![Representation of the tracked joints and anatomical landmarks (colored dots) for the calculated joint angles (α: shoulder flexion/extension; β: elbow flexion/extension); γ: shoulder abduction/adduction; and δ: lateral trunk sway) in (A) sagittal plane and (B) frontal plane. Part (A) is a modified and updated version of the one presented in Eken et al. 2020 [19].](https://www.researchgate.net/publication/390964345/figure/fig2/AS:11431281402331890@1745599867392/Representation-of-the-tracked-joints-and-anatomical-landmarks-colored-dots-for-the_Q320.jpg)
![Joint angles in sagittal and frontal planes presented for children with HIVE (red) and TD children (green) (thick lines = median; dashed lines = interquartile ranges (1st and 3rd percentile)) over the gait cycle. Parts A and B are modified and updated versions of those presented in Eken et al. 2020 [19].](https://www.researchgate.net/publication/390964345/figure/fig3/AS:11431281402351747@1745599868393/Joint-angles-in-sagittal-and-frontal-planes-presented-for-children-with-HIVE-red-and-TD_Q320.jpg)
Access to this full-text is provided by MDPI.
Content available from Applied Sciences
This content is subject to copyright.
Academic Editor: Arkady Voloshin
Received: 27 November 2024
Revised: 7 April 2025
Accepted: 19 April 2025
Published: 20 April 2025
Citation: Eken, M.M.; Meyns, P.;
Lamberts, R.P.; Langerak, N.G.
Markerless Upper Body Movement
Tracking During Gait in Children with
HIV Encephalopathy: A Pilot Study.
Appl. Sci. 2025,15, 4546. https://
doi.org/10.3390/app15084546
Copyright: © 2025 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license
(https://creativecommons.org/
licenses/by/4.0/).
Article
Markerless Upper Body Movement Tracking During Gait in
Children with HIV Encephalopathy: A Pilot Study
Maaike M. Eken 1, 2, † , Pieter Meyns 3, *,† , Robert P. Lamberts 4and Nelleke G. Langerak 5,6
1Institute of Sport and Exercise Medicine (ISEM), Department of Exercise, Sport and Lifestyle Medicine,
Faculty of Medicine and Health Sciences, Stellenbosch University, Tygerberg, Cape Town 7505, South Africa;
meken@sun.ac.za
2Institute of Orthopaedics and Rheumatology, Mediclinic Winelands Orthopaedic Hospital, Stellenbosch,
Cape Town 7600, South Africa
3REVAL Rehabilitation Research Center, Faculty of Rehabilitation Sciences, Hasselt University,
3500 Hasselt, Belgium
4
Department of Human Movement Sciences, University Medical Center Groningen, University of Groningen,
9713 Groningen, GZ, The Netherlands
5Neuroscience Institute and Division of Neurosurgery, Faculty of Health Sciences, University of Cape Town,
Cape Town 7935, South Africa; nellekelangerak@hotmail.com
6Department of Research, Sint Maartenskliniek, 6532 Nijmegen, SZ, The Netherlands
*Correspondence: pieter.meyns@uhasselt.be
†These authors contributed equally to this work—shared first authorship.
Abstract: The aim of this pilot study was to investigate the feasibility of markerless tracking
to assess upper body movements of children with and without human immunodeficiency
virus encephalopathy (HIV-E). Sagittal and frontal video recordings were used to track
anatomical landmarks with the DeepLabCut pre-trained human model in five children with
HIV-E and five typically developing (TD) children to calculate shoulder flexion/extension,
shoulder abduction/adduction, elbow flexion/extension and trunk lateral sway. Differ-
ences in joint angle trajectories of the two cohorts were investigated using a one-dimensional
statistical parametric mapping method. Children with HIV-E showed a larger range of
motion in shoulder abduction and trunk sway than TD children. In addition, they showed
more shoulder extension and more lateral trunk sway compared to TD children. Markerless
tracking was feasible for 2D movement analysis and sensitive to observe expected differ-
ences in upper limb and trunk sway movements between children with and without HIVE.
Therefore, it could serve as a useful alternative in settings where expensive gait laboratory
instruments are unavailable, for example, in clinical centers in low- to middle-income
countries. Future research is needed to explore 3D markerless movement analysis systems
and investigate the reliability and validity of these systems against the gold standard 3D
marker-based systems that are currently used in clinical practice.
Keywords: machine learning; deep learning; clinical gait analysis; arm swing; posture;
trunk; DeepLabCut
1. Introduction
Human movement analysis is imperative for tailoring evidence-based interventions
in individuals with and without a physical disability [
1
]. Two-dimensional movement
analysis has been used to assist in clinical decision making for the past decades [
2
]. This
process, however, requires researchers and/or clinicians to track defined markers on a
frame-to-frame basis using software packages, for example Kinovea [
3
,
4
], which is time-
consuming, and trained personnel are needed. In addition, skin markers are not always
Appl. Sci. 2025,15, 4546 https://doi.org/10.3390/app15084546
Appl. Sci. 2025,15, 4546 2 of 12
comfortable for individuals, particularly for children with or without disabilities, who
might find them interfering with their normal gait.
Quantitative 2D gait analysis is, however, particularly beneficial in low- to middle-
income countries (LMICs) since it is less expensive and easily accessible. An important
application of 2D movement analysis in LMICs has been in children with physical disabili-
ties, such as human immunodeficiency virus (HIV) encephalopathy (HIVE) [
5
]. HIVE is
a childhood disability particularly observed in LMICs such as South Africa. Most of the
children who suffer from HIV worldwide (approximately 1.8 million children) reside in
African sub-Saharan LMICs, such as South Africa [
6
,
7
]. Most children acquire HIV through
vertical transmission, including the transmission of the virus in utero, at birth, or through
breastfeeding. As it can affect the developing fetal and infant brain, the virus can induce
the neurological complication HIVE [
8
], which results in an impaired gait pattern [
9
] and
upper limb function [
10
]. To improve clinical care for children with HIVE, accessible, af-
fordable and easy-to-use tools are imperative to allow for tailored clinical decision making
in non-specialized, smaller clinical centers in remote, less populated areas in LIMCs.
Recent advances in biomedical engineering have resulted in new techniques based
on deep learning to track body landmarks in simple video recordings. Deep learning
approaches for human body landmark tracking in video recordings typically employ con-
volutional neural networks and pose estimation architectures that process sequential frames
to identify and localize anatomical key points. These models are trained on large datasets
of annotated images to learn hierarchical visual features that represent body parts and their
spatial relationships, enabling them to detect joint positions even under varying lighting
conditions, partial occlusions, and complex backgrounds. The use of this technique includes
a high degree of automation and allows for recordings in a natural
environment [8,11]
.
Cronin described the potential of markerless tracking using neural networks in the field of
human movement
science [8]
. More recently, DeepLabCut was introduced as a free and
open-source toolbox to track user-defined features in video
files [12,13]
. A great advan-
tage of DeepLabCut over other pose estimation software packages is that only a limited
number of labeled frames are required to obtain deep neural networks that match human
labeling accuracy [
8
]. Furthermore, DeepLabCut provides exceptional flexibility with a
user-friendly interface that allows researchers to define custom key points specific to their
research questions, rather than being limited to predefined landmark sets common in
other platforms.
An assessment tool based on markerless tracking of limb movement during gait may
provide a solution for cost-efficient movement analysis in LMICs, especially with a free
and open-source toolbox such as DeepLabCut. Therefore, the current study has several
aims. (1) It is important to investigate the feasibility of markerless tracking during gait in
a population specific to LMICs, such as children with HIVE. (2) We will assess whether
markerless tracking during gait is sensitive to observe expected differences in gait between
children with and without HIVE. (3) Given that previous research has focused primarily
on lower limb function during gait, and no research is available on upper body function
during gait of children with HIVE, we will specifically investigate the differences in upper
limbs between both groups. It is important to note that our previous work showed that
impaired upper body movements can have an effect on gait performance, which could be
related to gait instability, as previously shown in children with cerebral palsy (CP) [
14
–
17
].
Based on previous studies on lower limb function during gait, it is expected to also
observe differences in upper body function during gait between children with and without
HIVE. We hypothesize that the arm swing during gait in children with HIVE will show
alterations similar to children with CP, given their supraspinal/cortical neural impairments.
Therefore, the aim of this pilot study was to investigate the feasibility of using DeepLabCut
Appl. Sci. 2025,15, 4546 3 of 12
to assess upper body movements in children with and without HIVE, and whether these
expected differences could be quantified with DeepLabCut markerless tracking between
children with and without HIVE.
2. Materials and Methods
2.1. Presentation of Preliminary Data
A preliminary part of the data of this article was accepted as a short conference abstract
and presented at the European Society for Movement Analysis in Adults and Children
(ESMAC) Conference [VirtualESMAC 29th Annual Meeting on 17 September 2020] [authors:
M.M. Eken, P. Meyns, R.P. Lamberts, N.G. Langerak]. Markerless movement tracking using
a machine learning algorithm to assess arm movements during gait in children with HIV
encephalopathy, Gait & Posture 81 (2020) 85–86 [18].
2.2. Participants
This case-control study comprises a subset of a study previously published [
19
].
Children with HIVE were included based on the following inclusion criteria: (1) diagnosis
of HIVE according to Centers for Disease Control criteria [
20
]; (2) aged 5 to 12 years;
(3) girls
; (4) Gross Motor Function Classification System (GMFCS) [
21
] level II. Exclusion
criteria were: (1) significant prematurity (birth weight of
≤
2.0 kg and/or a gestational
age of
≤35 wks
); (2) additional neuromuscular or central nervous system disorders (e.g.,
tuberculosis meningitis);
(3) botulinum
neurotoxin treatment during the last 6 months;
(4) surgery
on the lower limbs during the last 12 months. Typically developing (TD)
children were included and matched based on gender and age. Given the goal of the
current study being to provide a proof-of-concept, a small but deliberate homogeneous
group of children (with fewer possible confounding factors) was recruited, resulting in the
inclusion of only children from one gender (in this case girls).
All the possible risks and benefits were explained to the participants and their care-
givers. Written informed consent was obtained from the caregivers before data collection,
and assent was obtained from the participants. Ethical approval for the study was granted
by the Human Research Ethics Committee of the University of Cape Town, South Africa,
while the study also followed the principles set out in the Declaration of Helsinki [22].
2.3. Assessment Procedures
The movement analyses were conducted in the Neuromechanics Laboratory of the
Central Analytics Facilities, Stellenbosch University, Tygerberg, Cape Town, South Africa,
which is a setting designed for clinical gait analysis using standard lighting with no
interfering daylight changes. All the participants were asked to walk at a self-selected
speed on a 10-m runway. Sagittal and frontal video recordings were obtained using static
Bonita video cameras (sampling frequency: 50 Hz). Only good-quality color videos were
used for the analyses.
Joint poses were tracked with the pre-trained human model of the DeepLabCut toolbox
(ResNet101 [
13
]) in both sagittal (view on the left-hand side; Figure 1) and frontal plane
(view on the front side). DeepLabCut is a free and open-source toolbox that can be used for
markerless pose estimation in a number of species and behaviors, in different situations
and diverse backgrounds. DeepLabCut is based on machine-learning algorithms with deep
neural networks programmed in Python 3, and pre-trained models are developed both
for humans and other animal species [
13
]. The following joints and anatomical landmarks
were tracked in the sagittal plane: left shoulder, left elbow, left wrist and left hip; and
in the frontal plane: chin, right and left hip, left shoulder and left elbow. The tracking
of joints and anatomical landmarks was visually inspected. In the case of inaccurate
Appl. Sci. 2025,15, 4546 4 of 12
marker tracking (pixel accuracy was above the accuracy threshold of 5 pixels), additional
frames were extracted and labeled manually, according to the protocol described by Mathis
et al. [
12
]. First, outliers were extracted to remove labels with a likelihood below 0.9. Second,
20 frames
were extracted from various participants and manually labeled. The data sets of
the original ResNet101 model and additionally labeled frames were then merged, and the
model was retrained until the loss plateaued [
12
]. No changes were made to the frames
regarding occlusions, motion blur or inconsistencies in frame quality.
Appl.Sci.2025,15,xFORPEERREVIEW4of12
anatomicallandmarksweretrackedinthesagialplane:leftshoulder,leftelbow,left
wristandlefthip;andinthefrontalplane:chin,rightandlefthip,leftshoulderandleft
elbow.Thetrackingofjointsandanatomicallandmarkswasvisuallyinspected.Inthecase
ofinaccuratemarkertracking(pixelaccuracywasabovetheaccuracythresholdof5pix-
els),additionalframeswereextractedandlabeledmanually,accordingtotheprotocol
describedbyMathisetal.[12].First,outlierswereextractedtoremovelabelswithalike-
lihoodbelow0.9.Second,20frameswereextractedfromvariousparticipantsandmanu-
allylabeled.ThedatasetsoftheoriginalResNet101modelandadditionallylabeled
frameswerethenmerged,andthemodelwasretraineduntilthelossplateaued[12].No
changesweremadetotheframesregardingocclusions,motionblurorinconsistenciesin
framequality.
Figure1.Representationofmarkerlessmovementtrackingusingthepre-trainedhumanmodelin
DeepLabCut.
Thetrajectoriesofjointlocationsandanatomicallandmarksweresubsequentlyim-
portedintoMATLAB(R2017b,MathworksInc.,Natick,MA,USA)andincludedwhen
thelikelihoodwasabove0.9.Thepositionsofthemarkerswereusedtocalculatejoint
anglesusingthecosinerule(Figure2):
𝑎𝑛𝑔𝑙𝑒 𝛼, 𝛽, 𝛾, 𝛿𝑎𝑐𝑜𝑠𝑎𝑏
𝑐
2𝑎𝑏
Theshoulderangle(α)inthesagialplane(i.e.,shoulderflexion/extension)wascal-
culatedastheanglebetweenthelinefromthemid-shoulderpointtothehipmarkerand
thelinefromthemid-shoulderpointtotheelbowmarker.Theelbowangle(β)inthesag-
ialplane(i.e.,elbowflexion/extension)wascalculatedastheanglebetweenthelinefrom
themid-shoulderpointtotheelbowmarkerandthelinefromtheelbowtothewrist
marker.Theshoulderangle(γ)inthefrontalplane(i.e.,shoulderabduction/adduction)
wascalculatedastheanglebetweenthelinefromthemid-shoulderpointtotheelbow
markerandthelinefromthechintothemiddleofthetwohipmarkers.Thetrunkangle
(δ)inthefrontalplane(i.e.,trunklateralsway)wascalculatedastheanglebetweenthe
linefromthemid-shoulderpointtothemiddleofthetwohipmarkersandthevertical.In
theformula,aandbrepresentthesegmentsthatformtheangle,whilecrepresentsthe
segmentoppositetotheangle.Theangleswerecalculatedperframe,andtheangle
Figure 1. Representation of markerless movement tracking using the pre-trained human model in
DeepLabCut.
The trajectories of joint locations and anatomical landmarks were subsequently im-
ported into MATLAB (R2017b, Mathworks Inc., Natick, MA, USA) and included when the
likelihood was above 0.9. The positions of the markers were used to calculate joint angles
using the cosine rule (Figure 2):
angle (α,β,γ,δ)=acos a2+b2−c2
2ab
The shoulder angle (
α
) in the sagittal plane (i.e., shoulder flexion/extension) was
calculated as the angle between the line from the mid-shoulder point to the hip marker
and the line from the mid-shoulder point to the elbow marker. The elbow angle (β) in the
sagittal plane (i.e., elbow flexion/extension) was calculated as the angle between the line
from the mid-shoulder point to the elbow marker and the line from the elbow to the wrist
marker. The shoulder angle (
γ
) in the frontal plane (i.e., shoulder abduction/adduction)
was calculated as the angle between the line from the mid-shoulder point to the elbow
marker and the line from the chin to the middle of the two hip markers. The trunk angle
(
δ
) in the frontal plane (i.e., trunk lateral sway) was calculated as the angle between the
line from the mid-shoulder point to the middle of the two hip markers and the vertical.
In the formula, aand brepresent the segments that form the angle, while crepresents
the segment opposite to the angle. The angles were calculated per frame, and the angle
trajectories were subsequently filtered using a Butterworth filter (fc = 10 Hz). Per video,
strides were manually extracted (heel strike to heel strike) along with the moment of
toe-off. After time normalization (0–100% gait cycle), the average trajectories of the three
Appl. Sci. 2025,15, 4546 5 of 12
strides were calculated per participant. Subsequently, the joint angle trajectories and the
range of motion (ROM) for shoulder flexion/extension, elbow flexion/extension, shoulder
abduction/adduction and trunk sway were determined and compared between children
with HIVE and TD children.
Appl. Sci. 2025, 15, x https://doi.org/10.3390/xxxxx
Figure 2. Representation of the tracked joints and anatomical landmarks (colored dots) for the cal-
culated joint angles (α: shoulder flexion/extension; β: elbow flexion/extension); γ: shoulder abduc-
tion/adduction; and δ: lateral trunk sway) in (A) sagial plane and (B) frontal plane. Part (A) is a
α
β
δ
γ
A
B
Figure 2. Representation of the tracked joints and anatomical landmarks (colored dots) for the
calculated joint angles (
α
: shoulder flexion/extension;
β
: elbow flexion/extension);
γ
: shoulder
abduction/adduction; and
δ
: lateral trunk sway) in (A) sagittal plane and (B) frontal plane. Part (A)
is a modified and updated version of the one presented in Eken et al. 2020 [19].
2.4. Statistical Analysis
The participants’ characteristics were presented using descriptive statistics and indi-
vidual data points. Normality was tested using Shapiro–Wilk tests. Given the small sample
size and as outcomes were not normally distributed, median values and interquartile
ranges (IQR) were determined. Differences in the ROM parameters between the children
with HIVE and their TD peers were investigated using non-parametric Mann–Whitney
U-tests. The joint angle trajectories of the two cohorts were investigated using the open-
source 1-dimensional statistical parametric mapping method (1DSPM) in MATLAB (R2017b,
Mathworks Inc., Natick, MA, USA). This is an open-source statistical method that allows
analysis of continuous biomechanical data (such as joint angles) across an entire time series
rather than at discrete points, enabling the identification of statistically significant differ-
ences between conditions or groups throughout the complete movement cycle. Herewith,
the non-parametric method was applied to compare the joint angle trajectories (vM0.1,
www.spm1d.org: download non-parametric toolbox), and median and IQR of the joint
angle trajectories were calculated and presented in figures (0–100% gait cycle). Significance
was set at p< 0.05 for all statistical analyses.
3. Results
3.1. Participants’ Background
Five girls with HIVE, classified as GMFCS level II, and five TD girls were included in
the study. The participants’ demographics are presented in Table 1. All the children with
HIVE were on antiretroviral therapy.
Appl. Sci. 2025,15, 4546 6 of 12
Table 1. Demographic characteristics of the participants with HIVE and TD participants.
No. Age
(Years)
BMI
(kg/m2)
BMI
Category
Videos Included (n)
Sagittal Frontal
HIVE
1 8.3 17.9 normal 3 3
2 10.8 21.7 overweight 3 3
3 10.5 15.9 normal 3 3
4 9.4 18.0 normal 3 2
5 12.8 22.5 overweight 3 3
TD
1 6.8 13.9 normal 3 3
2 7.9 13.6 normal 3 3
3 10.8 17.8 normal 3 2
4 10.8 15.0 normal 3 3
5 8.2 19.5 overweight 3 3
Abbreviations: HIVE, human immunodeficiency virus encephalopathy; TD, typically developing; and BMI, body
mass index.
3.2. Accuracy of Pre-Trained Human Network (ResNet101)
In the frontal plane, the pre-trained human network showed accurate tracking of the
movement without additional training of the network (pixel accuracy below 5 pixels). In the
sagittal plane, the pre-trained human network did not show accurate tracking of the move-
ments. After manually labeling 20 frames and retraining the network (
40.000 iterations
), the
training loss plateaued, resulting in a training error of 2.8 pixels and test error of
3.1 pixels
.
3.3. Arm Swing and Trunk Sway
The ROM per joint angle is reported in Table 2and shows that the ROM in shoulder
abduction of children with HIVE was significantly larger than TD children. In addition, the
ROM in trunk sway was significantly larger in children with HIVE than TD children. The
ROM for shoulder flexion/extension and for elbow flexion/extension in the sagittal plane
did not show significant differences between children with HIVE and TD children.
Table 2. Range of motion of upper limb movements in sagittal and frontal planes for HIVE and
TD children.
HIVE TD
Median IQR Median IQR pValue
Sagittal plane
Shoulder
flexion/extension (
◦
)
31.0 [24.4–51.5] 37.2 [25.2–61.8] 0.602
Elbow
flexion/extension (
◦
)
25.4 [20.1–38.7] 30.7 [24.8–36.1] 0.602
Frontal plane
Shoulder abduc-
tion/adduction (◦)41.5 [11.1–49.2] 6.5 [4.5–10.4] 0.028 *
Trunk sway (◦) 14.7 [8.5–24.7] 1.9 [1.6–2.9] 0.009 *
* Significantly different: p< 0.05.
The trajectories of the shoulder and elbow angles obtained in the sagittal plane, as
well as the trajectories of the shoulder and trunk angles obtained in the frontal plane, are
presented in Figure 3. Significant differences between the children with HIVE and the TD
children were observed in the shoulder angle in the sagittal plane between 20% and 59%
of the gait cycle (
p= 0.004
), showing that the shoulder angle of the children with HIVE
Appl. Sci. 2025,15, 4546 7 of 12
remained in extension, while the shoulder angle of the TD children moved to flexion during
the late stance phase. In the frontal plane, significant differences were observed between
the children with HIVE and the TD children in the shoulder angle (80–93% gait cycle;
p= 0.040
) and the trunk angle (44–65% gait cycle; p= 0.004). Children with HIVE showed
significantly more shoulder abduction during the mid-swing phase and lateral trunk sway
during late stance and early swing, while TD children showed overall minimal shoulder
abduction and trunk sway.
4. Discussion
The aim of this study was to investigate the feasibility of using markerless tracking
to assess upper body, including upper limb and trunk, movements in children with and
without HIVE, and whether expected differences could be quantified between children
with and without HIVE. The findings of this study indicate that it is feasible to obtain 2D
joint angles from videos using markerless tracking with DeepLabCut in children with and
without HIVE, and to obtain differences between children with and without HIVE in upper
limb and trunk movements during gait.
This study explored the potential of markerless movement tracking with DeepLabCut
software to use as a tool for clinical gait analysis in LMIC settings. The results of the study
showed that DeepLabCut can successfully be used to obtain 2D gait analysis, which can
save valuable time for clinical research. This observation confirms suggestions made by
Cronin (2021), who described the potential value of pose estimation software packages
for human movement analysis [
8
]. The implementation of markerless movement tracking
systems in clinical gait analysis can serve as a useful alternative in settings where expensive
gait laboratory instruments are unavailable, particularly in LMICs. In many clinical settings
in LMICs, there are no resources available (too expensive) for 3D gait analysis systems,
while resources are often limited, and only 2D gait analysis systems are available in high-
income countries as well.
It is important to note that the pre-trained human model (ResNet101), without ad-
ditional manual labeling of frames, did not achieve sufficient accuracy to track markers
in the sagittal plane, which could be explained by the fact that only children with and
without disabilities were included, instead of adults, suggesting that both children and
specifically children with motor disorders were most likely less or not present in the large
datasets of annotated images to learn hierarchical visual features that represent body
parts from which the anatomical joints are detected. Therefore, for the implementation of
DeepLabCut in clinical settings, researchers are advised to manually label 20–50 frames
with user-identified/user-defined markers specifically applicable for the laboratory and
retrain the ResNet101 model to extract body landmarks from new videos automatically,
especially when being used for pathological patient groups that have most likely not been
included in the training models of the pretrained human model (ResNet101).
In line with our hypothesis, which was based on a clinical study [
10
], differences
in upper limb and trunk movements between children with HIVE and spastic diplegia
and their TD peers were observed. The results showed that the shoulder movements of
children with HIVE obtained in both sagittal and frontal plane deviated from TD children,
showing more shoulder extension from mid-stance to late stance in the sagittal plane
and more shoulder abduction as indicated by a higher overall ROM, as well as more
abduction during the swing phase in the frontal plane (Figure 3). Trunk sway was also
considerably more pronounced in children with HIVE compared to the TD, showing more
lateral sway at initial contact, early stance and late swing. Both the deviated shoulder
and trunk movements observed in children with HIVE are consistent with upper body
movements that have previously been observed in children with spastic diplegia resulting
Appl. Sci. 2025,15, 4546 8 of 12
from CP [
23
], individuals after stroke [
24
–
26
] and individuals with Parkinson’s disease, who
show deviations in trunk sway only [
27
]. These altered movements observed during gait
in individuals with different neuromuscular conditions have been suggested to contribute
to compensating for gait instability by helping to maintain balance, redistribute weight
or adjust for deficits in lower limb control [
14
–
17
]. In some cases, excessive trunk sway
or shoulder deviations may serve as a strategy to counteract muscle weakness, impaired
coordination, or reduced proprioception, ultimately influencing overall gait efficiency
and energy expenditure. In addition, several children with HIVE appeared to show two
movements towards shoulder abduction during the gait cycle. These two ‘bumps’ (Figure 3),
which occurred during late stance (around 40% of gait cycle) and late swing (approximately
85% of gait cycle), could be related to the spastic gait pattern, as also shown in children
with spastic CP during gait [
28
,
29
]. A double bump pattern in the arm swing during one
gait cycle (i.e., stride) is often seen in the sagittal plane in healthy adults when walking
very slowly [
30
]. In patients with neurologic disorders, however, this double bump pattern
has been found to be present independent of the gait speed [
31
,
32
]. In patients with stroke,
it was found that the group with this altered arm swing coordination during gait showed
a greater level of upper limb impairment (reduced Fugl-Meyer upper extremity score)
and more spasticity in the internal shoulder rotators and elbow extensors, suggesting that
neuromechanical control might be an important factor contributing to the altered arm-to-leg
coordination pattern [
32
]. This should be further investigated in a larger sample of children
with HIVE and spastic diplegia.
Appl.Sci.2025,15,xFORPEERREVIEW9of12
Figure3.JointanglesinsagialandfrontalplanespresentedforchildrenwithHIVE(red)andTD
children(green)(thicklines=median;dashedlines=interquartileranges(1stand3rdpercentile))
overthegaitcycle.PartsAandBaremodifiedandupdatedversionsofthosepresentedinEkenet
al.2020[19].
Inthisstudy,theDeepLabCutpre-trainedhumanmodel(ResNet101)wasused[12].
Inthefrontalplane,trackingthepredefinedjointsandanatomicallandmarkswassuita-
ble,showingsufficientaccuracy.Oneofthelimitations,however,wasthatinthesagial
plane,trackingaccuracywaslimitedusingthepre-trainedhumanmodel,requiringre-
trainingofthenetworkwithadditionalself-labeledframes.Thiscouldbeexplainedby
thepre-trainedhumanmodel,whichwasbasedontypicaladultgait,where,inthisstudy,
childrenwithanimpairedgaitpaernwereincludedaswellasTDchildren.Itistherefore
importanttonotethatadditionaltraining,e.g.,onframesofindividualswithsimilarcon-
ditionsorgaitcomplications,orfine-tuningofpre-trainedhumanmodels,isnecessary
beforeimplementingmarkerlessposeestimationforclinicalgaitanalysisforindividuals
withatypicalgaitsandcompensatoryupperbodymovements.Previousresearchhas
shownthatmanualannotationofkeypointsinhumanmovementisavalidandreliable
methodintypicallydevelopingpopulations[33–38].Furthermore,arecentstudyshowed
thevalidityof2DmarkerlesstrackingusingDeepLabCutcomparedtogoldstandardla-
boratory-basedoptoelectronicthree-dimensionalmotioncapturetomeasurejointangles
inchildren[39].However,tothebestofourknowledge,nostudieshaveinvestigatedthe
reliabilityandvalidityof2Dmanualannotationforhumanmovementinindividualswith
atypicalgaits,whichisalimitationofthisstudy.Anotherlimitationofthestudywasthat
2Dmovementanalysiswasused,while3Dmovementanalysisprovidesessentialinfor-
mationrequiredtoprescribesuitabletreatmentinclinicalseings[1].Currently,3D
marker-basedmovementanalysissystemsthatusestereophotogrammetrymethods,such
asVicon,OptotrackandBTSsystems,areacknowledgedasthegoldstandard[40].How-
ever,thesesystemsarecurrentlyexpensive,requireadedicatedspaceandneedtrained
Figure 3. Joint angles in sagittal and frontal planes presented for children with HIVE (red) and TD
children (green) (thick lines = median; dashed lines = interquartile ranges (1st and 3rd percentile))
over the gait cycle. Parts A and B are modified and updated versions of those presented in Eken et al.
2020 [19].
In this study, the DeepLabCut pre-trained human model (ResNet101) was used [
12
].
In the frontal plane, tracking the predefined joints and anatomical landmarks was suitable,
Appl. Sci. 2025,15, 4546 9 of 12
showing sufficient accuracy. One of the limitations, however, was that in the sagittal plane,
tracking accuracy was limited using the pre-trained human model, requiring retraining of
the network with additional self-labeled frames. This could be explained by the pre-trained
human model, which was based on typical adult gait, where, in this study, children with
an impaired gait pattern were included as well as TD children. It is therefore important
to note that additional training, e.g., on frames of individuals with similar conditions
or gait complications, or fine-tuning of pre-trained human models, is necessary before
implementing markerless pose estimation for clinical gait analysis for individuals with
atypical gaits and compensatory upper body movements. Previous research has shown
that manual annotation of key points in human movement is a valid and reliable method in
typically developing populations [
33
–
38
]. Furthermore, a recent study showed the validity
of 2D markerless tracking using DeepLabCut compared to gold standard laboratory-based
optoelectronic three-dimensional motion capture to measure joint angles in children [
39
].
However, to the best of our knowledge, no studies have investigated the reliability and
validity of 2D manual annotation for human movement in individuals with atypical gaits,
which is a limitation of this study. Another limitation of the study was that 2D movement
analysis was used, while 3D movement analysis provides essential information required to
prescribe suitable treatment in clinical settings [
1
]. Currently, 3D marker-based movement
analysis systems that use stereophotogrammetry methods, such as Vicon, Optotrack and
BTS systems, are acknowledged as the gold standard [
40
]. However, these systems are
currently expensive, require a dedicated space and need trained personnel to be successfully
employed, which limits accessibility, particularly for LMICs. Three-dimensional markerless
movement analysis systems that use accessible cameras can provide solutions for these
regions specifically. Future research is needed to explore 3D markerless movement analysis
systems and investigate the reliability and validity of these systems against gold standard
3D marker-based systems. However, this comes with its challenges due to different factors,
such as constant updates to deep learning models, occlusion of body landmarks and
segments and high-resolution cameras, which are needed to approach the accuracy of
3D marker-based systems. In turn, these technical developments may also increase the
financial burden of these systems. Another limitation of this study is that only a small,
homogeneous sample of girls was included, which limits the generalizability of the results
presented in this pilot study. Therefore, future research is recommended to include a larger
and more heterogeneous sample of children with and without physical disabilities. This
study does however show that it is feasible to use DeepLabCut to assess upper limb and
trunk movements during gait in children with HIVE and spastic diplegia and provides
novel insights into the arm movements during walking in this population.
5. Conclusions
In conclusion, this pilot study showed that DeepLabCut could serve as a useful alterna-
tive for conventional gait analysis, with several advantages ranging from no dedicated and
expensive laboratory required, more time efficient, no need for highly skilled experts, and
does not affect the individual’s gait. Based on the markerless movement tracking system
DeepLabCut, children with HIVE showed deviations in their upper limb movements and
trunk sway. These abnormalities in upper limb behavior and trunk sway may be related to
strategies to compensate for impaired balance control due to spastic diplegia. However,
future research is needed to assess its validity and reliability.
Appl. Sci. 2025,15, 4546 10 of 12
Author Contributions: Conceptualization, M.M.E., P.M., N.G.L. and R.P.L.; methodology, M.M.E.
and P.M.; software, M.M.E.; validation, P.M.; resources, N.G.L.; writing—original draft preparation,
M.M.E. and P.M.; writing—review and editing, M.M.E., P.M., R.P.L. and N.G.L.; visualization, M.M.E.
and P.M.; supervision, R.P.L. and N.G.L.; funding acquisition, M.M.E., P.M. and N.G.L. All authors
have read and agreed to the published version of the manuscript.
Funding: This project was supported by the DIOS mobility programme 2019 (UHasselt & VLIR-UOS
[GM2019]), the International AIDS Society (Collaborative Initiative for Paediatric HIV Education
and Research (CIPHER) Grant Programme), the South African National Research Foundation, the
Neuroscience Institute at University of Cape Town and the special research fund (BOF20KV07) of
Hasselt University.
Institutional Review Board Statement: The study was conducted in accordance with the Declaration
of Helsinki and approved by the Human Research Ethics Committee of the University of Cape Town,
South Africa (HREC 447/2012).
Informed Consent Statement: Written informed consent was obtained from all the caregivers of
the participants before data collection, and assent was obtained from the participants involved in
the study.
Data Availability Statement: The data presented in this study may be made available on request from
the author in charge of the resources (NGL), depending on privacy, legal and ethical considerations.
Conflicts of Interest: The authors declare no conflicts of interest. The funders had no role in the design
of the study; in the collection, analyses, or interpretation of data; in the writing of the manuscript; or
in the decision to publish the results.
References
1.
Benedetti, M.G.; Beghi, E.; De Tanti, A.; Cappozzo, A.; Basaglia, N.; Cutti, A.G.; Cereatti, A.; Stagni, R.; Verdini, F.; Manca, M.;
et al. SIAMOC position paper on gait analysis in clinical practice: General requirements, methods and appropriateness. Results
of an Italian consensus conference. Gait Posture 2017,58, 252–260. [CrossRef]
2.
Castelli, A.; Paolini, G.; Cereatti, A.; Della Croce, U. A 2D Markerless Gait Analysis Methodology: Validation on Healthy Subjects.
Comput. Math. Methods Med. 2015,2015, 186780. [CrossRef]
3.
Fernández-González, P.; Koutsou, A.; Cuesta-Gómez, A.; Carratalá-Tejada, M.; Miangolarra-Page, J.C.; Molina-Rueda, F.Reliability
of Kinovea
®
Software and Agreement with a Three-Dimensional Motion System for Gait Analysis in Healthy Subjects. Sensors
2020,20, 3154. [CrossRef] [PubMed]
4.
Damsted, C.; Nielsen, R.O.; Larsen, L.H. Reliability of video-based quantification of the knee- and hip angle at foot strike during
running. Int. J. Sports Phys. Ther. 2015,10, 147–154.
5.
Van Rie, A.; Harrington, P.R.; Dow, A.; Robertson, K. Neurologic and neurodevelopmental manifestations of pediatric HIV/AIDS:
A global perspective. Eur. J. Paediatr. Neurol. 2007,11, 1–9. [CrossRef] [PubMed]
6.
UNAIDS. UNAIDS Data. 2019. Available online: https://www.unaids.org/en/resources/documents/2019/2019-UNAIDS-data
(accessed on 19 April 2025).
7.
Tindyebwa, D.; Kayita, J.; Musoke, P.; Eley, B.; Nduati, R.; Coovadia, H.; Bobart, R.; Mbori-Ngacha, D.; Kieffer, P.M. Handbook
on Paediatric AIDS in Africa, 4th ed.; African Network for the Care of Children Affected by HIV/AIDS—ANECCA Publishers:
Kampala, Uganda, 2017.
8.
Cronin, N.J.; Avela, J.; Finni, T.; Peltonen, J. Differences in contractile behaviour between the soleus and medial gastrocnemius
muscles during human walking. J. Exp. Biol. 2013,216, 909–914. [CrossRef]
9.
Langerak, N.G.; du Toit, J.; Burger, M.; Cotton, M.F.; Springer, P.E.; Laughton, B. Spastic diplegia in children with HIV
encephalopathy: First description of gait and physical status. Dev. Med. Child. Neurol. 2014,56, 686–694. [CrossRef] [PubMed]
10.
Mann, T.N.; Donald, K.A.; Laughton, B.; Lamberts, R.P.; Langerak, N.G. HIV encephalopathy with bilateral lower limb spasticity:
Upper limb motor function and level of activity and participation. Dev. Med. Child. Neurol. 2017,59, 412–419. [CrossRef]
11.
Kidzi ´nski, L.; Yang, B.; Hicks, J.L.; Rajagopal, A.; Delp, S.L.; Schwartz, M.H. Deep neural networks enable quantitative movement
analysis using single-camera videos. Nat. Commun. 2020,11, 4054. [CrossRef]
12.
Mathis, A.; Mamidanna, P.; Cury, K.M.; Abe, T.; Murthy, V.N.; Mathis, M.W.; Bethge, M. DeepLabCut: Markerless pose estimation
of user-defined body parts with deep learning. Nat. Neurosci. 2018,21, 1281–1289. [CrossRef]
13.
Nath, T.; Mathis, A.; Chen, A.C.; Patel, A.; Bethge, M.; Mathis, M.W. Using DeepLabCut for 3D markerless pose estimation across
species and behaviors. Nat. Protoc. 2019,14, 2152–2176. [CrossRef] [PubMed]
Appl. Sci. 2025,15, 4546 11 of 12
14.
Meyns, P.; Desloovere, K.; Van Gestel, L.; Massaad, F.; Smits-Engelsman, B.; Duysens, J. Altered arm posture in children with
cerebral palsy is related to instability during walking. Eur. J. Paediatr. Neurol. 2012,16, 528–535. [CrossRef] [PubMed]
15.
Meyns, P.; Kerkum, Y.L.; Brehm, M.A.; Becher, J.G.; Buizer, A.I.; Harlaar, J. Ankle foot orthoses in cerebral palsy: Effects of
ankle stiffness on trunk kinematics, gait stability and energy cost of walking. Eur. J. Paediatr. Neurol. 2020,26, 68–74. [CrossRef]
[PubMed]
16.
Meyns, P.; Duysens, J.; Desloovere, K. The arm posture in children with unilateral Cerebral Palsy is mainly related to antero-
posterior gait instability. Gait Posture 2016,49, 132–135. [CrossRef]
17.
Delabastita, T.; Desloovere, K.; Meyns, P. Restricted Arm Swing Affects Gait Stability and Increased Walking Speed Alters Trunk
Movements in Children with Cerebral Palsy. Front. Hum. Neurosc 2016,10, 354. [CrossRef]
18.
Eken, M.M.; Meyns, P.; Lamberts, R.P.; Langerak, N.G. Markerless movement tracking using a machine-learning algorithm to
assess arm movements during gait in children with HIV encephalopathy. Gait Posture 2020,81, 85–86. [CrossRef]
19.
Mann, T.N.; Laughton, B.; Donald, K.A.; Langerak, N.G. HIV encephalopathy with bilateral lower limb spasticity: Gross motor
function and antiretroviral therapy. Dev. Med. Child. Neurol. 2017,59, 407–411. [CrossRef]
20.
Caldwell, M.; Oxtoby, M.; Simonds, R.; Lindegren, M.; Rogers, M. Revised classification system for human immunodeficiency
virus infection in children less than 13 years of age. CDC Morbodity Mortal. Wkly. Rep. 1994,43, 1–10.
21.
Palisano, R.J.; Hanna, S.E.; Rosenbaum, P.L.; Russell, D.J.; Walter, S.D.; Wood, E.P.; Raina, P.S.; Galuppi, B.E. Validation of a Model
of Gross Motor Function for Children with Cerebral Palsy. Phys. Ther. 2000,80, 974–985. [CrossRef]
22.
World Medical Association. World Medical Association Declaration of Helsinki: Ethical principles for medical research involving
human subjects. JAMA 2013,310, 2191–2194. [CrossRef]
23.
Romkes, J.; Peeters, W.; Oosterom, A.M.; Molenaar, S.; Bakels, I.; Brunner, R. Evaluating upper body movements during gait in
healthy children and children with diplegic cerebral palsy. J. Pediatr. Orthop. B 2007,16, 175–180. [CrossRef] [PubMed]
24.
De Vlieger, D.; Defour, A.; Bar-On, L.; Cambier, D.; Swinnen, E.; Van der Looven, R.; Van Bladel, A. Speed-dependent changes in
the arm swing during independent walking in individuals after stroke. PLoS ONE 2025,20, e0315332. [CrossRef]
25.
Van Criekinge, T.; Saeys, W.; Hallemans, A.; Velghe, S.; Viskens, P.J.; Vereeck, L.; De Hertogh, W.; Truijen, S. Trunk biomechanics
during hemiplegic gait after stroke: A systematic review. Gait Posture 2017,54, 133–143. [CrossRef]
26.
Chaparro-Rico, B.D.M.; Cafolla, D.; Tortola, P.; Galardi, G. Assessing Stiffness, Joint Torque, and ROM for Paretic and Non-Paretic
Lower Limbs during the Subacute Phase of Stroke Using Lokomat Tools. Appl. Sci. 2020,10, 6168. [CrossRef]
27.
Wolfsegger, T.; Pichler, R.; Assar, H.; Topakian, R. Quantitative trunk sway analysis under challenging gait conditions in early
and untreated Parkinson’s disease. Neurol. Sci. 2022,43, 1411–1413. [CrossRef] [PubMed]
28.
Zhou, J.; Butler, E.E.; Rose, J. Neurologic Correlates of Gait Abnormalities in Cerebral Palsy: Implications for Treatment. Front.
Hum. Neurosci. 2017,11, 103. [CrossRef]
29.
Wang, K.K.; Munger, M.E.; Chen, B.P.J.; Novacheck, T.F. Selective dorsal rhizotomy in ambulant children with cerebral palsy.
J. Child. Orthop. 2018,12, 413–427. [CrossRef]
30.
Wagenaar, R.C.; Emmerik, R.E.A. Resonant frequencies of arms and legs identify different walking patterns. J. Biomech. 2000,33,
853–861. [CrossRef]
31. Wagenaar, R.C.; Emmerik, R.E.A. Dynamics of pathological gait. Hum. Mov. Sci. 1994,13, 441–471. [CrossRef]
32.
Van Bladel, A.; De Ridder, R.; Palmans, T.; Van Der Looven, R.; Verheyden, G.; Meyns, P.; Cambier, D. Defining characteristics
of independent walking persons after stroke presenting with different arm swing coordination patterns. Hum. Mov. Sci. 2024,
93, 103174. [CrossRef]
33.
Puig-Diví, A.; Escalona-Marfil, C.; Padullés-Riu, J.M.; Busquets, A.; Padullés-Chando, X.; Marcos-Ruiz, D. Validity and reliability
of the Kinovea program in obtaining angles and distances using coordinates in 4 perspectives. PLoS ONE 2019,14, e0216448.
[CrossRef]
34.
Iijima, S.; Shiomi, M.; Hara, T. Verification of Reliability and Validity of Trunk Forward Tilt Angle Measurement During Gait
Using 2-Dimensional Motion Analysis. J. Chiropr. Med. 2023,22, 89–95. [CrossRef] [PubMed]
35.
Ugbolue, U.C.; Papi, E.; Kaliarntas, K.T.; Kerr, A.; Earl, L.; Pomeroy, V.M.; Rowe, P.J. The evaluation of an inexpensive, 2D, video
based gait assessment system for clinical use. Gait Posture 2013,38, 483–489. [CrossRef]
36.
Fatone, S.; Stine, R. Capturing quality clinical videos for two-dimensional motion analysis. J. Prosthet. Orthot. 2015,27, 27–32.
[CrossRef]
37.
Wade, L.; Needham, L.; McGuigan, P.; Bilzon, J. Applications and limitations of current markerless motion capture methods for
clinical gait biomechanics. PeerJ 2022,10, e12995. [CrossRef]
38.
Needham, L.; Evans, M.; Cosker, D.P.; Wade, L.; McGuigan, P.M.; Bilzon, J.L.; Colyer, S.L. The accuracy of several pose estimation
methods for 3D joint centre localisation. Sci. Rep. 2021,11, 20673. [CrossRef] [PubMed]
Appl. Sci. 2025,15, 4546 12 of 12
39.
van den Bogaart, M.; Jacobs, N.; Hallemans, A.; Meyns, P. Validity of Deep Learning-Based Motion Capture Using DeepLabCut
to Assess Proprioception in Children. Appl. Sci. 2025,15, 3428. [CrossRef]
40.
Sipari, D.; Chaparro-Rico, B.D.M.; Cafolla, D. SANE (Easy Gait Analysis System): Towards an AI-Assisted Automatic Gait-
Analysis. Int. J. Environ. Res. Public Health 2022,19, 10032. [CrossRef]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.
