Predicting gait events from tibial acceleration in rearfoot running: a structured machine learning approach
Pieter Robberechts, Rud Derie, Pieter Van den Berghe, Joeri Gerlo, Dirk De Clercq, Veerle Segers, Jesse Davis
PPredicting gait events from tibial acceleration inrearfoot running: a structured machine learningapproach
Pieter Robberechts ∗ , Rud Derie (cid:63) , Pieter Van den Berghe , Joeri Gerlo , DirkDe Clercq , Veerle Segers , and Jesse Davis Department of Computer Science, KU LeuvenLeuven, Belgium [email protected] Department of Movement and Sport Sciences, Ghent UniversityGhent, Belgium [email protected]
Abstract.
Gait event detection of the initial contact and toe off is essen-tial for running gait analysis. Heuristic-based methods exist to estimatethese key gait events from tibial accelerometry. These heuristic-basedmethods are unfortunately tailored to very specific acceleration profiles,which may offer complications when dealing with larger data sets andinherent biological variability. Therefore, the purpose of this study wasto compare a previously utilised heuristic method of gait event detec-tion to an original proposed method using a structured machine learningapproach. Force-based event detection acted as the criterion measure inorder to assess the accuracy of the predicted gait events. 3D tibial ac-celeration and ground reaction force data from 93 rearfoot runners werecaptured. A heuristic method and two machine learning methods wereemployed to derive initial contact, toe off and stance time from tibialacceleration signals. Both machine learning methods significantly out-performed existing heuristic approaches. Furthermore, results indicatethat a structured recurrent neural network machine learning model of-fers the most accurate and consistent estimation of the gait events andits derived stance time during level overground running. The machinelearning methods seem less affected by intra- and inter-subject variationwithin the data, allowing for accurate and efficient automated data out-put with possibilities for real-time monitoring and biofeedback duringprolonged measurements.
The running gait comprises of the stance and swing phases, separated by twokey events: initial contact (IC) and toe off (TO) (Figure 1). Determining thetiming of these events allows performing a detailed stride-by-stride analysis ofa runner’s gait. Moreover, many variables relevant for gait analysis are defined (cid:63)
PR and RD contributed equally to this work. a r X i v : . [ c s . L G ] D ec P. Robberechts, R. Derie et al. with respect to either one or both of these gait events. For example, stance time(ST) is defined as the time span between both events. Therefore, accurate andconsistent detection of these gait events is crucial for any gait analysis.The criterion instrument for gait event detection is a force platform [15]. Thisexpensive device is part of an instrumented runway or treadmill [13,18], hencerestricted by its measurement zone. Therefore, body-worn accelerometers andmatching detection algorithms have been proposed as an ambulatory methodfor gait event detection [6].
Time [ms]
TOIC vGRF [BW]
ACC [g]
StancePhaseSwingPhase SwingPhase
Fig. 1.
The running gait cycle can be split in a stance phase and a swing phase,separated by the initial contact (IC) and toe-off (TO) events. The timing of these gaitevents are estimated from 3D tibial acceleration signals (ACC). Ground truth labelsare based on the vertical ground reaction force (vGRF).
However, these accelerometry-based event detection methods are all heuristic-based [10,11,14]. For example, Mercer et al. [11] used an accelerometer affixed tothe shin, identifying IC as a local minima before the axial peak tibial accelerationand TO as the second local maxima after that axial peak. These heuristic-basedmethods assume that both gait events are associated with typical accelerationfeatures, neglecting inter-subject variation. Recent advances in the field of ma-chine learning [9,13], specifically the success of neural networks [5], suggest thata data-driven approach for gait event detection using machine learning may leadto better accuracy and consistency.This study evaluated gait event detection (IC-TO) from 3D tibial accelera-tion signals using a heuristic-based method and two machine learning methods.Criterion validation happened by comparing the estimated timings to those de-termined using a force platform. We evaluated the success of event detection, theabsolute error of prediction and its variability to propose an accurate estimationmethod. ait Event Detection in Tibial Acceleration Profiles 3
This study recruited 93 rearfoot runners from the local running community (Ta-ble 1). These runners were free of running-related injuries during the last sixmonths, ran at least 15 km per week and signed an informed consent. Approvalfor the study was obtained from the local ethical committee (bimetra 2015/0864).
Table 1.
Characteristics of subjects, expressed as mean ± standard deviation. Men Women (n = 55) (n = 38)
Age (Yrs) 35 . ± . . ± . . ± .
07 1 . ± . . ± . . ± . . ± . . ± . Data collection took place during two different projects, but with the samemeasurement setup. The first cohort consisted of 13 subjects, wearing a stan-dardized neutral running shoe (Li Ning Magne, ARHF041), who were asked torun on a 30-m instrumented running track at multiple speeds (2.55 m.s -1 , 3.20m.s -1 , 5.10 m.s -1 and preferred running speed) [18]. The second cohort consistedof 80 runners, wearing their own regular training shoes, running at 3.20 m.s -1 only. The running speed was controlled with timing gates and runners receivedfeedback if they did not run within ± -1 of the target speed.All runners wore a backpack/tablet system to measure the tibial acceleration.Two tri-axial accelerometers (LIS331, Sparfkun, Colorado, USA;1000 Hz/axis),were tightly strapped with medical tape on the antero-medial side of both shins,eight centimeters above the medial malleolus [18]. Accelerometers were orien-tated along the longitudinal axis of the tibia. The skin around the lower legwas pre-stretched with sport tape to improve the rigid coupling between the ac-celerometers and the tibia. Simultaneously, ground reaction forces were measuredat 1000 Hz by two built-in force platforms (2 m and 1.2 m, AMTI, Watertown,MA). Tibial acceleration and force data were synchronized in time by meansof an infrared impulse sent from a motion capture system and captured by aninfrared sensor at the backpack system [18]. The vertical ground reaction force was filtered using a Butterworth second-order,zero-lag, low-pass filter with a 60 Hz cutoff frequency. For each trial containingat least three steps, the second step in the sequence was extracted using the forcedata. For each second step, a period ranging from 200 ms before IC to 200 ms
P. Robberechts, R. Derie et al. after TO was extracted. The data from the left and right leg were mirrored. Con-sequently, each of these steps starts with a right foot making ground contact.This procedure resulted in 1003 examples. Tibial acceleration signals were fil-tered using a second-order band-pass filter with cutoff frequencies of 0.8 Hz and45 Hz. Using the filter configuration as a hyper-parameter during the learningphase (Section 2.5), this configuration gave the best results.
For each sampled value of each example, a feature vector was constructed fromthe x, y and z components of the bi-lateral acceleration profiles. Below we de-scribe the features used in the final models. For the full list of considered features,we refer to the supplementary materials.
Filtered Acc { Left, Right } { x, y, z } The raw values in the filter accelerationsignals at each time step.
Filtered Acc { Left, Right } Total
The magnitude of the resultant accelera-tion.
Jerk { Left, Right } { x, y, z } The first derivative of the bandpass-filtered ac-celeration signals.
Jerk { Left, Right } Total
The magnitude of the resultant jerk.
Roll { Left, Right } The roll extracted from the acceleration signals. Here acustom second-order Butterworth low-pass filter at 60 Hz was applied.
Pitch { Left, Right } The pitch extracted from the same low-pass filtered ac-celeration signals.
Acc Right x Peak Min
A moving average filtered labeling of local minima inthe anterior-posterior x-component of the foot making ground contact. Thismarks the neighborhood of a clear peak value for the underlying accelerationsignal.All features were standardized by removing the mean and scaling to unit vari-ance. This scaling happened independently on each feature and independentlyfor each example.
Formally, the problem of gait event detection can be specified as:
Given:
A 3D tibial acceleration signal x ( t ) of length l , described by asequence x = ( x , . . . , x l ) of D-dimensional feature vectors x t ∈ R D with t = 1 , . . . l . Find:
The gait event or phase y t ∈ { Swing, IC, Stance, TO } for eachcorresponding x t , such that y ∗ = ( y , . . . , y l ) is the correct sequence ofgait events and phases. ait Event Detection in Tibial Acceleration Profiles 5 In machine learning, this type of problem is traditionally solved by multiclassclassification algorithms. In this setting, the task is to find the most likely outputlabel y ∈ { , , . . . , K } for each input x ∈ R D . Therefore, the algorithm learnsa scoring function f such that f ( x, y ∗ ) > f ( x, y ) for all y (cid:54) = y ∗ , where y ∗ is thetrue label and y is an imposter label. This scoring function is then evaluated foreach possible output label y and the sample is finally labelled with the highestscored output label y ∗ : y ∗ = argmax y ∈ [1 ,K ] f ( x, y ) . (1)For computational feasibility, the scoring functions generally are a linear formof a joint feature vector Φ ( x, y ): f ( x, y ) = w T · Φ ( x, y ) . Here, the features Φ ( x, y ) should quantify how “compatible” the input x iswith the output y . The vector w are parameters learned from the data thatcorrespond to the weight given to each feature in the computation of the finalscore. A natural way to represent Φ is an outer product between x and the labelspace. This yields the following representation: Φ ( x, k ) = < , , . . . , (cid:124) (cid:123)(cid:122) (cid:125) D ( k −
1) zeros , ∈ R D (cid:122)(cid:125)(cid:124)(cid:123) x , , , . . . , (cid:124) (cid:123)(cid:122) (cid:125) D ( K − k ) zeros > ∈ R DK , (2)with D the number of features and k ∈ y . In this representation, w effectivelyencodes a separate weight for every feature/label pair.In the case of gait event detection, however, the output has a natural struc-ture: IC and TO events alternate each other and the time difference betweenboth events is similar from stride to stride. We can benefit from this inherentstructure of the output to train a more accurate predictor [2].In this “structured prediction” setting a score is similarly assigned to eachpossible output, given the input. However, both input and output are now se-quences instead of individual samples. Specifically, given an input sequence x of a tibial acceleration signal and a corresponding possible segmentation y =(Swing, IC, Stance, . . . ), the task is to find the element y ∗ of all possible outputsequences γ that maximizes a scoring function w T · Φ ( x, y ): y ∗ = argmax y ∈ γ w T · Φ ( x, y ) . (3)However, in the structured setting, every input x has many possible segmen-tations (Figure 2). Therefore, the main challenge is how to efficiently search forthe optimal input-output combination. Below, we introduce two machine learn-ing algorithms to solve this problem. P. Robberechts, R. Derie et al.
FeaturerepresentationStructured prediction
Acc xAcc yAcc z TOIC
SwingICStanceTOSwingICStanceTO SwingICStanceTO SwingICStanceTO SwingICStanceTO SwingICStanceTO SwingICStanceTO SwingICStanceTOSwing IC Stance Stance Stance Stance TO Swing
Fig. 2.
A graphical illustration of the structured prediction task. First, the accelera-tion signals of a step are transformed into a feature vector representation. Next, thestructured prediction algorithm maps the sequence of input vectors to the most likelygait segmentation sequence. At each time step t the corresponding feature vector x t can have any of the four possible labels. Any path through this trellis corresponds to aunique labeling of this stride. The gold standard path is drawn with bold green arrows. The ground truth of the IC and TO timings was determined per vertical groundreaction force. The threshold for detection was set at 20 N. Additionally, weused the method defined by Mercer et al. [11] (hereafter referred to as the M-method [12]) to set a baseline for the machine learning models. The IC and TOtimings were herein defined as the minimum point before positive axial peaktibial acceleration and the minimum acceleration after a second local maximumafter positive axial peak tibial acceleration, respectively.
As a simple structured learning algorithm, we used the averaged structured per-ceptron algorithm [3,4] from the SeqLearn package. This is a generalization ofthe standard perceptron algorithm and the basic building block of the deep learn-ing algorithm introduced in the next section. The algorithm makes predictionsby choosing the output sequence y that maximizes a score given by w T Φ ( x, y ).And if this output sequence is incorrect, it will adjust the weights w toward thecorrect output sequence y ∗ .The key insight is that Φ ( x, y ) can be decomposed as (cid:80) pi =1 Φ (cid:48) i ( x, y i ). Thismakes it possible to search efficiently over all possible output sequences y us-ing a variant of the Viterbi algorithm [7]. The joint feature functions Φ (cid:48) i are a http://larsmans.github.io/seqlearn/ait Event Detection in Tibial Acceleration Profiles 7 combination of the unary features given by equation 2 and Markov features thatquantify the likelihood of transitioning from one state to another in the nextsample. These are learned from the data and represent for example, that an ICevent is always followed by the stance phase.
The joint feature vector Φ ( x, y i ) can be extracted using different techniques.Usually, this is done by handcrafting features as in the structured perceptronmodel. An interesting alternative for the gait event detection problem is theusage of a Recurrent Neural Network (RNN) model [1]. RNNs are a deep networkarchitecture that can model the behavior of dynamic temporal sequences usingan internal state which can be thought of as memory [8]. RNNs provide theability to predict the current frame based on the previous and/or next frames.As such, the model can learn which long-term patterns in the acceleration profilesare relevant for determining the timings of the gait events.We optimized a RNN model and a structured prediction model in an end-to-end fashion. Therefore, we use the structural hinge loss [17] l ( w, x, y ) = max y ∗ ∈ γ (cid:2) , l ( y, y ∗ ) − w T Φ ( x, y ) + w T Φ ( x, y ∗ ) (cid:3) (4)with l ( y, y ∗ ) = | y IC − y ∗ IC | + | y TO − y ∗ TO | . (5)Since both the loss function and the RNN are differentiable, we can optimizethem using stochastic gradient descent.Specifically, we use the RNN outputs as feature functions for a structuredprediction model (Figure 3). First, an RNN encodes the filtered accelerationsignals of an entire stride and outputs a new representation for each of thesamples. This new representation corresponds to an approximate likelihood ofan event. Then an efficient search is executed over all possible sample valuesso that the most probable one can be selected. Therefore, we use a constrainedpeak detection algorithm. This algorithm selects the local maxima that satisfythe following constraints: – A IC and TO event of opposing foot are separated by at least 35 ms and atmost 200 ms – A TO and IC event of the same feet are separated by at least 160 ms and atmost 350 ms
The machine learning models were trained and assessed in a two-step procedure.First, 5-fold cross-validation was used to obtain a good set of features and hyper-parameters for both models. For the perceptron model, the dataset was splitinto training (83 subjects) and test (10 subjects) sets. We found the 21 features
P. Robberechts, R. Derie et al.
TO TOIC IC0 100 200 300 400 500 600600
Time [ms] ...... ...... ...... ...... ............
True event timingPredicted event timing Vertical accleft foot Vertical accright foot Ф Fig. 3.
Visual representation of a two-layer bidirectional recurrent neural network withlong short-term memory cells to map the multivariate time series of tibial accelerationprofiles (top plot) to the likelihood of an event. Subsequently, a constrained peakdetection algorithm is used to determine the most likely timing of IC and TO events.For simplicity, we only plot the axial acceleration profiles.ait Event Detection in Tibial Acceleration Profiles 9 described in section 2.3 to give the best result. The learning rate was set to 0.1,which is the only parameter of this model. For the RNN model, the dataset wasrandomly split into training (73 runners), validation (10 runners), and test (10runners). The validation set was used for early stopping. The same features asin the perceptron model were used, excluding the
Acc Right x Peak Min feature.Furthermore, we achieved the best results with two bidirectional long short-termmemory layers with dropout 0.2 after each recurrent layer and 50 hidden units.Using these hyper-parameters and features, the models were retrained in aleave-one-out cross-validation analysis to evaluate the accuracy. Each model wasiteratively trained on 92 of the 93 test subjects and then the accuracy of themodel was tested on the 93 th subject. This procedure was repeated 93 times andeach time the data of a different subject was left out, obtaining an out-of-sampleprediction for each subject’s steps. Doing hyper-parameter tuning and featureselection for each of these 93 folds separately is not computationally feasible,hence the two-step procedure.For each step, the relative error and absolute error were determined for theestimated IC and TO event timings [12]. Relative errors were calculated as thearithmetic difference (ms) between the predicted event timings ( t acc ) obtainedthrough the acceleration profiles and reference timings ( t vgrf ) obtained throughthe force-platform method: Relative error = t acc − t vgrf . A positive value indicatesthat the detected event occurred after the reference (time lag). Absolute errors,indicating the error magnitude regardless of direction, were calculated as theabsolute value of relative errors: Absolute error = | t acc − t vgrf | .ST was determined from the estimated gait events. As for the event timings,relative and absolute errors on the estimated ST were calculated as the arith-metic difference between the estimated ST using the accelerometer-based methodand reference. Here, a positive relative error corresponds to an overestimationof the ST.The number of trials completed by each runner varied. In order to avoidthat one runner would excessively impact the accuracy of our models, we com-puted the global median relative error and median absolute error in a two-stepprocedure. First, for each runner, the average median absolute error and me-dian relative error were computed over all strides of that runner. Thereafter,the global metrics were calculated as the median values of these metrics over allrunners. Statistical analysis was executed in Python using the
Statsmodels and
Scipy libraries, with the significance level set at p = 0 .
05. A Shapiro-Wilk test fornormality was first performed on the relative difference of ST. Subsequently, aFriedman test (and Wilcoxon signed-rank tests for comparing pairs of models)and Levene’s test for non-normal distribution were used to examine whether thevarious prediction methods have significantly different accuracies and standarddeviations. Post-hoc testing was conducted using Bonferroni correction. Failed predictions were imputed with the subject’s average estimated ST at the cor-responding running speed. Statistical analysis on the IC and TO estimates wasnot possible since there exists no logical imputation.
The relative differences of ST showed a non-normal distribution (all p < . p < . p < . Table 2.
Median relative error (MRE) and median absolute error (MAE) between eachestimation method and the reference for initial contact and toe-off detection ( − : timelead; +: time lag) and stance time estimation ( − : shorter time; +: longer time). Stancetime was calculated as the timespan between the initial contact and toe-off events. Median ± standard deviation. Variable Method Initial Contact Toe-off Stance time
MRE(ms) M-method − . ± . − . ± . − . ± . . ± . . ± . − . ± . − . ± . . ± . . ± . . ± . . ± . . ± . . ± . . ± . . ± . . ± . . ± . . ± . Failedpredictions(%) M-method 00.60 21.93 22.53Structured perceptron - - 2.99Structured RNN - - 1.69
Bold : minimum MAE for detected IC and TO and estimated stance time.
Regarding the absolute error of estimation, our machine learning models bet-ter estimated the IC than the M-method (Table 2). The TO event was harderto estimate. Still, the RNN model clearly outperformed the other methods (Ta-ble 2). This RNN model also significantly outperformed the other methods interms of error variability ( p < . ait Event Detection in Tibial Acceleration Profiles 11 than a given threshold for the machine learning models and the heuristic M-method. In 83% of the examples, the RNN model was off by at most 10 mswhereas the perceptron model only attained this level of accuracy in 40% of theexamples. Error tolerance (Absolute error [ms]) 0% 10% 20% 30% 40% 50% 60% 70% 80% 90%100%
Sensitivity (% of examples)
RNN Perceptron M method
Fig. 4.
Sensitivity (true positive ratio) of the three methods as a function of errortolerance. The absolute error (ms) is the deviation from the stance time, as determinedby means of a force plate. Both of our machine learning models outperform the heuristicM-method.
This study examined gait event detection (IC-TO) in 3D tibial acceleration sig-nals of rearfoot runners. Three approaches for event detection were examined:a heuristic-based method [11], a structured perceptron model with hand-craftedfeatures and a deep learning structured RNN model. While the M-method de-tected the gait events consistently earlier with underestimation of the ST [12],our machine learning models better handled the estimation of IC and TO. TheM-method cannot deal with the variation observed in the acceleration signalsbetween subjects and even between strides of a subject. Heuristic-based meth-ods to determine gait events relying on a universal pattern in the accelerationprofiles results in a consistent lead (too soon) or lag (too late) in the estimates.Using a structured learning approach, our machine learning algorithms coulddeal with some variability in 3D tibial acceleration. Despite accurate predictionsfor most examples, the errors for an individual step can be quite large. For 3% ofthe examples, the RNN model predicted a ST that deviates at least 50 ms from the criterion reference (Figure 4). Most acceleration patterns of these examplesbelong to the same four subjects and were manually investigated. Unfortunately,no clear patterns were distinguished. One may further improve the models byadding data of a large number of runners with different, unique patterns. Alterna-tively, a model could be specifically trained for each pattern. Using principles oftransfer learning, one can learn such specific models from a limited dataset [16].The automated methods enable accurate and real-time detection of key eventswhilst running overground. The mean computation time of the perceptron andRNN models needed to go from the raw signals to estimated gait event timingswere respectively 4 ms and 142 ms (2.3 GHz Intel Core i5, 16 GB LPDDR3RAM), which is below a typical ST in overground rearfoot running at submaxi-mal speed. Given that the prediction models require data of a complete step tomake a prediction, estimates can be provided before the end of the next step.This ability to output running gait parameters accurately and promptly permitsthe development of an automated feedback system based on the consistency orfluctuation of spatio-temporal parameters. Further research should investigateour proposed method when applied outdoors on different terrains. Altogether,this study presents a structured RNN learning approach which accurately de-tects IC and TO events in 3D tibial acceleration profiles for rearfoot runnerswhen running indoor on a standard sports floor. This algorithm offers possibil-ities towards implementation in overground gait analysis or gait-retraining ofrearfoot runners without the need of an embedded force plate.
Acknowledgements
This work was supported by the H2020 Interreg EU (Nano4Sports project),the Research Foundation Flanders (FWO.3F0.2015.0048.01), the KU LeuvenResearch Fund (C32/17/036) and the International Society of Biomechanics(matching dissertation grant program 2019).
References
1. Adi, Y., Keshet, J., Cibelli, E., Goldrick, M.: Sequence segmentation using jointrnn and structured prediction models. In: 2017 IEEE International Conference onAcoustics, Speech and Signal Processing (ICASSP). pp. 2422–2426. IEEE (2017)2. Chang, K.W., Kundu, G., Roth, D., Srikumar, V.: Learning and inference in struc-tured prediction models. In: AAAI-16 Tutorial Forum (February 2016)3. Collins, M.: Discriminative training methods for hidden markov models: Theoryand experiments with perceptron algorithms. In: Proceedings of the ACL-02 Con-ference on Empirical Methods in Natural Language Processing - Volume 10. pp. 1–8. EMNLP ’02, Association for Computational Linguistics, Stroudsburg, PA, USA(2002). https://doi.org/10.3115/1118693.11186944. Daume, III, H.C.: Practical Structured Learning Techniques for Natural LanguageProcessing. Ph.D. thesis, University of Southern California, Los Angeles, CA, USA(2006)5. Deng, L., Yu, D., et al.: Deep learning: methods and applications. Foundations andTrends in Signal Processing (3–4), 197–387 (2014)ait Event Detection in Tibial Acceleration Profiles 136. Falbriard, M., Meyer, F., Mariani, B., Millet, G.P., Aminian, K.: Accurate estima-tion of running temporal parameters using foot-worn inertial sensors. Frontiers inPhysiology , 610 (Jun 2018). https://doi.org/10.3389/fphys.2018.006107. Forney, G.D.: The viterbi algorithm. Proceedings of the IEEE (3), 268–278(1973)8. Graves, A., Fernandez, S., Gomez, F., Schmidhuber, J.: Connectionist temporalclassification: Labelling unsegmented sequence data with recurrent neural net-works. In: Proceedings of the 23rd International Conference on Machine Learning.pp. 369–376. ACM (2006)9. Halilaj, E., Rajagopal, A., Fiterau, M., Hicks, J.L., Hastie, T.J., Delp, S.L.:Machine learning in human movement biomechanics: Best practices, commonpitfalls, and new opportunities. Journal of Biomechanics , 1–11 (Nov 2018).https://doi.org/10.1016/j.jbiomech.2018.09.00910. Lee, J.B., Mellifont, R.B., Burkett, B.J.: The use of a single iner-tial sensor to identify stride, step, and stance durations of running gait.Journal of Science and Medicine in Sport (2), 270–273 (Mar 2010).https://doi.org/10.1016/j.jsams.2009.01.00511. Mercer, J.A., Bates, B.T., Dufek, J.S., Hreljac, A.: Characteristics of shock atten-uation during fatigued running. Journal of Sports Sciences (11), 911–919 (Nov2003). https://doi.org/10.1080/026404103100014038312. Mo, S., Chow, D.H.K.: Accuracy of three methods in gait event detection dur-ing overground running. Gait & Posture (Supplement C), 93–98 (Jan 2018).https://doi.org/10.1016/j.gaitpost.2017.10.00913. Ngoh, K.J.H., Gouwanda, D., Gopalai, A.A., Chong, Y.Z.: Estimation ofvertical ground reaction force during running using neural network modeland uniaxial accelerometer. Journal of Biomechanics , 269–273 (Jul 2018).https://doi.org/10.1016/j.jbiomech.2018.06.00614. Norris, M., Kenny, I.C., Anderson, R.: Comparison of accelerometry stride timecalculation methods. Journal of Biomechanics (13), 3031–3034 (Sep 2016).https://doi.org/10.1016/j.jbiomech.2016.05.02915. Novacheck, T.F.: The biomechanics of running. Gait & posture (1), 77–95 (1998)16. Pan, S.J., Yang, Q.: A survey on transfer learning. IEEE Transac-tions on Knowledge and Data Engineering (10), 1345–1359 (Oct 2010).https://doi.org/10.1109/TKDE.2009.19117. Tsochantaridis, I., Joachims, T., Hofmann, T., Altun, Y.: Large margin methodsfor structured and interdependent output variables. Journal of Machine LearningResearch (Sep), 1453–1484 (2005)18. Van den Berghe, P., Six, J., Gerlo, J., Leman, M., De Clercq, D.: Validity andreliability of peak tibial accelerations as real-time measure of impact loading duringover-ground rearfoot running at different speeds. Journal of Biomechanics86