Table 4.

Key findings from studies that used the Kinect.

Study	Primary features	Main results
Cimolin et al (2022) [24]	Gait cadence, mediolateral sway, and step width	Strong positive correlation between Kinect and Vicon systems for gait cadence and mediolateral sway (ICCa 0.94-0.97) and a weak correlation for step width (ICC 0.44) in people with PDb
Kondragunta et al (2020) [19]	Gait cycle (dynamic time warping)	SVMc for classifying between controls, persons with possible MCId, and persons with MCI: 74.6%-87.3%
Lai et al (2022) [28]	Stride length, straight walking speed, and turning speed	Mediation analysis demonstrates decreased stride length, walking speed, and turning speed are associated with increased falls prediction model score (r=–0.58, r=–0.52, and r=–0.46, respectively; P<.001) UPDRSe negatively correlated with features (r=–0.65, r=–0.56, and r=–0.37, respectively; P<.001) but positively with fall prediction model score (r=.53, P<.001) UPDRS serves as a mediator for features and higher fall prediction model scores
Mehdizadeh et al (2021) [9]	Gait stability, step time, step length, step time variability, and step length variability	Mixed effects models over 10 weeks show: Decrease in primary features and an increase in variability over time for people with dementia Gait stability decreased more in men Mediolateral range of motion decreased in those with mild neuropsychiatric symptoms but increased in those with more severe symptoms
Mehdizadeh et al (2021) [15]	Gait stability.	Cox proportional hazard regressions show gait stability predicts time to fall in people with dementia (ROCf 0.80 at 7 days, 0.67 at 30 days)
Muñoz-Ospina et al (2022) [31]	Left and right arm and ankle swing (magnitude and speed), stance time, gait speed, total time, and number of steps	Random forest model was most accurate for discriminating between people with PD and controls (85% using all gait features)
Ng et al (2020) [16]	Gait: cadence, symmetry, CVg of step time, step width (average and CV), and eMOSh	Univariate linear regression: cadence associated with POMAi-gait scores (P<.001) Poisson regression: cadence, eMOS, average step width associated with the number of future falls (P<.001)
Ospina et al (2021) [32]	Arm swing: magnitude, time, and arm swing asymmetry	Age influenced arm movement People with PD showed significant reductions in arm swing magnitude (left, P=.002; right, P=.006) and speed (left, P=.002; right, P=.004) Arm swing asymmetry differentiated people living with PD from controls (ROC: 78%)
Otte et al (2020) [33]	Cadence, knee amplitude, asymmetry, average step time, longest step time, arrhythmicity, average stance time, and longest stance time	Knee amplitude and longest stance time correlated with UPDRS (–0.51, P=.003 and 0.52, P=.002, respectively) Postural instability (pull test) correlated with longest stance time (0.47, P=.008) Knee amplitude, asymmetry, and average step time differed between on- and off-medication states (P=.002, P=.007, and P=.007, respectively)
Pedro et al (2020) [34]	Step length	In comparison with the GAITRite (CIR Systems, Inc) system, the Kinect camera overestimated the average variation in step length for the 2 people with PD potentially due to inherent smoothing
Procházka et al (2015) [35]	Average step length	In total, 91.7% classification accuracy for determining between controls and those with people with PD. Decrease in step length (regression coefficient=–0.0082 m/year)
Sabo et al (2022) [18]	Number of steps, cadence, velocity, step length, CV of stride width, and step and swing time	Moderate or strong positive correlations between steps, cadence, step width from 2D pose-estimation, and Zeno in people with PD Automated heel strike algorithm struggled to identify short steps
Sabo et al (2021) [17]	Cadence, steps, average step width, average margin of stability, CV of step width and time, and symmetry	ST-GCNj using 2D joint trajectories and gait features outperforms ST-GCN using only gait features Regression models for predicting UPDRS-gait over 94% if off by 1 is allowed
Sabo et al (2020) [38]	2D: steps, cadence, symmetry, and CV of step time 3D: walking speed, step length or width, step width, step length symmetry angle, RMSk of MLl velocity, margin of stability, and CV step width	Multivariate ordinal logistic regression models achieved 61.4% and 62.1% for 2D and 3D features for predicting UPDRS-gait in people with dementia
Seifallahi et al (2022) [39]	Steps and stride	Adaptive neuro-fuzzy inference system classifier accuracy >90% for differentiating between MCI and controls
Soltaninejad et al (2018) [41]	Stride and tremor	Random forest classifier accuracy for differentiating controls and people with dementia: 93.33% stride and 81% tremor
Tan et al (2019) [42]	Step length, step time, vertical pelvic displacement, and gait speed	Multivariable regression: step length during TUGm and vertical pelvic displacement during the gait speed were associated with postural instability and gait disorder (P=.01 and P<.05, respectively) in people with PD
Ťupa et al (2015) [43]	Step length and average speed	Combining gait features improves classification accuracy relative to single features 2-layer neural network achieved an accuracy of 97.2% in classifying people with PD from controls

^aICC: intraclass correlation coefficient.

^bPD: Parkinson disease.

^cSVM: support vector machine.

^dMCI: mild cognitive impairment.

^eUPDRS: Unified Parkinson’s Disease Rating Scale.

^fROC: receiver operating characteristic.

^gCV: coefficient of variation.

^heMOS: estimated margin of stability.

ⁱPOMA: Tinetti Performance Oriented Mobility Assessment.

^jST-GCN: spatiotemporal graph convolutional networks.

^kRMS: root mean squared.

^lML: mediolateral.

^mTUG: Timed Up and Go.