Automatic Detection of Gait Asymmetry

Maciej Cwierlikowski and Mercedes Torres Torres

School of Computer Science, University of Nottingham, Nottingham, U.K.

Keywords:

Machine Learning, Biomechanics, Gait Symmetry Assessment, Gait Classiﬁcation.

Abstract:

Gait analysis, and gait symmetry assessment in particular, are commonly adopted in clinical settings to deter-

mine sensorimotor ﬁtness reﬂecting body’s ability to integrate multi-sensory stimuli, and use this information

to induce ongoing motor commands. Inter-limb deviation can serve as a non-invasive marker of gait function

to identify health conditions and monitor the effects of rehabilitation regimen. This paper examines the perfor-

mance of machine learning methods (decision trees, k-NN, SVMs, ANNs) to learn and predict gait symmetry

from kinetic and kinematic data of 42 participants walking across a range of speeds on treadmill. Classiﬁca-

tion was conducted for each speed independently with several feature extraction techniques applied. Subjects

elicited gait asymmetry, yet ground reaction forces were more discriminative than joint angles. Walking speed

affected gait symmetry with larger discrepancies registered at slower speeds; the highest F1 scores were noted

at the slowest condition (decision trees: 87.35%, k-NN: 91.46%, SVMs: 88.88%, ANNs: 87.22%). None of

the existing research has yet addressed ML-assisted assessment of gait symmetry across a range of walking

speeds using both, kinetic and kinematic information. The proposed methodology was sufﬁciently sensitive to

discern subtle deviations in healthy subjects, hence could facilitate an early diagnosis when anomalies in gait

patterns emerge.

1 INTRODUCTION

Gait is scientiﬁcally recognised as a highly coordi-

nated cooperation between the central nervous and

musculoskeletal systems. Its classiﬁcation has the po-

tential to provide an insight into patient’s quality of

life, predict cognitive decline (Verghese et al., 2007)

and estimate fall risk (Beauchet et al., 2009). Healthy

gait can be viewed as a recurring and rhythmic pro-

cess of translating the centre of mass (COM) through

space whilst keeping the energy expenditure at min-

imum implying optimal path of the COM motion

(Zhang et al., 2010). Conversely, pathological gait

can be conceptualised as an attempt to maintain body

economy by introducing compensatory exaggerations

of movements at unaffected levels (Inman and Eber-

hart, 1953) which may be manifested by deviations in

gait patterns that are indicative for health problems of

various origins including neurological (e.g. Parkin-

son, multiple sclerosis, post-stroke hemiplegic gait)

and systemic (e.g. cardiopathies, osteoporosis) dis-

eases (Muro et al., 2014).

Gait (a)symmetry, deﬁned as a measure of the par-

allels of lower extremities with respect to a selected

gait parameter (Viteckova et al., 2018), is one of the

markers argued as capable of capturing these devia-

tions thereby reﬂecting the (dis)order within/between

systems controlling the human locomotion. Up to

date, gait symmetry assessment relied predominantly

on algebraic formulas e.g. symmetry index (Robinson

et al., 1987), expressing bilateral difference of spatio-

temporal (e.g. stride length), kinetic (e.g. ground re-

action forces (GRF)) or kinematic (e.g. joint angles)

variables. However, it has been proven that these tra-

ditional approaches have limitations related to the de-

gree of reported detail due to its discrete nature and

that they are insufﬁcient to capture complex discrim-

inative information contained in patterns over a com-

plete gait cycle (Kutilek et al., 2014).

In this paper, we present, to our knowledge, the

ﬁrst machine learning-assisted gait asymmetry as-

sessment carried out across a range of speeds. Our

study includes a population of 42 healthy subjects.

Four traditional classiﬁers were trained to discrim-

inate between left/right gait patterns based on the

lower-limb kinematics or kinetics. Results showed

that gait asymmetry was discernible although its mag-

nitude varied as a function of the walking speed with

larger asymmetries occurring at slower speeds for

which classiﬁcation performance was superior.

Cwierlikowski, M. and Torres, M.

Automatic Detection of Gait Asymmetry.

DOI: 10.5220/0009787501690176

In Proceedings of the 9th International Conference on Data Science, Technology and Applications (DATA 2020), pages 169-176

ISBN: 978-989-758-440-4

169

2 RELATED WORK

Scarcity of quantitative tools to diagnose motor symp-

toms popularised automated movement classiﬁcation

which already enriched medical practice and selected

studies are presented in section 2.1. However, only

few attempts (section 2.2) speciﬁcally targeted gait

asymmetry detection by means of ML techniques.

2.1 Diagnosis of Gait Disorders using

Machine Learning

Several studies have explored machine learning for

gait disorder prediction. Mezghani et al. (2016) used

classiﬁcation and regression tree algorithm to im-

prove objectivity of joint replacement surgery recom-

mendation for osteoarthritis (OA) patients. 3D knee

angles time-series of each patient with a moderate-to-

severe knee OA were used to predict surgical recom-

mendation. Results showed high ability of the model

to differentiate between surgical and non-surgical

candidates as expressed by sensitivity (79.5%), speci-

ﬁcity (90%), accuracy (84.7%) and ROC curve

(0.8477) metrics. However, authors did not report

whether any hyperparameter tuning was performed.

In fact, by optimising model’s structure, potential

gains in performance might be medically meaningful.

Salazar et al. (2004) addressed the limitations of a

manual classiﬁcation of the Spastic Hemiplegia (SH)

varieties. Joint angles of the hemiparetic limb were

used to train 7 binary SVM models and one-against-

all scheme was adopted to account for 7 different

classes (6 SH types + 1 healthy). The speciﬁcity met-

ric ranged between 84.61% and 98.69% yet the study

lacked a systematic testing to prevent the overﬁtting.

Alam et al. (2017) extracted 13 statistical features

from vertical ground reaction force (VGRF) and cen-

tre of pressure time-series of 29 Parkinson’s disease

(PD) sufferers and 18 age-matched healthy subjects

and trained several ML models to diagnose Parkin-

sonian gait. Gait of patients afﬂicted by a Parkinson’s

disease was identiﬁed with an accuracy of 85.1%, and

91.6% obtained by k-nn and SVMs with a linear ker-

nel, respectively. However, considering complexity

of human gait, other dimensions of the force vector or

alternative movement descriptions e.g. joint angles,

may contain diagnostically meaningful information.

2.2 Machine Learning-based Detection

of Gait Asymmetry

Most gait symmetry research aims to (1) determine

the magnitude and locations of asymmetries in gait-

impaired subjects as well as (2) investigate if healthy

population exhibits inter-leg differences. In diag-

nosed locomotor dysfunctions, gait asymmetry reduc-

tion is used as a tool attesting to the appropriateness

of rehabilitation regimen (Lewek et al., 2014). The

literature regarding gait asymmetry in able bodied in-

dividuals is equivocal as no signiﬁcant differences in

ground reaction forces (Hamill et al., 1984) or joint

angles (Hannah et al., 1984), are confronted by con-

trary ﬁndings (Vaughan et al., 1992). The inherent

asymmetry of healthy gait is becoming increasingly

evidenced (Sadeghi et al., 2000); although, its cause

is unknown, methods detecting minor deviations in

normal gait should emerge.

2.2.1 Naturally Occurring Gait Asymmetry

Wu and Wu (2015) used SVMs to assess gait sym-

metry. Researchers gathered 10 trials from left and

right legs of 60 elderly individuals. First, the rela-

tive variability of collected parameters (3 peak forces

and corresponding times) was obtained for each leg

by calculating the Coefﬁcient of Variation (CV ≤

12.5%), followed by Absolute Symmetry Index (ASI

≤ 10%), and t-test with a null hypothesis indicat-

ing gait symmetry. The t-test demonstrated signiﬁ-

cant differences (p ≤ 0.05%. ) for 3 out of 6 pa-

rameters. Despite managing to report explicit inter-

limb asymmetries, SI proved inadequate in revealing

more complex relationships. To confront the standard

method a SVM model was developed. The collec-

tive input (120 samples) containing gait variables for

both legs undergone a 6-fold cross-validation result-

ing in the creation of corresponding subsets. Fea-

ture inputs sets comprised of 101 data points vector

of a normalised gait waveform, PCA extracted com-

ponents and 6 aforementioned gait variables with 3

kernel functions (i.e. polynomial, radial basis (RBF)

and linear applied to each). SVM was superior to

SI in determining subtle asymmetries in gait. More-

over, polynomial and RBF outperformed the linear

kernel most signiﬁcantly when carried out on PCA-

based input features. For this conﬁguration (kernel:

Polynomial/RBF, input: PCA-derived), external met-

rics reported were highest with accuracy, sensitivity

and speciﬁcity of 90%, 88-90%, 80%, respectively.

PCA minimised potential redundancies of the origi-

nal 101 dimensions vector, thus boosting classiﬁca-

tion performance; and provided more gait symmetry

information compared to the discretised set of 6 gait

attributes. The ﬁndings can be criticised as the ran-

dom division of the gait pattern pool spread subject-

speciﬁc data across training and testing subsets, hence

the results were affected by the information leakage

(Halilaj et al., 2018).

DATA 2020 - 9th International Conference on Data Science, Technology and Applications

170

2.2.2 Artiﬁcially Induced Gait Asymmetry

Schlaﬂy et al. (2019) aimed to determine the impor-

tance of feature selection for classiﬁcations of poten-

tially overlapping gait pathologies. Artiﬁcially in-

duced locomotion asymmetries interspersed normal

walking trials of 20 healthy subjects. The 10 pertur-

bations resulted from a combinations of shoe heights

(2.7/5.2 cm) and ankle weights (2.3/4.6 kg), and were

grouped into 3 leg length categories (none, small,

large) and 5 distal mass classes: large/small ankle

weights attached to either leg (4 conditions) or a no-

added weight condition (1). The recordings (max.

20 steps for both limbs) provided 21 discrete vari-

ables/step which encapsulated three feature modal-

ities (2 spatio-temporal, 12 kinematic, 7 kinetic).

SVMs, k-NN, and ANNs were trained on all trials be-

longing to 16/20 subjects whereas the data of the 4

remaining participants was used to evaluate the per-

formance of each classiﬁer. For each experimen-

tal condition, classiﬁcation with a PCA-transformed

features was repeated 100 times equating to 4000

tested instances. Subsequently, the results derived

from all 21 features were contrasted with those gen-

erated by possible combinations of 3-element feature

subsets. Overall, SVMs (Gaussian and polynomial

kernels) outperformed other models across all condi-

tions. Lower leg lengths were best classiﬁed by all

21 features (72.9%); the combination of the 3 fea-

ture modalities (64.8%) was more accurate than ki-

netic (61.1%), kinematic (55.9%) and spatio-temporal

(32.9%) results separately. The distal mass was ex-

amined most accurately by kinetic (68%) parame-

ters alone. This outperformed the collective 21 at-

tributes (66.3%), the best performing subset (66.9%)

and other feature modalities (kinematic (43%) and

and spatio-temporal (43.7%)).

Authors emphasised that the most discriminative

variables may be directly dictated by the underlying

cause of the gait impairment. Nonetheless, the study

did not include speed variations which is an important

factor that may inﬂuence gait dynamics and amplify

or nullify inter-limb deviations (Fukuchi et al., 2019).

3 DATASET

The dataset used in this paper containing gait ki-

netics and kinematics of 42 healthy subjects walk-

ing at a range of speeds was developed by Fukuchi

et al. (2018). For each signal, authors speciﬁed the

body side recorded, thereby providing a class label

(left/right) for a gait pattern classiﬁcation.

• Kinetics. Force precedes and causes motion and

the branch of Newtonian mechanics that explores

these vector quantities is known as kinetics. Con-

tact forces must be measured and the measure-

ment relies on Newton’s Third Law of Motion.

Speciﬁcally, by applying the force to the ground

(force platform) at each step, the ground (load

cells) reacts by exerting the ground reaction force

that acts on the body with the same magnitude but

opposite in direction. The raw kinetics comprised

of numerous GRFs traces generated at each step

over the ﬁnal 30 s treadmill-based conditions and

while walking across ﬁve force platforms during

the overground trials.

• Kinematics. On the other hand, kinematics is the

study of motion without regard to causes. The un-

processed kinematics contained 3-D trajectories

of each marker that was afﬁxed to a distinct body

landmark and these were are used to obtain po-

sitions of anatomical segments allowing relevant

variables e.g. joint angles to be computed. As a

result, 3-D joint angles for the hip, knee and ankle

segments were obtained.

Each of the 11 experimental conditions were ex-

pressed as a percentage of the pre-determined com-

fortable speed and reﬂected treadmill (T01 to T08) or

overground (OS, OC, OF) trials. However, to eluci-

date the main trends of speed’s effect on gait asym-

metry, only the results for the three boundary speeds

i.e. slow: T01 (40%), self-selected: T05 (100%) and

fast: T08 (145%) on treadmill are reported.

3.1 Data Processing

To present an overall trend of a signal across mul-

tiple trials (or cycles as in gait) for a given individ-

ual thereby discarding within-subject variance, an en-

semble average was obtained. All ipsilateral gait pat-

terns recorded during a given speed condition were

extracted and linearly interpolated to a 101-point se-

ries. Then, values at each time-point across all derived

gait cycles were averaged yielding three representa-

tive signals (XYZ GRF or joint angle) for each limb

per speed condition/subject.

3.2 Analysis

Initial data exploratory attempts led to a hypothesis

that, in the context of the present dataset, gait sym-

metry might vary as a function of the walking speed.

Also, it can be argued that kinetics might be more

discriminative with respect to limb classiﬁcation than

kinematics, and that walking speed is a potentially im-

portant factor that should be controlled.

Automatic Detection of Gait Asymmetry

171

Figure 1: The shaded area represents the standard deviations across all subjects.

3.2.1 Pooled Analysis of Between-limb

Differences

All overground and most distinct treadmill conditions

(slow, comfortable, fast) of the entire population were

ensemble-averaged into 3-D signals (i.e. ground reac-

tion forces, joint angles) for both limbs, and colour-

coded according to body side. Overall, an inter-limb

similarity was present across the subject pool. In

treadmill trials (see Figure 1), vertical GRF was least

symmetrical at the slowest (T01, 40%) speed (the

red curve protrusion spanned approx. 20-60% of the

gait cycle) yet much more similar at faster conditions

(T05, 100% and T08, 145%). For the overground

walking, curves largely overlapped across all speeds

and the same merely discernible trends characterised

kinematics regardless of speed and walking environ-

ment, hence their visualisations were omitted.

3.2.2 Principal Component Analysis

PCA was applied to standardised kinetic and kine-

matic time-series. Biplots of the ﬁrst two PCs showed

discernible clusters of left/right kinetic gait patterns

for the treadmill walking at slow (T01) and comfort-

able (T05) speeds. No clear groupings emerged for

the fastest treadmill (T08) and overground (b) condi-

tions. PCA-reduced kinematic waveforms provided

lesser insight into gait asymmetry, regardless of the

walking environment.

3.3 Feature Extraction

Classiﬁcation results obtained by using the high-

dimensional gait waveforms were contrasted with its

discretised and reduced (PCA) representations. Each

feature vector was standardised and this was carried

out on each training and test folds during CV cycle.

Mean and variance computed for the training parti-

tion were used to obtain z-scores for the test split.

3.3.1 Baseline Discrete Gait Parameters

Nine parameters corresponding to forces at three dis-

crete instants (see Figure 2) as well as their respec-

tive chronological times of occurrence expressed as a

percentage of the gait cycle were extracted from the

vertical GRF. First and second peaks denote weight

acceptance at heel-strike impact push-off preceding

toe-off, respectively whilst the minimum in-between

refers to the maximum knee ﬂexion prior to forward

propulsion; these points are considered as representa-

tive characteristics reﬂecting an overall quality of gait

(Chao et al., 1983), thus were used as a feature vector

for classiﬁcation.

0 10 20 30 40 50 60 70 80 90 100

Gait Cycle (%)

Normalised Vertical GRF (N/BM)

Vertical GRF (Y)

Limb

left

right

Figure 2: Discrete features extracted from the vertical GRF.

3.3.2 Complete Waveforms

For the kinetics data, all three components (Y - ver-

tical, X - anterior-posterior, Z - medial-lateral) of

the GRF were horizontally concatenated yielding the

2x303 matrix for each subject (2 - left and right limb,

303 - combined GRFs). Given three planes of motion

(XYZ) and three joints (hip, knee, ankle), kinematic

feature matrix was much larger (2x909).

DATA 2020 - 9th International Conference on Data Science, Technology and Applications

172

Dataset

Data

Pre-

Processing

Feature

Extraction

Data Preparation Model Development

3-Fold

Cross-

Validation

Machine

Learning

Algorithm

Hyperparameter

Optimisation

Figure 3: Classiﬁcation Pipeline.

3.3.3 PCA-processed Gait Patterns

PCA was applied to the complete kinetic/kinematic

time-series. Number of principal components used as

an input varied at each CV cycle, and these were pro-

grammatically added until at least 90% of the variance

was explained.

4 AUTOMATIC GAIT

ASYMMETRY DETECTION

The methodology workﬂow is shown in Figure 3.

4.1 Decision Trees

Decision trees were trained by recursive binary strat-

iﬁcation of the predictor space X (root) into a set

of hyper-rectangular regions (inferior nodes) with the

goal to ﬁnd clearer class separation according to the

Information Gain as per equation 1:

∆I(N, A

) = I(N) −

∑

i=1

· I(S

) (1)

where:

• N ⊆ T ← training set (if a root node) or a subset

thereof (if an inferior node)

• A

← logical test that divides N, yielding k candi-

date child nodes S

, S, ..., S

• I ← an arbitrary measure of impurity (e.g. Gini

Index, Entropy)

Recursive grow continued until either a subset

contained only homogenous samples or there was no

further beneﬁt of segmentation in terms of purity. To

optimise the tree, two hyperparameters: 1) minimum

leaf size (1-25) and 2) splitting criterion (Gini Diver-

sity Index, Entropy) were varied.

4.2 k-Nearest Neighbours (k-NN)

k-nearest neighbours uses a proximity measure to

gather the most frequent label of k closest instances in

a local neighbourhood of the unseen record and clas-

siﬁes it accordingly (Hastie et al., 2013).

Two hyperparameters: 1) k-neighbours (1-20) and

2) distance metric (Euclidean, Manhattan) were var-

ied to optimise the model.

4.3 Support Vector Machines (SVM)

SVM is a hyperplane-based discriminator which ad-

dressed limitations of previous approaches such as a

linear separability constraint imposed by the maximal

margin classiﬁer and consolidated the concept of soft

margin introduced in the support vector classiﬁer.

Classiﬁcation was tested with a linear kernel, an

RBF kernel, and a polynomial kernel.

4.4 Artiﬁcial Neural Netoworks (ANN)

Artiﬁcial Neural Networks are non-linear black-box

models and proven to be capable of approximating

any function given that the optimised architecture has

been established (Velten, 2009).

Single-hidden layer feedforward network with a

sigmoid activation function was ﬁtted and trained for

1000 epochs. The number of hidden neurons was var-

ied (1-30) to optimise the network.

5 EXPERIMENTAL SETUP

All the data for a given subject (i.e. two gait patterns)

were placed within either training or a test subset.

Thus, 3-fold cross-validation was carried out divid-

ing 42 subjects into three groups so that at each itera-

tion the model was ﬁtted on a training data comprising

56 gait patterns of 28 subjects (66.6%) and evaluated

on the 28 gait patterns belonging to 14 left out cases

(33.3%).

Automatic Detection of Gait Asymmetry

173

5.1 Evaluation Metrics

To enable comparison between speed conditions and

experimental scenarios, F1 score denoting a weighted

average of precision and recall was chosen to assess

the performance of each of the classiﬁers:

Precision =

T P

T P + FP

(2)

Recall =

T P

T P + FN

(3)

F1 = 2 ∗

Precision ∗ Recall

Precision + Recall

(4)

where:

• TP: cases correctly classiﬁed as a left limb

• TN: cases correctly identiﬁed as a right limb

• FP: cases falsely classiﬁed as left

• FN: cases falsely classiﬁed as right

6 RESULTS

6.1 Kinetics

The results outlined in Tables 1-3 show that on aggre-

gate level (column-wise mean), classiﬁcation perfor-

mance dropped as the walking speed increased with

the highest average F1 score (86.71%) achieved using

discrete kinetic parameters as the feature vector and

this inverse correlation between speed condition and

F1 score was also noted across majority of models and

feature extraction methods.

At the individual model level (row-wise mean),

on average across all speeds ANNs trained on com-

plete kinetic waveforms outperformed other algo-

rithms (83.59%) yet k-NN was superior in identifying

correctly left/right gait patterns at the slowest (T01)

speed condition (91.46%).

Table 1: F1 scores - discrete kinetic parameters.

T01 T05 T08 Mean SD

DT 84.28% 55.72% 62.63% 67.54% 14.90%

k-NN 91.46% 72.08% 66.83% 76.79% 12.97%

SVM 83.88% 79.07% 63.13% 75.36% 10.86%

ANN 87.22% 74.83% 65.65% 75.90% 10.82%

Mean 86.71% 70.42% 64.56% 73.90% 5.24%

SD 3.50% 10.22% 2.01% 5.24% 4.37%

Table 2: F1 scores - complete kinetic waveforms.

T01 T05 T08 Mean SD

DT 87.35% 64.21% 69.46% 73.67% 12.13%

k-NN 84.77% 77.14% 68.89% 76.93% 7.94%

SVM 81.76% 75.08% 71.39% 76.08% 5.26%

ANN 82.46% 90.03% 78.29% 83.59% 5.95%

Mean 84.08% 76.61% 72.01% 77.57% 6.09%

SD 2.53% 10.59% 4.32% 5.81% 4.23%

Table 3: F1 scores - PCA-reduced kinetic waveforms.

T01 T05 T08 Mean SD

DT 76.37% 76.80% 69.22% 74.13% 4.26%

k-NN 82.46% 76.70% 69.94% 76.37% 6.27%

SVM 88.88% 80.08% 68.39% 79.12% 10.28%

ANN 86.32% 77.08% 78.26% 80.55% 5.03%

Mean 83.51% 77.66% 71.45% 77.54% 6.03%

SD 5.44% 1.62% 4.58% 3.88% 2.00%

6.2 Kinematics

As shown in Tables 4-5, in contrary to kinetics-driven

classiﬁcation models’ performance for this modality

was not inversely correlated with the walking speed.

Surprisingly, the highest F1 score (70.77%) was

yielded by k-NN at the fastest condition (T08) yet on

average across all speeds it ranged between 65.33-

68.23%. Hence, joint angles were signiﬁcantly less

discriminative than VGRFs.

Table 4: F1 scores - complete kinematic waveforms.

T01 T05 T08 Mean SD

DT 64.93% 64.89% 68.73% 66.18% 2.21%

k-NN 66.46% 66.07% 66.91% 66.48% 0.42%

SVM 66.67% 66.67% 67.36% 66.90% 0.40%

ANN 64.59% 63.70% 63.82% 64.04% 0.48%

Mean 65.66% 65.33% 66.70% 65.90% 0.72%

SD 1.05% 1.32% 2.07% 1.48% % 0.53

Table 5: F1 scores - PCA-reduced kinematic waveforms.

T01 T05 T08 Mean SD

DT 65.87% 66.70% 65.50% 66.02% 0.61%

k-NN 68.69% 68.34% 70.77% 69.27% 1.31%

SVM 65.76% 66.51% 67.91% 66.73% 1.09%

ANN 66.83% 67.66% 68.75% 67.75% 0.96%

Mean 66.79% 67.30% 68.23% 67.44% 0.73%

SD 1.36% 0.86% 2.18% 1.47% 0.67%

7 DISCUSSION

Human locomotion is elicited through a collaboration

of various systems and it is assumed that gait sym-

DATA 2020 - 9th International Conference on Data Science, Technology and Applications

174

metry can be viewed as the degree of order/disorder

or control of this interplay (Viteckova et al., 2018).

To assess inter-limb differences, gait pattern classiﬁ-

cation was carried out based on kinetic and kinematic

data derived from treadmill and overground walking

at a range of speeds using different ML algorithms

and feature extraction approaches.

Gait asymmetry in healthy subjects was identiﬁed

and generally high classiﬁcation accuracy was noted

for all models, however, k-NN and ANNs on average

outperformed SVMs and decision trees. By examin-

ing F1 maxima from the kinetic-based classiﬁcation,

it can be argued that gait asymmetry gradually dissi-

pated as the walking speed increased. This resonates

with the dynamical systems theory and motor con-

trol research proposing a plausible explanation of an

enhanced movement coordination at faster velocities

of execution. According to these domains, limbs can

be modelled as coupled oscillators that are more eas-

ily controlled at higher movement frequencies whilst

when afforded to decouple at slower speeds, limbs

might employ distinct functional strategies leading to

the hypothesis of a non-dominant/dominant roles of

lower extremities (Gobble et al., 2003).

Another observation was a noticeable drop in clas-

siﬁcation performance when models relied on kine-

matic as opposed to kinetic data. This can be inter-

preted within the context of the study conducted by

Schlaﬂy et al. (2019) in which a spectrum of artiﬁ-

cially induced gait asymmetries was investigated. It

was concluded that the discriminatory power of fea-

tures of various modalities and combinations thereof

may depend on a particular type of gait asymmetry

being examined. Authors demonstrated that features

extracted from the force plate data (kinetic) exclu-

sively, best classiﬁed gait asymmetries caused by dis-

tally afﬁxed mass to either limb. Conversely, when

a limb length and not the mass was manipulated, the

best result was obtained by incorporating both, kinet-

ics and kinematics. Thus, considering the fact that

forces provide an insight into the musculoskeletal dy-

namics that generate the movement and that these

were found to be more effective in differentiating be-

tween left/right gait patterns, it is then possible that

inter-limb differences present in healthy individuals

are manifested in kinetic inequalities.

When applied to the kinetic waveform, PCA had

a varied inﬂuence on the classiﬁcation performance.

For example, SVM and ANN beneﬁted from the di-

mensionality reduction for the slowest walking speed.

On the other hand, it had a negative or no effect on

the k-NN’s and DT’s results overall. As to the hyper-

parameters tested, both distance metrics examined in

k-NN produced similar results whereas the value of k

for the best F1 scores noted ranged predominantly be-

tween 8-14. This parameter is highly dataset-speciﬁc,

however, usually profoundly small values magnify the

effects of outliers whilst increasingly larger settings

yield smoother decision boundaries thereby introduc-

ing bias. In terms of SVMs, its performance was

superior when the linear kernel was adopted whilst

the least favourable results were observed for the rbf

kernel. It is possible that non-linear kernels did not

improve classiﬁcations performance due to already

large dimensionality of the data (Ben-Hur and We-

ston, 2010). To ﬁnd the best decision tree, minimum

number of instances per leaf (MLS) and split crite-

rion (SC) were varied. The MLS imposes a rule on

how many instances in a given node are required to

justify the split, thereby providing a mean to control

a tree depth. In general, the best gait pattern classi-

ﬁcation accuracy was obtained when MLS was kept

either very low (1-4) or just above/below ten (7-13).

With regard to the branching criterion, trends gener-

ated by varying MLS were also remarkably congruent

between the Gini Diversity Index and Entropy. Fi-

nally, potentially due to a small sample size as for

ANNs, the number of hidden neurons did not have

a visible effect on classiﬁcation performance.

8 CONCLUSION

Machine learning and gait analysis co-occur more fre-

quently in the literature than ever before. This study

used an alternative approach to quantify gait asym-

metry in which its assessment was considered as a bi-

nary classiﬁcation problem where a given gait pattern

was assigned to one of the two available categories

namely, left or right limb. Consequently, F1 score

was computed and regarded as a quantitative mea-

sure of gait symmetry with low and high values denot-

ing symmetrical and asymmetrical gaits, respectively.

Future research should also examine ensemble meth-

ods and other cross-validation approaches. Addition-

ally, the role of laterality should be addressed but this

would require a dataset comprising a balanced num-

ber of right- and left- footed subjects. Also, provided

that sufﬁciently large dataset becomes available, so-

phisticated techniques e.g LSTM coupled with ad-

vanced methods to extract discriminative information

from signals such as time-frequency spectral analysis

via the wavelet transform could boost classiﬁcation

performance.

Automatic Detection of Gait Asymmetry

175

REFERENCES

Alam, M. N., Garg, A., Munia, T. T. K., Fazel-Rezai, R.,

and Tavakolian, K. (2017). Vertical ground reaction

force marker for parkinson’s disease. PLOS One,

12(5):p.e0175951.

Beauchet, O., Annweiler, C., Dubost, V., and Allali, G.

(2009). Stops walking when talking: A predictor of

falls in older adults? European Journal of Neurology,

16(7):786–795.

Ben-Hur, A. and Weston, J. (2010). A user’s guide to sup-

port vector machines. In Carugo, O. and Eisenhaber,

F., editors, Data Mining Techniques for the Life Sci-

ences, pages 223–239. Humana Press.

Chao, Y., Laughman, R., Schneider, E., and Stauffer, R.

(1983). Normative data of knee joint motion and

ground reaction forces in adult level walking. Jour-

nal of Biomechanics, 16(3):219–233.

Fukuchi, C. A., Fukuchi, R. K., and Duarte, M. (2018).

A public dataset of overground and treadmill walking

kinematics and kinetics in healthy individuals. PeerJ,

6:e4640.

Fukuchi, C. A., Fukuchi, R. K., and Duarte, M. (2019). Ef-

fects of walking speed on gait biomechanics in healthy

participants: a systematic review and meta-analysis.

Systematic Reviews, 8(1):153.

Gobble, D. J., Marino, W. G., and Potvin, J. R. (2003). The

inﬂuence of horizontal velocity on interlimb symme-

try in normal walking. Human Movement Science,

22(3):271–283.

Halilaj, E., Rajagopal, A., Fiterau, M., Hicks, J. L., Hastie,

T. J., and Delp, S. L. (2018). Machine learning in

human movement biomechanics. Journal of biome-

chanics, 81:1–11.

Hamill, J., Bates, B. T., and M.Knutzen, K. (1984). Ground

reaction force symmetry during walking and run-

ning. Research Quarterly for Exercise and Sport,

55(3):289–293.

Hannah, R., Morrison, J., and Chapman, A. (1984). Kine-

matic symmetry of the lower limbs. Archives of phys-

ical medicine and rehabilitation, 65(4):155–158.

Hastie, T., Tibshirani, R., Witten, D., and James, G. (2013).

An Introduction to Statistical Learning: With Applica-

tions in R. Springer.

Inman, V. and Eberhart, H. (1953). The major determinants

in normal and pathological gait. The Journal of Bone

& Joint Surgery, 35(3):543–558.

Kutilek, P., Viteckova, S., Svoboda, Z., and Socha, V.

(2014). Kinematic quantiﬁcation of gait asymmetry

based on characteristics of angle-angle diagrams. Acta

Polytechnica Hungarica, 11(5):25–28.

Lewek, M., Bradley, C. E., Wutzke, C., and Zinder, S. M.

(2014). The relationship between spatiotemporal gait

asymmetry and balance in individuals with chronic

stroke. Journal of Applied Biomechanics, 30(1):31–

36.

Mezghani, N., Dunbar, M. J., Ouakrim, Y., and Fuentes, A.

(2016). Biomechanical signal classiﬁcation of surgi-

cal and non-surgical candidates for knee arthroplasty.

In 2016 International Symposium on Signal, Image,

Video and Communications, pages 287–290.

Muro, A., Zapirain, B. G., and Mendez-Zorrilla, A. (2014).

Gait analysis methods: An overview of wearable and

non-wearable systems, highlighting clinical applica-

tions. Sensors, 14(2):3362–3394.

Robinson, R. H., Herzog, W., and Nigg, B. M. (1987). Use

of force platform variables to quantify the effects of

chiropractic manipulation on gait symmetry. Jour-

nal of Manipulative and Physiological Therapeutics,

10(4):172–176.

Sadeghi, H., Allard, P., Prince, F., and Labelle, H. (2000).

Symmetry and limb dominance in able-bodied gait: a

review. Gait & Posture, 12(1):34–45.

Salazar, A. J., Cuevas, O. C. D. C., and Bravo, R. J. (2004).

Novel approach for spastic hemiplegia classiﬁcation

through the use of support vector machines. In The

26th Annual International Conference of the IEEE

Engineering in Medicine and Biology Society, pages

466–469.

Schlaﬂy, M., Yilmaz, Y., and Reed, K. B. (2019). Feature

selection in gait classiﬁcation of leg length and distal

mass. Informatics in Medicine Unlocked, 15:100163.

Vaughan, C. L., Davis, B. L., and O’Connor, J. C. (1992).

Dynamics of human gait. Human Kinetics Publishers.

Velten, K. (2009). Mathematical Modeling and Simulation:

Introduction for Scientists and Engineers. John Wiley

& Sons.

Verghese, J., Wang, C., Lipton, R., and Holtzer, R. (2007).

Quantitative gaitdysfunction and risk of cognitive de-

cline and dementia. Journal of neurology, neuro-

surgery, and psychiatry, 78(9):929–935.

Viteckova, S., Kutilek, P., Svoboda, Z., and Krupicka, R.

(2018). Gait symmetry measures: A review of cur-

rent and prospective methods. Journal of the Indian

Medical Association, 42:89–100.

Wu, J. and Wu, B. (2015). The novel quantitative technique

for assessment of gait symmetry using advanced sta-

tistical learning algorithm. BioMed Research Interna-

tional, 2015(528971).

Zhang, J., Zhang, K., Feng, J., and Small, M. (2010).

Rhythmic dynamics and synchronization via dimen-

sionality reduction: application to human gait. PLOS

Computational Biology, 6(12):p.e1001033.

DATA 2020 - 9th International Conference on Data Science, Technology and Applications

176