Stroke and cerebrovascular diseases are leading causes of both death and long-term disabilities worldwide ^[1][2][1,2], and hemiplegia is the most common impairment in the survivors ^[3][4][3,4]. Stroke results in a wide range of neurological deficits [5], and it severely affects motor skills, causing muscular weakness or partial hemi-paralysis that compromises the full arm function [6] and the mobility of the lower limbs ^[5][7][5,7].

With respect to lower limbs, hemiplegic gait is the most common manifestation of stroke: it is characterized by inter-limb asymmetry in walking or dragging gait due to the unilateral weakness of the affected side, together with abnormal torso tilting rotation ^{[5][8][9][10][11][12][13][14]}[5,8,9,10,11,12,13,14]. The decline in functional abilities and the impaired motor performance greatly affect the post-stroke patient’s quality of life ^{[9][15][16][17]}[9,15,16,17]. In order to limit the consequent impact on people’s daily life, a timely assessment is crucial for proper monitoring of the improvement or worsening of motor skills, and to set up effective rehabilitation protocols fitting each individual’s condition ^[18][19][20][18,19,20]. In this context, specific rehabilitation protocols that assess the functionality of the trunk/pelvis and the joints of the lower limbs are desirable.

The assessment of gait-related impairments in hemiplegic post-stroke patients is commonly performed using standardized clinical scales (e.g., the Fugl-Meyer Assessment score) [21] and specific walking tests. Among them, there are the 10-meter walk test (10MWT) [22], the 6-minute walk test [23], and the Timed Up and Go (TUG) test [24], which are often performed as part of the same examination session to gain a complete description of the patient’s walking behavior [25]. In the clinical field, optoelectronic 3D motion capture systems and force plates are the broadly acknowledged gold standard tools to assess gait patterns in a laboratory setting due to their consistency and measurement accuracy ^{[5][9][26][27]}[5,9,26,27]. Such systems provide spatiotemporal parameters, kinematics and kinetics that quantitatively describe the main features of the gait cycle, thus allowing for the functional performance assessment of patients and the identification of atypical patterns ^[5][28][5,28]. However, optoelectronic systems cannot be extensively used due to their high cost, complexity and often-troublesome equipment wearing requirements ^[29][30][29,30], together with the dependency on dedicated laboratory settings and specifically trained operators ^[26][31][32][26,31,32] and the need to wear limited clothing, which represents a great limitation for the application in many types of patients (for example, in patients with eating disorders or neurological patients with great functional problems in dressing).

Over the last decade, different technologies and methods ^{[33][34][35][36][37][38][39][40][41][42]}[33,34,35,36,37,38,39,40,41,42] have been proposed as an alternative to the optoelectronic systems for the analysis of body movements and functions as a medical diagnostic tool as well as in sports assessment. Among them, low-cost optical body-tracking sensors, such as the Microsoft Kinect, have proved to be particularly promising in order to assess both healthy and pathological gait, posture, postural instability, and balance in a non-invasive way ^{[5][9][43][44][45][46][47]}[5,9,43,44,45,46,47]. Originally featured for entertainment together with the Xbox^TM video game consoles, Microsoft Kinect^® (Microsoft Corporation, Redmond, WA, USA) has become a ground-breaking vision-based motion capture system, based on its color and depth sensors, finding application in new contexts including medical–clinical and rehabilitation settings ^[15][48][15,48]. Acknowledged as a non-intrusive tracking device [49], it requires neither any subject preparation nor attachment of markers to the patient’s body, nor a dedicated handheld controller [15]. In fact, its inherent technology is able to detect and capture the movements of the body in real-time by estimating the positions of the main joints through the anatomical landmarks of a skeletal model in the 3D space [50]. As it does not require any additional equipment, people are free to move with their natural patterns as they perform various tasks inside the device’s field of view, and their movements can be reproduced in real-time on the computer screen, for example, to obtain visual feedback [15]. The device is small, portable, and does not require a complex laboratory setup [26], enabling its use even as part of virtual home tele-monitoring and tele-rehabilitation systems ^{[51][52][53][54][55][56]}[51,52,53,54,55,56] that may allow the patients to practice exercises in a private environment [3]. In addition, it allows for training in specific motor and non-motor tasks with an amusing game-based approach that may increase the motivation and engagement of the patients ^{[15][57][58][59][60][61][62][63]}[15,57,58,59,60,61,62,63], in particular those in outpatient settings. This could permit an adequate and prolonged clinical monitoring in real-life contexts, thus replacing dedicated hospitalizations [64].

In recent years, several studies have investigated the accuracy and effectiveness of Microsoft Kinect for the assessment of posture, gesture, lower limbs and gait performance in several pathological states, such as stroke, Parkinson’s disease ^{[26][65][66][67][68][69][70]}[26,65,66,67,68,69,70] and other pathologies ^{[71][72][73][74]}[71,72,73,74]. Different studies have reported its reliability for the assessment of spatiotemporal gait parameters (e.g., step length and gait speed) and kinematic variables (e.g., trunk angle) in healthy individuals, with results comparable to those of laboratory-grade systems ^{[25][75][76][77][78][79][80]}[25,75,76,77,78,79,80] using both the first and second model of the device. The last version of the device, the Azure Kinect DK, was released in 2019 and, thanks to its new body tracking algorithm based on deep learning and convolutional neural networks [81], shows improved features in terms of depth accuracy and the number of joints tracked compared to the previous generations. Pilot studies on gait analysis ^[81][82][81,82] highlight higher accuracy of the device in the estimation of spatial gait parameters and kinematics compared to the previous models. Recent studies found that Azure Kinect had good agreement with a traditional motion capture system setup, indicating that the sensor could provide clinically relevant measurement of spatiotemporal parameters during gait [83], postural control [84], and sit-to-stand movement strategies, allowing for improved precision in clinical decision-making across multiple clinical populations [85]. The results showed high levels of agreement in evaluating spatiotemporal and kinematic variables during walking, sit-to-stand, and functional balance tasks, indicating that this technology is capable of accurate, and clinically relevant, assessment of motion data while performing these tasks. However, to date, no studies have explored its applicability in the characterization of a pathological condition.

Kinect sensor has been extensively used in Parkinson’s disease gait analysis ^[65][86][65,86] and postural control tests [66], such as the single-leg eyes-closed standing balance [87], claiming concurrent validity with the gold standard systems [88] and showing its ability to accurately measure some temporal and clinically relevant spatial features [49]. This non-invasive optical sensor increased the odds of virtual reality in the rehabilitation. The use of exergames is an innovative viable strategy for rehabilitation purposes, because it is not only recreational, but it also allows one to stimulate cognitive and motor functions and to promote physical activities through a more engaged game interaction ^{[89][90][91][92]}[89,90,91,92]. However, at present, few studies have used this technology in specific neurological treatments, such as in patients with stroke. The use of Kinect for the rehabilitation of post-stroke patients is in fact a recent topic. The first controlled and randomized studies were published in 2013 ^[93][94][93,94]. Moreover, the studies selected here represented small samples, and the majority comprised less than 30 volunteers. The greater use of Kinect with more significant results in the treatment of stroke patients was in the recovery of motor function and postural balance [95]. Nevertheless, conclusive findings on these and other variables were not yet possible, which increased the necessity for caution with this device in rehabilitation. With respect to the post-stroke population, however, just a few studies have focused on characterizing pathological gait patterns, but the lack of homogeneity among the characteristics of the cohorts, the selected pool of gait parameters, and the methodologies and objectives of the research make it difficult to directly compare their results.

2. Estimated Gait Parameters

Gait assessment is commonly based on spatiotemporal parameters, as in traditional gait analysis with gold-standard systems. However, there is a certain level of inhomogeneity among studies in this regard. In particular, the spatiotemporal parameters considered are generally not the same. In addition, the method used to estimate gait parameters varies from study to study: this introduces a bias in the results that makes direct comparisons between studies complicated. For example, in Latorre et al. [28] and Latorre et al. [18], some of the most commonly used spatiotemporal parameters in traditional gait analysis were estimated: gait speed, step information (distance and time), stride information (distance and time), asymmetry, double support, and swing time. In Latorre et al. [18], in addition to spatiotemporal parameters, some kinematic parameters were estimated and analyzed. In contrast, Clark et al. [87] and Ferraris et al. [5] estimated only a subset of spatiotemporal parameters related to the body’s single side (right and left sides) and the overall walk. In addition, Ferraris et al. [5] included parameters related to the body’s center of mass, which could be relevant for identifying specific abnormalities during walking associated with increased risk of fall (e.g., walking patterns with relevant lateral body sways). For the same reason, Luo et al. [26] also used parameters related to the body center of mass, derived with a methodology similar to Ferraris et al. [5], as they could be significant for classifying subjects with and without hemiplegia. Vernon et al. [25], on the other hand, estimated a small number of spatiotemporal parameters related only to the first step, the first stride, and the walking speed, as the setup used limited the space available for the gait analysis. Along a different line was Gao et al. [9], where no spatiotemporal parameters were estimated. Instead, a gait index was estimated from the analysis and overall motion (i.e., 3D trajectories) of specific joints in the skeletal model.

3. Statistical Analysis Methods

The statistical analysis is also closely linked to the objectives of a study, and for this reason, it generally includes several statistical tests that can, consequently, make the comparison of results more or less complex. For example, in studies involving a validation procedure, the goal is to demonstrate the accuracy of the obtained measures compared to a gold-standard system. In fact, in Latorre et al. [28], the average values (with standard deviation) of the estimated parameters for the control group and the post-stroke subjects were reported, as well as the estimated mean square error compared to the video analysis, that is, the gold-standard for this study. In Ferraris et al. [5], a series of statistical (Wilcoxon test) and correlation (Spearman correlation and ICC) tests were considered to demonstrate the accuracy of the estimated parameters on post-stroke subjects: estimated parameters were reported as median values with first quartile, compared to an optoelectronic system, the gold-standard ford this study. In addition, Spearman’s correlation was used to investigate the correlation between the estimated parameters and the TUG test, administered before the walking trials. Other studies mainly focused on the correlation between walking parameters and tests commonly used in clinical practice on post-stroke subjects, consequently using a statistical analysis that was more or less complex but appropriate to the study’s objectives. For example, because the study had multiple objectives, many statistical and correlation tests were used in Latorre et al. [18]: paired t-test to evaluate the significance of the statistical difference between populations divided by decade (control group and post-stroke subjects) and to identify subjects at risk of falling compared to the Berg balance scale; Pearson’s correlation coefficient to evaluate the effect of age on the estimated parameters and to validate the proposed solution against a set of clinical tests; and two-way random effect ICC to evaluate the inter-rater reliability between two raters and the intra-rater reliability for each individual rater. In contrast, Vernon et al. [25] included only two statistical tests: ICC for the test-retest reliability between the estimated parameters and the clinical tests performed on post-stroke subjects in the two planned sessions; and Spearman’s correlation to evaluate the correlation between the estimated parameters and the TUG test. The same statistical tests were also included in Clark et al. [87], where the Spearman’s correlation was used to evaluate the correlation of estimated gait parameters with static and dynamic balance. In contrast, the studies by Luo et al. [26] and Gao et al. [9], which focused on classification, did not report information on statistical analysis.

4. Findings and Data Availability

TAs discussed in the previous points, the inhomogeneity of the study objectives led to different methodological approaches and thus to different results and findings. Latorre et al. [28] characterized gait by comparing five different video analysis methodologies to estimate spatiotemporal parameters during gait, using video analysis for validation purposes. Following this approach, the overall results revealed limited accuracy between the Kinect-based and the video-based measurements in both the healthy and post-stroke groups. The researcheuthors hypothesized that such inaccuracies may be due to the speed and jitter of the tracking of ankle and toe ^[96][97][98,99], thus explaining the poor results achieved on the estimation of short duration and length event (e.g., double support) and the event that required the toe-off detection (e.g., swing time). Additionally, even though the use of video analysis is acknowledged as a valid approach in clinical setting, it may introduce additional errors in the measurements with respect to laboratory-grade systems ^[98][100]. According to the overall results, they concluded that the Kinect v2 could be used as a complementary tool to support the gait analysis in the estimation of events with a certain duration and length. With regard to the validation of the optical device with respect to the gold-standard optoelectronic systems, the results achieved by Ferraris et al. [5] show good agreement, accuracy, and correlation between the subset of spatiotemporal parameters estimated by the two systems, and compared with a clinical assessment test (i.e., Timed Up and Go Test). The results suggest the reliability of an optical-based system for the evaluation of gait impairments, even though some aspects need to be further explored. For instance, the restricted number of participants may have biased the robustness of the characterization of the gait parameters, thus affecting the clinical and statistical findings. Regarding the correlation of the Kinect-derived gait parameters with clinical tests for post-stroke patient assessment, the study presented by Latorre et al. [18] showed excellent intra-reliability between the clinical test and almost all gait measures, also allowing for the identification of patterns exposing the patient to fall risk. However, the minimal detectable change was inconstant among the measured gait parameters, resulting in a poor estimation of the kinematic parameters. Similar results on the estimation of gait parameters were achieved by Vernon et al. [25] and Clark et al. [87]. In particular, in Vernon et al. [25], all Kinect-estimated variables showed excellent reliability (ICC > 0.90), with the exception of the trunk flexion angle. Similarly, as the two researcheuthors worked on the same dataset, the estimated variables confirmed the previously found high reliability (ICC > 0.80), even though many of the results were redundant. The remaining studies focused on gait pattern characterization and classification through objective parameters using machine learning techniques. In particular, Luo et al. [26] worked on the development of a random forest method for the classification and analysis of hemiplegic gait. The method was developed starting from a pool of gait features (e.g., stride length, gait speed, left/right moving distances) acquired via Kinect, resulting in the achievement of an averaged classification accuracy of 90.65% among all the combinations of gait features. Conversely, whereas Luo et al. [26] highlighted the usability of Kinect-derived data in machine learning techniques, Gao et al. [9] used the Kinect to record kinematic data during walking to obtain a quantitative evaluation of the gait quality of the hemiplegia group according to a gait quality index (GQI) based on a radar map. The final results show a significant negative correlation between the GQI and the Fugl-Meyer Assessment score for lower limbs, together with a significant statistical difference in lower limb joint movement quality between the healthy and the hemiplegia groups, reflecting the differences in motion quality of the joints for the two groups. The results thus highlight the reliability of the GQI, estimated from joint trajectories, as an assessment tool to support clinical decisions on rehabilitation programs.