
Abbreviations represent root mean square of accelerations (RMS), economy of accelerations (ECON), ratio of accelerations (RATIO), discrete wavelet transformation (WAVE),step regularity (REG1), stride regularity (REG2), and symmetry (SYM) in the vertical (V), mediolateral (ML), anteroposterior (AP), and resultant (R) direction and gait speed(Speed).
D. Kobsar et al. / Journal of Biomechanics ∎ (∎∎∎∎) ∎∎∎–∎∎∎2
Please cite this article as: Kobsar, D., et al., Classification accuracy of a single tri-axial accelerometer for training background andexperience level in runners. Journal of Biomechanics (2014), http://dx.doi.org/10.1016/j.jbiomech.2014.04.017i
confirmed to be the only PC that displayed a significant correlation(po0.001) and covariance (po0.001) with gait speed. Therefore,Model 1 consisted of all eight PCs, while Model 2 consisted ofseven PCs, excluding Speed.
For training background (soccer players vs. experienced mara-thon runners), Model 1 revealed a significant discriminant func-tion (Λ¼0.24, χ2 (2, N¼26)¼32.46, po0.001) that explained75.7% of variation between groups using Speed (0.63) and Reg-ularity (�0.28). Model 1 displayed a cross-validated classificationaccuracy of 96.2% for training background. Model 2 revealed asignificant discriminant function (Λ¼0.55, χ2 (3, N¼26)¼13.55,p¼0.004) that explained 45.3% of variation between groups usingVertical (0.57), Regularity (�0.54), and Symmetry (0.23). Model2 displayed a cross-validated classification accuracy of 84.6% fortraining background.
For experience level (experienced marathon runners vs. first-time marathon runners), Model 1 revealed a significant discrimi-nant function (Λ¼0.25, χ2 (3, N¼28)¼34.02, po0.001) thatexplained 75.0% of variation between groups using Speed (0.66),Regularity (�0.27), and WaveletLevel1 (0.16). Model 1 displayed across-validated classification accuracy of 96.4% for experiencelevel. Model 2 revealed a significant discriminant function (Λ¼0.83, χ2 (1, N¼28)¼4.87, p¼0.027) that explained 17.4% ofvariation between groups using Regularity (1.00). Model 2 dis-played a cross-validated classification accuracy of 67.9% forexperience level.
4. Discussion
The aim of this study was to examine the classification ability ofa single tri-axial accelerometer in different groups of runners. Theresults of the present investigations reveal that in combinationwith Speed (Model 1), the accelerometer demonstrated a classifi-cation accuracy of 96% for both training background and experi-ence level. When the effect of Speed was removed (Model 2), theaccelerometer was able to correctly classify 84.6% of the runnerswhen examining training background, but only 67.9% of runnerswhen examining experience level. These findings suggest a singleaccelerometer can effectively classify individuals based on trainingbackgrounds (soccer players vs. experienced marathon runners),but without the effect of gait speed, classification accuracy is much
lower for experience level (experienced marathon runners vs.first-time marathon runners).
In addition to Speed, the PCs labelled Vertical, Regularity, andSymmetry were the most effective classifiers for training back-ground. Fig. 1 illustrates that competitive soccer players were bestdistinguished from experienced marathon runners using theVertical and Regularity PCs, with further distinction occurringwith the addition of the Symmetry PC. These scores suggest thatcompetitive soccer players, trained in explosive, sport-specificmovements, displayed a pattern of motion involving a greatervertical component (Fig. 2), along with less symmetry and reg-ularity in comparison to experienced marathon runners, trained inlong-distance running. This finding is similar to McGregor et al.(2009) who reported that trained long-distance runners displayedsignificantly less accelerations near the centre of mass, includingthose in the vertical axis, compared to recreationally activesubjects. The current results were able to demonstrate the dis-criminatory ability of the vertical acceleration component, as wellas the additional discriminatory ability of symmetry and regularitycomponents not previously studied in running.
Unlike training background, the classification on experiencelevel was not as clear. Although Speed was an excellent classifier of
Fig. 1. Soccer players (red) were most distinguishable from experienced marathonrunners (blue) using Vertical and Regularity PCs (2D projection), but the optimalclassification of these runners occurred with the addition of the Symmetry PC (3D).
Fig. 2. Average accelerometer signal for a one gait cycle in the vertical (top),mediolateral (middle), and anteroposterior (bottom) direction providing a visuali-zation of a typical accelerometer trace for each group. While a greater verticalcomponent can be observed in Soccer players compared to experienced marathonrunners (*), the measurement of symmetry and regularity can be difficult tointerpret when signals are averaged across all participants.
D. Kobsar et al. / Journal of Biomechanics ∎ (∎∎∎∎) ∎∎∎–∎∎∎ 3
Please cite this article as: Kobsar, D., et al., Classification accuracy of a single tri-axial accelerometer for training background andexperience level in runners. Journal of Biomechanics (2014), http://dx.doi.org/10.1016/j.jbiomech.2014.04.017i
experience level, the pattern of motion measured by the accel-erometer displayed limited discriminatory ability. Experiencedmarathon runners generally exhibited greater regularity in theirpattern of motion compared to first-time marathon runners, butRegularity alone classified less than 70% of these runners correctly.(FIG) This finding is similar to research in which biomechanicaldifferences pertaining to running experience can be subtle(Fukuchi et al., 2011, 2014) and often times contradictory, suggest-ing an individualized, longitudinal approach may be required toobserved differences (Ferber et al., 2009; Williams, 2007). There-fore, the current results do not necessarily reflect limitations of theaccelerometer itself, but rather the limitations in the cross-sectional nature of the study.
In conclusion, the current study provides promising results forthe application of accelerometry to the classification of runners, aswell as indications for future research in this area. First, variablesrelating to the vertical, symmetry, and regularity components ofaccelerations near the centre of mass appear to be effectivepredictors of training background, and gait speed remains animportant classifier of both training background and experiencelevel. Therefore, stratifying or matching runners based on gaitspeed may be important in creating homogenous groups ofrunners (e.g., experience vs. first-time runners), but furtherclassifying or modelling of the pattern of motion near the centreof mass with an accelerometer could be useful for coaches andtrainers in tracking rehabilitation or performance. For example,Fig. 2 provides a representative trace of the accelerometer patterndemonstrated by each group. This information could be used intracking the rehabilitation of athletes as they return to the“typical” pattern for their population or tracking the changes thatmay occur as a runner moves from a less experienced runner to amore experienced runner. The current study also provides evi-dence for these applications and the use of a single tri-axialaccelerometer as a simple and accessible classification tool forcentre of mass movement patterns in running.
Conflict of interest statement
There are no conflicts of interest associated with this research.
Acknowledgements
This research was supported by funding from the Universityof Calgary Athletics Department (Grant no. 63020-KIS020000) andAlberta Innovates: Health Solutions (Grant no. 20080449).
References
Altun, K., Barshan, B., Tunçel, O., 2010. Comparative study on classifying humanactivities with miniature inertial and magnetic sensors. Pattern Recognit. 43(10), 3605–3620.
Chen, K.Y., Janz, K.F., Zhu, W., Brychta, R.J., 2012. Re-defining the roles of sensors inobjective physical activity monitoring. Med. Sci. Sports Exerc. 44 (1), S13–S23.
Dalton, A., Khalil, H., Busse, M., Rosser, A., van Deursen, R., ÓLaighin, G., 2013.Analysis of gait and balance through a single triaxial accelerometer inpresymptomatic and symptomatic Huntington's disease. Gait Posture 37,49–54.
Ferber, R., Hreljac, A., Kendall, K.D., 2009. Suspected mechanisms in the cause ofoveruse running injuries: a clinical review. Sports Health 1 (3), 242–246.
Fukuchi, R.K., Eskofier, B.M., Duarte, M., Ferber, R., 2011. Support vector machinesfor detecting age-related changes in running kinematics. J. Biomech. 44 (3),540–542.
Fukuchi, R.K., Stefanyshyn, D.J., Stirling, L., Duarte, M., Ferber, R., 2014. Flexibility,muscle strength and running biomechanical adaptations in older runners. Clin.Biomech. 29 (3), 304–310.
Godfrey, A., Bourke, A.K., Olajghin, G.M., van de Ven, P., Nelson, J., 2011. Activityclassification using a single chest mounted tri-axial accelerometer. Med. Eng.Phys. 33 (9), 1127–1135.
Jolliffe, I.T., 2002. Principal Component Analysis, Springer Series in Statistics, 2nded. Springer, New York.
Jombart, T., Devillard, S., Balloux, F., 2010. Discriminant analysis of principalcomponents: a new method for the analysis of genetically structured popula-tions. BMC Genet. 11 (1), 94.
Kettaneh, N., Berglund, A., Wold, S., 2005. PCA and PLS with very large data sets.Comput. Stat. Data Anal. 48, 69–85.
Kobsar, D., Olson, C., Paranjape, R., Hadjistavropoulos, T., Barden, J.M., 2014.Evaluation of age-related differences in the stride-to-stride fluctuations, reg-ularity and symmetry of gait using a waist mounted tri-axial accelerometer.Gait Posture 39 (1), 553–557.
McGregor, S.J., Busa, M.A., Yaggie, J.A., Bollt, E.M., 2009. High resolution MEMSaccelerometers to estimate VO2 and compare running mechanics betweenhighly trained inter-collegiate and untrained runners. PloS One 4 (10), e7355.
Moe-Nilssen, R., Helbostad, J.L., 1998. A new method for evaluating in gait underreal-life environmental conditions. Clin. Biomech. 34, 320–327.
Moe-Nilssen, R., Helbostad, J.L., 2004. Estimation of gait cycle characteristics bytruck accelerometry. J. Biomech. 37, 121–126.
Preece, S.J., Goulermas, J.Y., Kenney, L.P.J., Howard, D., 2009. A comparison of featureextraction methods for the classification of dynamic activities from acceler-ometer data. IEEE Trans. Biomed. Eng. 56 (3), 871–879.
Worthington, R.L., Whittaker, T.A., 2006. Scale development research: a contentanalysis and recommendations for best practices. Couns. Psychol. 34 (6),806–838.
Yang, M., Zheng, H., Wang, H., McClean, S., Hall, J., Harris, N.A., 2012. Machinelearning approach to assessing gait patterns for complex regional painsyndrome. Med. Eng. Phys. 34 (6), 740–746.
D. Kobsar et al. / Journal of Biomechanics ∎ (∎∎∎∎) ∎∎∎–∎∎∎4
Please cite this article as: Kobsar, D., et al., Classification accuracy of a single tri-axial accelerometer for training background andexperience level in runners. Journal of Biomechanics (2014), http://dx.doi.org/10.1016/j.jbiomech.2014.04.017i
