The kinectome: A comprehensive kinematic map of human motion in health and disease

Abstract Human voluntary movement stems from the coordinated activations in space and time of many musculoskeletal segments. However, the current methodological approaches to study human movement are still limited to the evaluation of the synergies among a few body elements. Network science can be a useful approach to describe movement as a whole and to extract features that are relevant to understanding both its complex physiology and the pathophysiology of movement disorders. Here, we propose to represent human movement as a network (that we named the kinectome), where nodes represent body points, and edges are defined as the correlations of the accelerations between each pair of them. We applied this framework to healthy individuals and patients with Parkinson's disease, observing that the patients’ kinectomes display less symmetrical patterns as compared to healthy controls. Furthermore, we used the kinectomes to successfully identify both healthy and diseased subjects using short gait recordings. Finally, we highlighted topological features that predict the individual clinical impairment in patients. Our results define a novel approach to study human movement. While deceptively simple, this approach is well‐grounded, and represents a powerful tool that may be applied to a wide spectrum of frameworks.


INTRODUCTION
Movement is essential to human life and survival. 1 As a consequence, movement impairment significantly reduces quality of life and individual autonomy. Thus, the study of the characteristics of the voluntary movement is of broad interest in multiple frameworks. However, an appropriate description of human movement requires taking into account multiple simultaneous interactions, 2 stemming from the coordinated activations of several musculoskeletal segments, 3,4 and resulting in complex patterns. 5 Such patterns are fine-tuned, and also small changes can lead to physiologically relevant effects. 6 Therefore, an accurate characterization of these patterns requires precise measurements and appropriate mathematical methods, in order to capture and describe their complex interactions.
To date, human movement kinematics, and notably gait kinematics, has been approached focusing on specific body segments, or conveying complex patterns into a few synthetic parameters. [7][8][9][10] While this approach is useful, it inevitably leads to loss of information, providing a summary picture of human movement rather than a comprehensive description of the complex patterns of interactions that generated it.
Yet, whole-body interactions are needed for a comprehensive account of movement dynamics. 11,12 Complex network theory is a methodological approach to integrate into a unique explanatory framework, complex systems consisting of a large number of interconnected elements. 13 This approach allows the description of the properties of the network and, ultimately, of its functioning. The complex networks may be analyzed using a host of mathematical techniques, such as graph theory, a branch of mathematics dedicated to the study of the topological properties of the networks.
Algebraically, graphs are represented as adjacency matrices, square arrays of numbers wherein rows and columns correspond to nodes and individual entries give the connection between each node and all the others. 14 Network analysis, given its ability to capture the properties of the network as a whole, but also to analyze the contribution of each individual element to the organization of the entire network, finds wide applications in a large number of disciplines (e.g., physics, sociology, epidemiology, climatology, and neuroscience). [15][16][17][18] As a consequence, network science may lend itself nicely to describe the complex patterns generated by motor behaviors, and extract features that are relevant to the pathophysiology of movement disorders. 19 Indeed, in the last decades, network science has been extensively applied to characterize the aspects of neurological disorders. 13 Recently, the first applications of network analysis to the study of human movement proved successful. Utilizing electromyography, Boonstra et al. analyzed the network of the leg muscles, detecting the presence of lower and higher frequency components related to between and within legs connectivity, respectively. 20 The authors suggested that network analysis may be suitable to study the motor system also in a clinical setting. In another study, Kerkman et al. investigated a combined musculoskeletal network structure. They examined the different frequency-specific muscle networks during postural control. 21 The study showed that the examined networks presented frequency-specific relationships with the synaptic input to motor neurons.
Despite these first efforts, to date a comprehensive network description of the kinematics of movement is still lacking. To overcome this deficiency, borrowing concepts from network science, we set out to represent certain anatomical points as nodes, and their coaccelerations throughout gait as edges, thereby defining the network of human movement. With this approach, we aimed at identifying the large-scale characteristics of the human gait. Hence, we considered the whole body as an integrated and synergistic system, whose individual musculoskeletal segments are in a constant and reciprocal biomechanical relationship constrained by the individual anatomical characteristics.
To this end, we utilized a three-dimensional motion analysis stereophotogrammetric system, which is the gold standard for quantitative analysis of movement, 22 and is widely applied for the assessment of motor skills in health and disease. [23][24][25][26][27] Specifically, we captured the position of reflective markers applied on specific bone reference points during gait. Each bone marker was considered as a node, and the edges linking the nodes were defined by the covariance of the acceleration and jerk (i.e., the first derivative of acceleration with respect to time) between each pair of bone markers. We named the resulting network the human kinectome. We focused on acceleration and its derivative (i.e., jerk) for our analysis since those kinematic measures are mainly associated with smoothness of gait and quality of movement control. 28,29 Indeed, through acceleration and its tuning, we are able to properly control speed.
Then, we characterized the human kinectome in a cohort of healthy subjects (HS). Furthermore, in order to explore the clinical relevance of our framework, we compared the kinectomes of individuals affected by Parkinson's disease (PD), a neurodegenerative disorder which disrupts the motor patterns of the patient, 30 to those of matched healthy controls (HC). Kinematics in PD patients has been widely investigated, with several studies focusing on different aspects of motor impairment, including variability, asymmetry, smoothness, and stability of gait. 31,32 Hence, PD kinematics emerges as a natural scope for the proposed approach. We hypothesized that PD patients would be less capable of maintaining an optimal motor strategy, as opposed to HC. According to this hypothesis, we first explored the structure of the kinectomes, expecting a dysregulated (i.e., more variable) organization in patients with respect to controls. Then, to test the reliability of the kinectome, we performed an identifiability analysis, 33 identifying subjects based on their kinectomes, similarly to a motion fingerprint. This idea is in analogy to recent evidence showing that dysregulated activity would make brain network identifiability harder in patients as compared to healthy people. 34  To test the validity of our methods in a clinical setting, we used the data of 23 patients (mean age 65.3 ± 11.6) affected by PD and 23 HC, matched for age, sex, and education. The subjects included in this study are partially overlapping with those included in a previous study. 39 Parkinsonians were tested in off-medicament state. Inclusion criteria were: (1) Hoehn and Yahr score ≤ 3 while off-medicament; 40 (2) disease duration < 10 years; and (3) antiparkinsonian treatment at a stable dosage. All participants signed an informed consent in accordance with the declaration of Helsinki. The study was approved by the "Azienda Ospedaliera di Rilievo Nazionale A. Cardarelli'" Ethic Committee (protocol number: 00019628).

Stereophotogrammetric acquisition
The acquisitions were carried out in the Motion Analysis Laboratory of the University of Naples Parthenope. Gait data were recorded through a stereophotogrammetric system for motion analysis composed of eight infrared cameras (ProReflex Unit-Qualisys Inc., Gothenburg, Sweden), capturing (at 120 frame per second) the light reflected by 21 passive markers positioned on the naked skin of the participants. The markers were placed in correspondence of bone landmarks, based on a modified version of the Davis protocol. 41 We asked the participants to walk in a straight path choosing their preferred walking speed. For each participant, two gait acquisitions were performed, each of which included one complete left and right gait cycle. A complete gait cycle is defined as starting with the heel touching the ground, and finishing with the next contact with the ground of the same heel. Through the Qualisys Track Manager software, we obtained the three-dimensional position of each bone marker during the gait cycle. Hence, we could calculate the time series for acceleration and jerk (the first derivative of acceleration with respect to time) of each bone marker.

Introducing the kinectome
We computed the Pearson's correlation coefficients between each pair of the time series representing the bone markers (see also Figure 1A,B), and defined the kinectome as the covariance matrix, which conveys whole-body interactions in a pairwise fashion. Hence, using 21 markers as nodes, we obtained a symmetric matrix containing 420 edges (excluding the main diagonal elements which represent the correlation of a node with itself). Only 210 edges (since the kinectome is symmetric) were used in the subsequent analyses.
Time series from acceleration and jerk of the 21 markers along the three axes of movement (i.e., mediolateral, anteroposterior, and vertical) were used to build six kinectomes for each subject (2 kinematic units × 3 axes of movement). First, we explored the kinectomes heterogeneity within and between groups (PD patients and controls), by comparing mean and standard deviations of the kinectomes. On the one hand, the analysis of the mean allows to understand the level of motor synchronization, and may help to understand whether a clinical condition is able to alter it. On the other hand, the analysis of the standard deviation allows to assess the variability of the motor patterns within a group, which in turn highlights anatomical elements affected by the disease. After those preliminary investigations, we then characterized the kinectomes utilizing a graph-theoretical approach, as detailed in the next sections, and shown in the flowchart ( Figure 2).

Modularity analysis
Modularity measures the strength of division of a network into modules or communities. We assessed the community structure (i.e., partition) of each group-averaged kinectome (anteroposterior, mediolateral, and vertical, separately), in both healthy and PD patients, by using the Louvain (with consensus clustering across 100 iterations) method for identifying communities in large networks ( Figure 1C). 42,43 This is a method that detects communities by optimizing the modularity (Q) of the graph, defined as: where m is the sum of weights of all the edges in the network, A ij is the weight of the edge between i and j, k i and k j are the sum of the weights of the edges connected to nodes i and j, respectively, and δ(c i , c j ) is the Kronecker delta between community c i and c j . 42 With this approach, we were able to determine which body elements were recognized as belonging to the same group, based on their acceleration and jerk motor patterns, creating an allegiance matrix. 44 The aim was to identify, in a data-driven fashion, functional dynamical clusters within healthy and diseased kinectomes during gait. The modularity analysis allows to divide the body into sets of anatomical elements working together toward a motor task. While symmetrical organization of these groups is expected in health, individuals with PD might show altered distributions, highlighting which are the most affected anatomical elements.

Fingerprint analysis
Can we identify individuals based solely on their motion patterns, that is, their kinectomes? To address this question, we took inspiration from previous studies on fingerprint in human functional brain connectomes extracted from functional magnetic resonance imaging and magnetoencephalography data. 33,34 In a recent work, 33  which encodes the information about the self-similarity (I-self, main diagonal elements) of each subject with herself/himself, comparing two recording sessions, and the similarity of each subject with the others. To build an IM based on the kinectomes, we first considered two gait cycle registrations for each individual, called "test" and "retest," respectively.
We then obtained the IM through Pearson's correlation between the test and the retest of our subjects ( Figure 1D). The main diagonal of this matrix contains the similarity between two separate acquisitions of the same subject (self-similarity or I-self); the off-diagonal elements contain the similarity between different subjects (I-others). Furthermore, the difference between I-self and I-others, also known as differential identifiability (I-diff), provides a robust score of the overall fingerprinting assessment of a dataset. Finally, we estimated the identification rate (IR) as the percentage of times a subject was identified when compared to different subjects, as: where N is the sample size of the group, Iself n is the similarity between two connectomes of the same individual, and Iothers n is an array of elements representing the similarity between an individual with every other individual of the same group.
The fingerprint analysis assesses the uniqueness of the individual movement patterns. Moreover, if applied within a longitudinal framework, our approach might show individual changes over time.

F I G U R E 2
Schematic description of the network analysis. The flowchart describes the methodological approaches applied to the kinectomes. Three main network frameworks were explored: modularity, 42 fingerprint, 33,45 and topology. 46 For each of them, the methodological approach and the aim of the analysis have been highlighted.

Edge-based identification
We repeated the IR analysis on subsets of edges, based on their contribution to fingerprinting. To obtain this information, similarly to Sorrentino et al., 34 we used the intraclass correlation (ICC): 45 where MSA is the among-clusters mean square, MSW is the withinclusters mean square, and k represents the number of observations. 45 It is an approach that assesses how stable an edge value is across test-retest kinectomes. The higher the stability of an edge between the two kinectomes, the higher the contribution to identifiability. We performed this analysis in PD and HC groups separately, obtaining two ICC matrices. Based on this information, we calculated the IR of the two groups at each step utilizing an iterative model in which we added the edges in descending ICC order from the most to the least contributing to the identifiability. We started with the three edges and kept adding one edge at each iteration, up to including the complete kinectome. We obtained a curve displaying the IR of each group each time an edge was added to the analysis. To confirm the validity of the chosen ordering, for each curve of IR based on ICC ordered edges, we built 100 null curves obtained by calculating the IR based on randomly selected edges. To highlight the nodes that significantly contribute to subject identifiability, we checked how many edges were needed to exceed the 99% IR, and considered those edges as of interest. Then, we checked the distribution of the occurrences of the nodes over which the edges of interest hinge. Nodes whose occurrences exceeded chance level (confidence interval set to 99%) were considered significant and were considered for further investigation. This approach highlights which elements of the motor patterns are specific to the single subject (hence allowing identification). When applied to diseased motor patterns, this analysis may point to anatomical/functional elements of clinical interest.

Topological feature for the motor impairment prediction
We conceptualized the body as a network, where body parts are nodes and their correlations form the edges, thereby obtaining one weighted undirected graph per subject ( Figure 1E). For each node of a graph, we estimated the weighted degree (s), a centrality parameter, 13,46 defined as the sum of the absolute value of the edge weights for each node: 47 where i and j are two nodes of the network, w is the edge connecting them, and N is the number of nodes.
The degree of the nodes of interest was used to predict clinical impairment in patients. To this aim, we built a multilinear regression model to predict the UPDRS scores from the degree of the nodes of interest. 48 We added further predictors to the analysis to account for the effect of age, education, and gender. Multicollinearity was assessed through the variance inflation factor. 49,50 Furthermore, we improved the robustness of our approach using the k-fold cross-validation, with k = 5. 51 In particular, k iterations were performed and at each iteration, the kth subgroup was used as a test set.
Topological analysis elucidates the role of a single element with respect to the whole body. If a node presents an altered degree (with respect to healthy individuals), then that anatomical element may be of particular interest to understand how the disease affects the motor patterns.

Statistics
Statistical and data analysis were carried out in MATLAB 2020a.
Significance of the between groups (PD and HC) differences in the kinectomes standard deviation, fingerprint values (I-self, I-other, and I-diff), and topological parameter (degree) were assessed through permutation testing, by randomly shuffling group labels 10,000 times. At each permutation, the absolute value of the difference was computed, obtaining a distribution of the differences that are to be expected by chance alone. 52 This distribution was compared to the observed differences to retrieve a statistical significance. Correlation analysis between nodal degree and motor scores was performed through the Spearman correlation test. The significance threshold was set at p < 0.05, and was Bonferroni corrected in each analysis.

Group-specific characteristics of the kinectomes
We started from a group-level analysis comparing the average kinectomes of HS and those of PD patients that were recorded during gait.
Specifically, after building the subject-specific kinectomes ( Figure 3A), we averaged them within each group, obtaining the group-specific (i.e., HS, HC, and PD) kinectomes. Then, using permutation analysis, we compared the average values of the kinectomes in HC and PD patients ( Figure 3B). However, neither acceleration nor jerk kinectomes highlighted any significant difference between the two groups in any axes of movement. That is, the acceleration and jerk patterns of the two groups were similar to each other. Note that Figure

Modularity analysis
We then set out to provide a principled description of the kinectomes' topological structure, by investigating the emerging modular structure of the kinectomes. To this end, we computed the allegiance matrices, 44 which contain the probability of any two bone markers being clustered in the same community across individuals. This means that two or more bone markers belonging to the same cluster refer to body parts which are likely to coordinate themselves toward the same motor pattern, for each specific group. Figure 5 shows that the HS and HC groups This approach unraveled the kinematic structure of gait in the healthy, and its alterations in disease. It is noteworthy that the algorithm calculating the communities split the body parts symmetrically in healthy individuals, while the same result was not achieved for the PD patients, which might be due to the typically asymmetrical motor impairments occurring in PD, which is even a diagnostic criterion.

Fingerprint of human movement
Based on these results, we wondered whether it was possible to identify each individual based on their motion patterns. To answer this question, we tried to identify individuals through their kinectomes, obtained from different gait sessions recorded the same day. To this aim, we started by building an IM based on the kinectomes. 33 In the IM, the rows refer to the kinectomes from the first recording session (test kinectomes in Figure 1D), and the entries on the columns refer to the kinectomes from the second recording session (retest kinectomes in Figure 1D)  (Figure 7). This effect is mainly driven by the difference in the I-others scores. In fact, the PD group showed lower I-others scores as compared to HC patients in AP acceleration (p < 0.0001), ML jerk (p < 0.0001), and AP jerk (p < 0.0001) (Figure 7). This implies that the PD patients have more heterogeneous motor patterns with respect to the HC groups; hence, their kinectomes differ more with respect to each other. Nonetheless, both groups expressed similar IRs, which were above 95%. This result highlighted that an almost perfect identification is possible for both PD patients and controls.

F I G U R E 5
Kinematic modular organization of the kinectomes. Allegiance matrices for cluster analysis, based on the Louvain method and consensus-clustered through 100 iterations. The algorithm automatically defines which body parts belong to the same community, suggesting a functional relationship among those elements. Each matrix includes clustering information from both accelerations and jerks. Healthy subjects (HS) and healthy controls (HC) share the same communities in both mediolateral (ML) and anteroposterior (AP) axes. Parkinson's disease (PD) patients' matrices show different structural organizations. Body parts depicted with the same color belong to the same functional community.

F I G U R E 6
Motion fingerprinting: identifiability based on kinectomes. Identifiability matrices of healthy controls (HC) and Parkinson's disease (PD) patients, based on jerk and acceleration kinectomes in mediolateral (ML) and anteroposterior (AP) axes. The highest values within the main diagonal (I-self) convey great self-similarity. Off diagonal elements (I-others) are representative of the similarity between different subjects. IR, identification rate; I-diff is the differential identifiability scores of the dataset and is defined as I-self-I-others.

F I G U R E 7 Identifiability comparison between healthy controls and patients. Box plot for the comparison of I-diff and I-others between healthy controls (HC) and patients with Parkinson's disease (PD)
. High I-diff values imply that individuals are more similar to themselves than they are to the other subjects of the same group. High I-others values indicate high within-group similarity among the subjects of a group. The box represents data from the 25th to the 75th percentiles; the horizontal line shows the median; error lines indicate the 10th and 90th percentiles, and values falling beyond them are represented by colored dots. *, represents significant Bonferroni-corrected p-values.

F I G U R E 8
Edge-based identification rate. Identification rate (IR) for healthy controls (HC) and patients with Parkinson's disease (PD) kinectomes, for acceleration and jerk parameters in mediolateral (ML) and anteroposterior (AP) axes. The IR is computed in an iterative fashion: starting from three edges, at each iteration, one edge is added and the IR is computed. The edges were included following an order based on their contribution to the identifiability (from the most to the least contributing), as measured by the intraclass correlation analysis. The HC group exceeded the 99% identification threshold with a smaller number of edges (roughly 12), as compared to the PD patients (about 30).

Edges contribution to subjects' identification
We wondered if the loss of coordination observed in PD is a generalized phenomenon or, rather, affects the interactions of specific body segments. In the latter case, the different IR between the PD and HC groups should be due to a subset of specific edges. To test this, we calculated the IR iteratively, using each time a different number of edges (from 3 to 210) ordered from the highest to the lowest ICC rank.
Based on this, we built an IR curve describing the identification rate as a function of the number of added edges, as described. The analysis was

Nodal relevance for clinical evaluation
Following the previous analysis, we focused on the 30 edges with highest ICC value of the PD group. Indeed, since those edges were sufficient to maximize the identifiability of each patient for each parameter (AP/ML acceleration and AP/ML jerk), we assumed that they carried most subject-specific information. Hence, we counted, for each parameter, how often (across subjects) the 30 edges would be incident on any given node (i.e., body element). We observed that the T10 node in the

Network-based clinical prediction
We then wondered whether MLA-T10 could predict patient-specific motor impairment. To this end, we performed a multilinear regression analysis to predict the UPDRS scores based on the MLA-T10 degree values. 48 We included three nuisance variables in the regression model to account for confounds, such as age, sex, and education. The prediction model was validated through k-fold cross validation (k = 5), 51 to test its specificity and generalization capacity. We found that the model based on the MLA-T10 degree significantly predicts the UPDRS (p = 0.003, R 2 = 0.44) with a positive beta coefficient ( Figure 9B, left panel). That is, the higher the degree, the higher the UPDRS score.
None of the remaining predictors was significant. The agreement of the actual/predicted UPDRS scores, and the distribution of the residuals can be observed in Figure 9B (middle and right panels, respectively).

DISCUSSION
In this paper, we propose a novel approach to analyze human movement, based on the kinectome, a mathematical structure containing the pairwise interactions between different body segments during gait.
In fact, the kinectome consists of the covariance matrix of the accelerations (or the jerks) of all body segments. First, we show that the kinectome provides a thorough description of gait and distinguishes population-specific features (Figures 3 and 4). Second, the kinectome captures symmetries in the modularity of the human motion patterns, which are lost in PD patients ( Figure 5). Third, through the kinectome analysis, it is possible to identify subjects based on their gait data, using only short (∼2 s) recordings (Figures 6-8). Finally, and most importantly, the topological analysis of the kinectomes allows us to explore the role of individual biomechanical elements of the human kinematic network within a holistic, complex system perspective, and to use this information to predict clinical impairment (Figure 9). These With regard to the AP axis, the allegiance matrix of the healthy groups distinguished the passenger unit from the locomotor unit. 62 The former is composed of the head, the trunk (including the pelvis), and the arms, the latter includes the lower limbs. However, our analysis grouped together the accelerations of the arms and the legs, defining two separate communities encompassing contralateral arms and legs.
This shows the fact that our approach goes beyond purely anatomical arguments, and defines modules on functional grounds. In fact, our definition of the modules captures the fact that arms oscillate in antiphase with respect to the contralateral legs. 63  and retest sessions, we were able to identify subjects with an accuracy rate above 99%. Note that the jerk kinectomes were the most reliable in identifying individuals ( Figure 6). PD patients exhibited identifiability rates similar to those of the controls. However, the similarity within the PD group (as measured by the I-others) was lower than that within the control group (Figure 7). In other words, controls are more similar to each other than PD patients. One might speculate that a correct motor control imposes stricter constraints to the kinectome structure, which in turn produces more similar motion patterns. In pathological conditions, such as PD, such control mechanisms would fail, the constraints on the gait pattern would become looser and, hence, the patterns would be less similar to each other. This alteration in motor patterns could be related to the increase in gait variability, which, as mentioned above, is one of the gait features most commonly found in Parkinson's patients.
Next, we started exploring the contribution of individual edges to the identification (Figure 8 and Figure S8). First, it can be noticed that only a few edges suffice for an optimal recognition. Indeed, human movement interactions are fine-tuned, and only a few are sufficient to define a fingerprint of movement and, hence, to identify individuals (even when they are affected by a motor disease to assess the smoothness of trunk acceleration in PD. 29 The authors reported reduced smoothness in patients as compared to HC and, notably, a correlation with the UPDRS. All these studies, regardless of the methodological approach, highlighted the trunk acceleration as an informative element, and our results are in line with these findings. Further longitudinal studies may explore the potential of this approach in diagnostics and assessment/prediction of therapeutic responsiveness. Furthermore, we found that the MLA-T10 degree of the patients was significantly correlated with the motor impairment evaluated through the UPDRS scores ( Figure 9A). This correlation showed that the greater the motor impairment, the more the mediolateral accelerations of the upper trunk were coherent with those of the other body segments. Axial rigidity and postural abnormalities are typical features of PD that might reflect themselves into "hyperconnected" patterns. [73][74][75] Finally, we observed that the MLA-T10 degree could predict the UPDRS score at individual level, even after taking into consideration confounding variables, such as age, gender, and education.
Notably, the prediction has been validated with k-fold cross validation. This result highlighted that the hyper connectedness shown by the mediolateral trunk acceleration was strongly related to the subjectspecific impairment level. This outcome may open the possibility to use this approach to monitor at individual level both the disease progression and the effects of pharmacological therapies and rehabilitation protocols.
It should be stressed that this is the first time that the kinectome is defined. Hence, its ability to convey the individual clinical condition needs to be tested in samples, including more PD patients, as well as both in different movement disorders and in more neurological diseases. In our analysis, it was observed a lack of significant results concerning the vertical axis. One explanation could be that the vertical axis may be less informative when assessing the synchronization of movement during gait. Furthermore, the amplitude of the movements along the vertical axis is smaller as compared to the other direction, thereby making measurements less precise and, hence, less reliable. Methodologically, further analysis should be performed to explore the required number of bone markers for an optimal spatial resolution of the kinectome. In this work, we utilized 21 markers to sample the whole body. However, according to the research question at hand, it may be useful to vary the number and locations of the markers. As an example, one may also consider to focus on specific subnetworks (e.g., the lower limb network), as appropriate to test specific working hypotheses. Finally, in this work, we only consider pairwise interactions, future work should also consider higher-order interactions. 76 In conclusion, we have proposed a network approach to motion analysis, which identified several disease-related features, both at global and individual levels. It must be noted that the methodology underlying our approach is grounded in network science, which is a solid branch of mathematics. We believe that the application of such methodology in the motion analysis opens new possibilities, especially in the domain of clinical applications, where it allows disease-specific motor features examination. In particular, we think that the kinectome represents an informative tool, which encodes a large number of features that can be extracted, depending on the specific research question at hand. In this work, we started from a global examination, we then zoomed in to describe the characteristics of the kinectome that emphasize individual features, up to revealing a kinematic element (i.e., the MLA-T10), which captures many of the significant features of PD. Further studies may focus on the usefulness of exploiting the kinectome to support pharmacological and rehabilitative treatments.
Finally, this study limited itself to a first application, but the kinectome is a powerful tool that can be applied in a wide spectrum of frameworks. Useful applications could be in the differential diagnosis between PD and atypical parkinsonisms (e.g., multiple system atrophy, progressive supranuclear palsy, and corticobasal syndrome). More in general, this approach might be useful to capture pathophysiological changes that manifest in subtle alterations of the interactions among multiple body parts. These kinds of patterns are easily missed by the human eye, and might greatly help doctors to refine diagnosis by taking this information into account. Finally, other fields of applications will be in biomedical engineering for the construction of robots, or to design motor assistance devices.

ACKNOWLEDGMENTS
G.S. acknowledges the financial support from the University of Naples