Skip to main content

Table 1 Details of the selected studies

From: A systematic review of the applications of markerless motion capture (MMC) technology for clinical measurement in rehabilitation

Study

Patient types

Sample size (n)

MMC system

Measurement items

Content of measurement

Context of measurement

Primary results

Results interpretation

Cho et al. 2009 [28]

PD

Patients with PD (7); healthy controls (7)

Sony HDR-HC3 camcorder

Gait pattern

Recognition of PD gait by algorithm combining PCA with LDA

Laboratory

The proposed system can identify healthy adults and patients with PD by their gaits with high reliability

Video-based analysis helps in discriminating the gait patterns of PD patients and healthy adults

Adde et al. 2010 [41]

CP

Infants with high risk of CP (30)

Digital video camera

Quantity of motion, velocity and acceleration of the centroid of motion

Comparison of quantity of motion and centroid of motion in infants who developed into CP with those who did not develop into CP

Hospital

Quantity of motion mean, median, and standard deviation were significantly higher in the group of infants who did not develop CP than in the group who did develop CP

Quantitative variables related to the variability of the center of infant movement and to the amount of motion predicted later CP in young infants with high sensitivity and specificity

Bahat et al. 

[61]

Chronic neck pain

Patients with chronic neck pain (25); asymptomatic participants (42)

Customized VR assessment system

Cervical ROM (flexion, extension, rotation, and lateral flexion)

Comparison of cervical movement in patients with chronic neck pain, versus in healthy controls

Laboratory

Significant group differences for 3 of the kinematic measures: Vpeak, Vmean, and number of velocity peaks

“Goal-directed fast cervical movements performed by patients with chronic neck pain were characterized by lower velocity and decreased smoothness compared with asymptomatic participants” [61]

Chen et al. 2011 [29]

PD

Patients with PD (12); healthy adults (12)

CCD video camera

Gait parameters including gait cycle time, stride length, walking velocity, and cadence

Quantification of gait parameters

Structured environment

KPCA-based method

achieved a classification accuracy of 80.51% in identifying different gaits

Kinematic data extracted from video might allow clinicians to obtain the quantitative gait parameters and assess the progression of PD

Khan et al. 2013 [14]

PD

Patients diagnosed with advanced PD (13); healthy controls (6)

Video recordings, analyzed by CV algorithm

Index-finger motion in finger tapping, features including speed, amplitude, rhythm, and fatigue in tapping were computed

SVM classification to categorize the patient group between UPDRS-FT symptom severity levels, and to discriminate between PD patients and healthy controls

Medical facility

The proposed CV-based SVM scheme discriminated between control and patient group with an average of 94.5% accuracy

The ML framework offers good classification performance in distinguishing symptom severity levels based on clinical ratings, as well as in identifying PD patients and the healthy controls

Lowes et al. 2013 [65]

Dystrophinopathy

Patients with dystrophinopathy (5); healthy controls (5)

Kinect

Upper extremity functional reaching volume, velocity, and rate of fatigue

Validity and Reliability of the MMC system in capturing upper extremity functional reaching volume, movement velocity, and rate of UE fatigue in individuals with dystrophinopathy

Laboratory

Preliminary test-retest reliability of the MMC method for 2 sequential trials was excellent for functional reaching volume

“The newly available gaming technology has potential to be used to create a low-cost, accessible, and functional upper extremity outcome measure for use with children and adults with dystrophinopathy” [65]

O’Keefe et al. 2013 [60]

FXS

Males with FXS (13); healthy controls (7)

BioStage™

Motion parameters (frequency and total traveled distance) of body segments during 30 s of story listening while standing in the observation space

Comparison between groups, MMC system results were compared with scores on video-capture methodology and behavioral rating scales

Laboratory

Arm and foot travel distances were significantly greater in the FXS group compared with the controls

“Motion parameters obtained from the markerless system can quantify increased movement in subjects with FXS relative to controls” [60]

Olesh et al. 2014 [46]

Stroke

Patients with stroke (9)

Kinect

10 movements of the upper extremity

Quantitative scores derived from motion capture were compared to qualitative clinical scores produced by trained human raters

Laboratory

Strong linear relationship was found between qualitative scores and quantitative scores derived from both standard and low-cost motion capture system

“The low-cost motion capture combined with an automated scoring algorithm is a feasible method to assess objectively upper-arm impairment post stroke” [46]

Gritsenko et al. 2015 [9]

Breast cancer

Women with mastectomy (4) or lumpectomy (16) for breast cancer

Kinect

Active and passive shoulder motions

Regression coefficients for active movements were used to identify participants with clinically significant shoulder ROM limitation

Laboratory

Participants had good ROM in the shoulder ipsilateral to the breast surgery at the time of testing. Three participants showed clinically significant shoulder motion limitations

Findings support the use of MMC approach as part of the automated screening tool to identify people who have shoulder motion impairment

Lee et al. 2015 [64]

AC of shoulder

Healthy volunteers (15); patients with AC (12)

Kinect

Shoulder ROM

Validity of measure shoulder ROM in AC by calculating the agreement of Kinect measurements with measurements obtained using a goniometer, and assessment of its utility for the diagnosis of AC

Laboratory

Measurements of the shoulder ROM using Kinect showed excellent agreement with those taken using a goniometer

“Kinect can be used to measure shoulder ROM and to diagnose AC as an alternative to a goniometer” [64]

Tupa et al. 2015 [30]

PD

Patients with PD (18); healthy age-matched individuals (18); students (15)

Kinect

Leg length, normalized average stride length, and gait velocity

A two-layer sigmoidal neural network was used for the classification of gait features (stride length and gait velocity)

Laboratory

Results showed high classification accuracy for the given set of individuals with PD and the age-matched controls

Kinect has potential to be used in the detection of abnormal gait and the recognition of PD

Sá et al. 2015 [56]

Schizophrenia

Clinically stable outpatients with schizophrenia (13); healthy controls (16)

BioStage™

Kinematic parameters and motor patterns during a functional task

Comparison of the kinematic parameters and motor patterns of patients with schizophrenia and those of healthy subjects

Laboratory

Patients with schizophrenia displayed a less developed movement pattern during performance of overarm throwing

“The presence of a less mature movement pattern can be an indicator of neuro-immaturity and a marker for atypical neurological development in schizophrenia” [56]

Kim et al. 2016 [47]

Stroke

Patients with hemiplegic stroke (41)

Kinect

Upper extremity motion of 13 of 33 items of upper extremity motor FMA

Correlation of the prediction accuracy for each of the 13 items between real FMA scores and scores using Kinect were analyzed

Laboratory

Prediction accuracies ranged from moderate to good in each item. Correlations were high for the summed score for the 13 items between real FMA scores and scores obtained using Kinect

“Kinect can be a valid way to assess upper extremity function, which may be useful in the setting of unsupervised home-based rehabilitation” [47]

Matsenet al. 2016 [75]

Variety of diagnoses (cuff disease, instability, arthritis)

Patients with a variety of diagnoses, including cuff disease, instability, arthritis (32); control healthy subjects (10)

Kinect

Shoulder active ROM

Correlation of Kinect shoulder active ROM measurement with SST

Laboratory

The total SST score was strongly correlated with the range of active abduction. The ability to perform each of the individual SST functions was strongly correlated with active motion

“Kinect provides a clinically practical method for objective measurement of active shoulder motion” [75]

Chin et al. 2017 [10]

CP

Children with USCP (31); typically developing children (21)

Kinect v2

Proprioception

Comparison of proprioceptive ability in children with USCP versus that in typically developing children

Laboratory

Children with USCP showed significant impairments in proprioception compared with typically developing children

The use of MMC technology can clearly identify differences in proprioceptive ability between typically developing children and children with UCSP

de Bie et al. 2017 [63]

ALS

Patients diagnosed with ALS (10)

Kinect

Upper extremity reachable workspace RSA

Evaluation of longitudinal changes in upper extremity reachable workspace RSA versus the ALSFRS-R, ALSFRS-R upper extremity sub-scale and FVC

Laboratory

RSA measures were able to detect changes in the upper limbs while the ALSFRS-R could not. The RSA measures were also able to detect a declining trend similar to that of FVC

“Kinect-measured RSA can detect declines in upper extremity ability with more granularity than current tools” [63]

Bakhti et al. 2018 [48]

Stroke

Individuals with hemiparetic stroke (19)

Kinect

Movements of 25 predefined body “joints” that approximately correspond to the center of the anatomical joint or body part

Use of ICC and linear regression analysis to quantify the degree to which an ultrasound 3D motion capture system motion capture system and Kinect measurements were related

Laboratory

PANU scores determined by the Kinect were similar to those determined by the ultrasound 3D motion capture system

“The Kinect sensor can accurately and reliably determine the PANU score in clinical routine” [48]

Bonnechère et al. 2018 [49]

Stroke

Healthy young adults (40); elderly adults (22); and patients with chronic stroke (10)

Kinect

Parameters including length, angle, velocity, angular velocity, volume, sphere, and surface of upper limb motion

The different scores and parameters were compared for the three groups

Laboratory

Highly significant differences were found for both the shoulders’ total angle, the velocity for young adults and elderly individuals, and patients with stroke

Results of the evaluation could be useful in monitoring patients’ conditions during rehabilitation, while further studies are needed to select which parameters are the most relevant

Butt et al. 2018 [15]

PD

Participants with PD (16); healthy people (12)

LMC

PSUP, OPCL, THFF, and POST

Comparison of parameters between a PD group and control group; Supervised learning methods SVM, LR, and NB for classification of patients with PD and healthy subjects

Laboratory

The best performing classifier was the NB. All the other subset features selected by the other feature selection methods, showed the worst classification performance in all ML classifiers (LR, NB, SVM)

“LMC is not yet able to track motor dysfunction characteristics from all MDS- UPDRS proposed exercises” [15]

Dranca et al. 2018 [31]

PD

Patients with PD (30)

Kinect

Gait step, limbs angle, and bent angles related to Parkinson disease

Classification of different PD stages by the features from FoG using classification algorithms

Hospital

The accuracy obtained for a particular case of a Bayesian Network classifier built from a set of 7 relevant features is 93.40%

“Using Kinect is adequate to build an inexpensive and comfortable system that classifies PD into three different stages related to FoG” [31]

Li et al. 2018 [25]

PD

Patients with PD (9)

Consumer grade video camera

416 features including kinematics, frequency distribution extracted from 14 joint angle positions

Quantifying the severity of levodopa-induced dyskinesia by video-based features

Laboratory

Features achieved similar or superior performance to the UDysRS for detecting the onset and remission of dyskinesia

“The proposed system provides insight into the potential of computer vision and deep learning for clinical application in PD ” [25]

Li et al. 2018 [32]

PD

Patients with PD after DBS (24)

Ordinary 2D video camera

TUG sub-task segmentation

Frame classification algorithm to classify video frame in sub tasks of TUG test

Semi-controlled environments

Classification accuracies for the sub-tasks ‘Walk,’ ‘Walk-Back,’ and ‘Sit-Back’ are apparently higher than that of the other three sub-tasks

The results support that clinical parameters for the assessment of PD can be automatically acquired from TUG videos

Martinez et al. 2018 [26]

PD

Patients with PD (6); healthy subjects (6)

DARI system

BME of 16 different movements

UPDRS-III and BME of 16 different movements in six controls paired by age and sex were compared with those in PD populations with DBS in ‘on’ and ‘off’ states

Laboratory

A better performance in the BME was correlated with a lower UPDRS-III score. No statistically significant difference between patients in ‘on’ and ‘off’ states of DBS regarding BME

The DARI MMC system is accurate in PD classification

Pantzar-Castilla et al. 2018 [45]

CP

Participants with CP (18)

Kinect 2 for Xbox One

Gait variables (i.e., Knee flexion at initial contact; Maximum knee flexion at loading response; Minimum knee flexion in stance; Maximum knee flexion in swing)

Comparison of 2D MMC and 3D marker-based gait analysis methods for the selected variables

Laboratory

The reliability within 2D Markerless and 3D gait analysis was mostly good to excellent

2D MMC is a convenient tool that could be used to assess the gait in children with CP

Rammer et al. 2018 [67]

Pediatric manual wheelchair users

Pediatric manual wheelchair users (30)

Kinect 2.0

Upper extremity kinematics during manual wheelchair propulsion (i.e., joint range of motion and musculotendon excursion)

Kinematic parameters were used to develop and evaluate a markerless wheelchair propulsion biomechanical assessment system

Laboratory

Inter-trial repeatability of spatiotemporal parameters, joint range of motion, and musculotendon excursion were all found to be significant

“A markerless wheelchair propulsion kinematic assessment system is a repeatable measurement tool for pediatric manual wheelchair users” [67]

Langevin et al. 2019 [16]

PD

Patients with PD (127); healthy controls (127)

Webcam

Frequencies of hand movement in hand motor task

Comparison of the differences in the hand motion between the groups with and without PD

Home Setting

PD group had a mean frequency that is lower than the control group in the hand motor tasks

“Online framework that assesses features of PD could be introduced during a clinic visit to initially supplement the tool with personal support” [16]

Lee et al. 2019 [17]

PD

Participants with PD that are receiving benefit from DBS (8)

LMC

PSUP, OPCL, and THFF tasks during ‘on’ and ‘off’ condition, amplitude, frequency, velocity, slope, and variance were extracted from each movement

Correlation of the kinematic features with the overall bradykinesia severity score (average MDS-UPDRS ratings across tasks)

Laboratory

An exhaustive LOSOCV assessment identified PSUP, OPCL, and THFF as the best task combination for predicting overall bradykinesia severity

“Data obtained from the LMC can predict the overall bradykinesia severity in agreement with clinical observations and can provide reliable measurements over time” [17]

Liu et al. 2019 [18]

PD

Patients with PD (60)

Camera

Periodic pattern of hand movements in finger tapping, hand clasping and hand pro/supination

Correlation analysis on each feature parameter and clinical assessment scores; Classification of bradykinesia

Semi-controlled environment

Classification accuracy in 360 examination videos is 89.7%

Reliable assessment results on Parkinsonian bradykinesia can be produced from video with minimal device requirement

Sato et al. 2019 [33]

PD

Patients with PD (117 in phase I and 2 in phase II); healthy controls (117)

Home video camera

Cadence

, gait frequency, gait speed, step length, step width, foot clearance

Estimation of cadence of periodic gait steps from the sequential gait features using the short-time pitch detection approach

Structured environment

Cadence estimation of gait in its coronal plane in the daily clinical setting was successfully conducted in normal gait movies using ST-ACF

2D movies recorded with a home video camera is helpful in identifying an effective gait and calculate its cadence in normal and pathological gaits

Vivar et al. 2019 [8]

PD

Patients with PD (20)

LMC

Tremor levels measured during hand extension and pushing the ball action

Classification of tremor level in PD according to the MDS-UPDRS standard

Laboratory

The proposed method classified the patient measurements following MDS-UPDRS in tremor levels 0, 1, and 2 with high accuracy

“It is possible to classify the different levels of tremor in patients with PD using only two statistical features, such as homogeneity and contrast” [8]

Caruso et al. 2020 [52]

ASD

Infants with high risk of ASD (50); infants with low risk of ASD (53)

Video recording

Quantity of motion, centroid of motion, presence of repetitive movements in the motion of limbs

Kinematic parameters related to upper and lower limb movements in infants with low risk and high risk of ASD

Bed

Early developmental trajectories of specific motor parameters were different in high-risk infants later diagnosed with neurodevelopmental diseases from those of infants developing typically

“Computer-based analysis of infants’ movements may support and integrate the analysis of motor patterns of infants at risk of neurodevelopmental diseases in research settings” [52]

Chambers et al. 2020 [66]

Neuromotor disease

Infants at risk of neuromotor impairment (19); healthy infants (85)

GoPro cameras, YouTube video

Absolute position and angle, variability of posture, velocity of movement, variability of movement, complexity, left-right symmetry of movement

Extent of kinematic features from infants at risk deviate from the group of healthy infants as reflected by Naïve Gaussian Bayesian Surprise metric

Childcare facility, hospital, natural environment

Infants who are at high risk for impairments deviate considerably from the healthy group

“Markerless tracking promises to improve accessibility to diagnostics, monitor naturalistic movements, and provide a quantitative understanding of infant neuromotor disorders” [66]

Fujii et al. 2020 [70]

Patients with gait disturbance

Patients with gait ataxia (6); control subjects (6)

Kinect 2, migrated to Azure Kinect

Gait parameters (e.g., walking speed and stride length)

Gait comparison between the patient group and the healthy subject group

Laboratory

Significant differences were observed between the patient group and the healthy subject group in terms of the mean value and variation of stride length

“A low-cost noninvasive motion capture device can be used for the objective clinical assessment of patients with stroke and PD who display manifestations of gait and motor deficits” [70]

Hu et al. 2020 [34]

PD

Patients with PD (45)

Video

Gait parameters, motion patterns

Automatic FoG detection by fine-grained human action recognition method

Structured environment

The experimental results demonstrate the superior performance of the proposed method over the state-of-the-art methods

“Anatomic joint graph representation provides clinicians an intuitive interpretation of the detection results by localizing key vertices in a FoG video” [34]

Krasowicz et al. 2020 [42]

CP

Patients with diagnosed ICP (8)

4DBODY system

TMFPI developed based on movement sequences

TMFPI compared with the assessment made according to the GMFM-88 scale

Laboratory

The system provided results agreeable with the clinical indicator GMFM-88 and with clinical observations of a PT

“The conducted assessments indicated that the use of dynamic 3D surface measurements is a promising direction of research and can provide valuable information on patient movement patterns” [42]

Lin et al. 2020 [19]

PD

Patients with PD (121)

iPhone 6s Plus

Motor behaviors, including stability, completeness, and self-similarity

Quantification of motor behaviors in patients with PD and bradykinesia recognition by a periodic motion-based network consisting of an autoencoder and fully connected neural network

Laboratory

The proposed periodic motion model delivers the F-score of 0.7778 for bradykinesia recognition

Using single RGB video for bradykinesia recognition is easy and convenient for patients and doctors

Oña et al. 2020 [39]

PD

Patients with PD (20)

LMC

Manual dexterity in BBT

Evaluation the validity of VR-BBT to reliably measure the manual dexterity

Laboratory

VR-BBT significantly correlated with the conventional assessment of the BBT

“VR-BBT could be used as a reliable indicator for health improvements in patients with PD” [39]

Pang et al. 2020 [20]

PD

Patients with PD; healthy controls (22)

Logitech HD Pro C920 webcams

Hand motion in tap thumb to the finger, creating a fist, pronation and supination of hand and resting state

Measurement of parkinsonian symptomology using automated analysis of hand gestures

Structured environment

Behavior of patients with PD and control subjects can be distinguished by analyzing the detailed motion features of their hands/fingers

Automatic hand movement detection method may help clinicians to identify tremor and bradykinesia in PD

Sabo et al. 2020 [58]

Dementia

Older adults with dementia (14)

Kinect

Gait parameters including cadence, average and minimum margin of stability per step, average step width, coefficient of variation of step width and time, the symmetry index of the step times, number of steps in the walking bout

Correlation and regression of gait features with clinical scores UPDRS and SAS

Hospital

Gait features extracted from both 2D and 3D videos are correlated to UPDRS-gait and SAS-gait scores of parkinsonism severity in gait

“Vision-based systems have the potential to be used as tools for longitudinal monitoring of parkinsonism in residential settings” [58]

Schroeder et al. 2020 [43]

CP

High-risk infants (29)

Kinect v1

Infants’ general movement

Correlation of expert GMA ratings of standard RGB videos with GMA ratings on SMIL motion videos of the same sequence

Clinical environment

GMA based on computer-generated virtual 3D infant body models closely corresponded to the established gold standard based on conventional RGB videos

SMIL motion video might capture the movement characteristics required for GMA of infants

Williams et al. 2020 [21]

PD

Patients with PD (20); control participants (15)

Smartphone

Bradykinesia assessed by finger tapping

ML models to predict no/slight bradykinesia or mild/moderate/severe bradykinesia, and presence or absence of Parkinson’s diagnosis

Clinical setting

SVM with radial basis function kernels predicted presence of mild/moderate/

severe bradykinesia with good accuracy. NB model predicted the presence of PD with moderate accuracy

The proposed approach supports the detection of bradykinesia without purchasing extra hardware devices

Williams  al. 2020 [22]

PD

Patients with idiopathic PD (39); healthy controls (30)

Smartphone

Bradykinesia assessed by finger tapping

Correlation of machine learning models with clinical ratings of bradykinesia

Clinical setting

Computer measures correlated well with clinical ratings of bradykinesia

“The research provides a new tool to quantify bradykinesia. It could potentially be used to support diagnosis and monitoring of PD” [22]

Zefinetti et al. 2020 [62]

SCI patients using a wheelchair

Patients with SCI (60)

Kinect v2

Kinematic data, including humeral elevation, horizontal abduction of humerus, humeral rotation, elbow flexion, trunk flexion/extension of wheelchair propulsion

Correlation between the movements and the patients’ assessment

Laboratory

The measurements computed by the proposed system showed a good reliability for analyzing the movements of SCI patients’ wheelchair propulsion

“The proposed markerless solutions are useful for an adequate evaluation of wheelchair propulsion” [62]

Abbas et al. 2021 [57]

Schizophrenia

Patients with Schizophrenia (18); healthy controls (9)

Smartphone

Head movement

Comparison of head movement measurements between patients and healthy controls, relationship of head movement to schizophrenia symptom severity

Home setting/ Natural environment

Rate of head movement in participants with schizophrenia and those without differed significantly; head movement was a significant predictor of schizophrenia diagnosis

“Remote, smartphone- based assessments were able to capture meaningful visual behavior for computer vision-based objective measurement of head movement” [57]

Ardalan et al. 2021 [71]

Neurodevelopmental Disorders

Children with 16p11.2 mutation (15); TD children (12)

A single point-and-shoot camera

Gait synchrony, balance parameters

Comparison of gait synchrony and balance in children with 16p11.2 mutation and TD children

Natural environment

Children with 16p11.2 mutation had significantly less whole-body gait synchrony and poorer balance compared to TD children

Remote video analysis approach facilitates the research in motor analysis in children with developmental disorders

Cao et al. 2021 [35]

PD

Patients with PD (18); healthy controls (42)

RGB camera

Shuffling step

Detection of shuffling step and severity assessment

Hospital

3D convolution on videos achieves an average shuffling step detection accuracy of 90.8%

Video-based detection method might facilitate more frequent assessment of FoG in a more cost-effective way

Hurley et al. 2021 [69]

Patients awaiting TKR who were attending POAC

Patients awaiting unilateral primary TKR (23)

BioStage™

LLM, VVM

Comparison of LLM and VVM performed clinically, radiologically, and using MMA

Laboratory

Discrepancies existed in LLM and VVM when evaluated using clinical, radiological, and MMA modalities

The MMC system should not be the only method to assess the patients for TKR

Kojovic et al. 2021 [55]

ASD

Children with ASD (169); TD children (68)

2D camera

Patterns of atypical postures and movements

Differentiation between children with ASD and TD using non-verbal aspects of social interaction by deep neural network

Clinical setting

The classification accuracy is 80.9% with the prediction probability positively correlated to the overall level of symptoms of autism in social affect and repetitive and restricted behaviors domain

Remote machine learning-based ASD screening might be possible in the future

Lee et al. 2021 [50]

Stroke

Patient with stroke (206)

Smartphone

Swing time asymmetry between paretic and non-paretic lower limbs while walking

Classification of dependence in ambulation by employing a deep model in 3D-CNN

Hospital

The trained 3D-CNN performed with 86.3% accuracy, 87.4% precision

“Monitoring ambulation using videos may facilitate the design of personalized rehabilitation strategies for stroke patients with ambulatory and balance deficits in the community” [50]

Li et al. 2021 [23]

PD

Patients with PD (157)

Video

Skeleton sequence from finger-tapping test

Classification of finger tapping performance according to MDS-UPDRS score

Hospital

Fine-grained classification net- work achieved an accuracy of 72.4% and an acceptable accuracy of 98.3%

Vision-based assessment method has potential for remote monitoring of PD patients in the future

Mehdizadeh et al. 2021 [59]

Dementia

Individuals admitted to a specialized dementia inpatient unit (54)

Kinect v2

Gait variables, including gait stability, step length, step time variability, step length variability

Changes in quantitative gait measured over a period during a psychogeriatric admission

Laboratory

Results showed that there was deterioration of gait in this cohort of participants, with men exhibiting greater decline in gait stability

“Quantitative gait monitoring in hospital environments may provide opportunities to intervene to prevent adverse events, decelerate mobility decline, and monitor rehabilitation outcomes” [59]

Negin et al. 2021 [53]

ASD

Children with or without ASD (108)

YouTube video

Spinning, head banging, hand action, arm flapping

Recognition of ASD associated behaviors

Natural environment

HOF descriptor achieves the best results when used with MLP classifier

“An action-recognition-based system can be potentially used to assist clinicians to provide a reliable, accurate, and timely diagnosis of ASD disorder” [53]

Nguyen-Thai et al. 2021 [44]

CP

Videos of infants who were at 14–15 weeks post-term age (235)

Smartphone

FM

Predicted the risk of CP by FM

Natural environment

Pose sequences were strong signals that retained motion information of joints and limbs while ignoring irrelevant, distracting visual artifacts

A STAM model can be used to identify infants at risk of cerebral palsy via video-based infant movement assessment

Rupprechter et al. 2021 [36]

PD

Patients with PD (729)

Smartphone

Leg ratio difference, vertical angle of the body, horizontal angle of the ankles and wrists, horizontal distance between the heels, speed of the ankles, step frequency

Estimation of severity of gait impairment in Parkinson’s disease using a computer vision-based methodology

Hospital and offices

Step frequency point estimates from the Bayesian model were highly correlated with manually labelled step frequencies

“Automated systems for quantifying Parkinsonian gait have great potential to be used in combination with, or the absence of, trained assessors, during assessments in the clinic or at home” [36]

Stricker et al. 2021 [37]

PD

Patients with PD (24)

Standard camera

Step length

Reliability of step length measurements from 2D video in patients with stroke; comparison of the step lengths of patients with/without a recent history of falls

Structured environment

Step length measurements from the video demonstrated excellent intra- and inter-rater reliability; patients with PD who had experienced a fall within the previous year demonstrated shorter step lengths

“Quantification of step length from 2D video may offer a feasible method for clinical use” [37]

Wei et al. 2021 [68]

Wheelchair user

Full-time wheelchair users (91)

Kinect

Wheelchair transfer motions including joint angles and positions

ML algorithm for evaluation of the quality of independent wheelchair sitting pivot transfers

Structured environment

Accuracies of the ML classifier were over 71%

“The results show promise for the objective assessment of the transfer technique using a low cost camera and machine learning classifiers” [68]

Williams et al. 2021 [72]

Tremor

Patients with PD (9); patients with essential tremor (5); patient with functional tremor (1)

Smartphone

Hand tremor at rest and in posture

Measurement of hand tremor frequency

Clinical setting

There was less than 0.5 Hz difference between the computer vision and accelerometer frequency measurements in 97% of the videos

“The study suggests a potential new, contactless point-and-press measure of tremor frequency within standard clinical settings, research studies, or telemedicine” [72]

Wu et al. 2021 [40]

PD

Patients with PD (7)

LMC

Hand kinematic in finger tapping hand opening and closing, and hand pronation and supination

Quantification of the motor component of bradykinesia

Laboratory

Average velocity and average amplitude of pronation/supination isolate the bradykinetic feature

“The LMC achieved promising results in evaluating PD patients’ hand and finger bradykinesia” [40]

Ferrer-Mallol et al. 2022 [73]

DMD

Patients with DMD (8)

Smartphone

Time, pattern of movement trajectory, smoothness and symmetry of movement

Quantitative measurement of the motor performance of the patients in the functional tasks

Home

Computer vision analysis allowed characterization of movement in an objective manner

“Video technology offers the possibility to perform clinical assessments and capture how patients function at home, causing minimal disruption to their lives” [73]

Guo et al. 2022 [24]

PD

Patients with PD (48); healthy controls (11)

RGB camera

Finger movement in finger tapping test

Classification of PD from finger tapping video

Structured environment

Classification accuracy is of 81.2% on a newly established 3D PD hand dataset of 59 subjects

Novel computer-vision approach could be effective in capturing and evaluating the 3D hand movement in patients with PD

Lonini et al., 2022 [51]

Stroke

Patients with stroke (8)

Digital RGB video camera

Gait parameters including cadence, double support time, swing time, stance time, and walking speed

Comparison of gait parameters obtained from clinical system and video-based method for gait analysis

Laboratory

Absolute accuracy and precision for swing, stance, and double support time were within 0.04 ± 0.11 s

“Single camera videos and pose estimation models based on deep networks could be used to quantify clinically relevant gait metrics in individuals poststroke” [51]

Morinan et al. 2022 [38]

PD

Videos from patients with PD (447)

Smartphone

Body kinematics including movement, velocity variation and smoothness

Estimation of ‘arising from chair’ task score in MDS-UPDRS

Clinical setting

Compute-vision based method can accurately quantify PD patients’ ability to perform the arising from chair action

Computer-vision based approach might be used for quality control and reduction of human error by identifying unusual clinician ratings

Vu et al. 2022 [74]

CD

Patients with CD (93)

Video recording

Peak power, frequency, and directional dominance of head movement

Quantification of oscillatory and directional aspects of HT

Structured environment

Computer-vision based method of quantification of HT exhibits convergent validity with clinical severity ratings

“Objective methods for quantifying HT can provide a reliable outcome measure for clinical trials” [74]

Morinan et al. 2023 [27]

PD

Patients with PD (628)

Consumer-grade hand- held devices

Movements during the bradykinesia examinations including finger tapping, hand movement, pronation-supination, toe tapping, leg agility

Quantification of bradykinesia according to clinician ratings

Clinical setting and laboratory

Classification model estimate of composite bradykinesia had high agreement with the clinician ratings

Computer vision technology with smartphone/ tablet devices can be adopted in the current clinical workflows

Song et al. 2023 [54]

ASD

Children with ASD (29); TD child (1)

RGB camera

Head and body movement during response to name behavior

Prediction of ASD by response to name behavior

Structured environment

The prediction method is highly consistent with clinical diagnosis

Automatic detection method can help to carry out remote autism screening in the early developmental stage of children

  1. 3D-CNN: 3D Convolutional Neural Network; AC: Adhesive Capsulitis; ALS: Amyotrophic Lateral Sclerosis; ALSFRS-R: Revised Amyotrophic Lateral Sclerosis Functional Rating Scale; ASD: Autism Spectrum Disorder; BME: Body Motion Evaluation; CCD: Commercial Digital Charge-coupled Device; CD: cervical dystonia; CP: Cerebral Palsy; CV: Computer Vision; DBS: Deep Brain Stimulation; DMD: Duchenne muscular dystrophy; FM: Fidgety Movement; FMA: Fugl-Meyer Assessment; FoG: Freezing of Gait; FoG: Freezing of gait; SAS: Simpson- Angus Scale; FVC: Forced Vital Capacity; FXS: Fragile X Syndrome; GMA: General Movement Assessment; GMFM-88: Gross Motor Function Measure-88; HOF: Histogram of Optical Flow; HT: Head Tremor; ICC: Intra-Class Correlation Coefficient; ICP: Infantile Cerebral Palsy; KPCA: Kernel-based Principal Component Analysis; LDA: Linear Discriminant Analysis; LLM: Leg Length Measurement; LMC: Leap Motion Controller; LOSOCV: Leave-One-Subject-Out Cross-Validation; LR: Logistic Regression; MDS-UPDRS: Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale; ML: Machine Learning; MLP: Multi-layer Perceptron; MMA: Markerless Motion Analysis; MMC: Markerless Motion Capture; NB: Naïve Bayes; NN: Neural Network; OPCL: Hand Opening/Closing; PANU: Proximal Arm Non-Use; PCA: Principal Component Analysis; PD: Parkinson’s Disease; PFP: Patellofemoral pain; POAC: Pre-Operative Assessment Clinic; POST: Postural Tremor; PSUP: Forearm Pronation-Supination; PT: Physiotherapist; RGB: Red Green Blue; ROM: Range of Motion; RSA: Relative Surface Area; SCI: Spinal Cord Injured; SDK: Software Development Kit; SMIL: Skinned Multi-Infant Linear Body Model; SST: Simple Shoulder Test; ST-ACF: short-time autocorrelation function; STAM: Spatio-Temporal Attention-Based Model; SVM: Support Vector Machine; TD: Typically Developing; THFF: Thumb Forefinger Tapping; TKR: Total Knee Arthroplasty; TMFPI: Trunk Mobility in the Frontal Plane Index; UDysPS: Unified Dyskinesia Rating Scale; UPDRS: Unified Parkinson’s Disease Rating Scale; UPDRS-FT: Unified Parkinson’s Disease Rating Scale-Finger Tapping; USCP: Unilateral Spastic Cerebral Palsy; VR: Virtual Reality; VVM: Varus/Valgus Knee Measurements