A summary index derived from Kinect to evaluate postural abnormalities severity in Parkinson’s Disease patients

Postural abnormalities are common disabling motor complications affecting patients with Parkinson’s disease (PD). We proposed a summary index for postural abnormalities (IPA) based on Kinect depth camera and explored the clinical value of this indicator. Seventy individuals with PD and thirty age-matched healthy controls (HCs) were enrolled. All participants were tested using a Kinect-based system with IPA automatically obtained by algorithms. Significant correlations were detected between IPA and the Movement Disorder Society-Sponsored Revision of the Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) total score (rs = 0.369, p = 0.002), MDS-UPDRS-III total score (rs = 0.431, p < 0.001), MDS-UPDRS-III 3.13 score (rs = 0.573, p < 0.001), MDS-UPDRS-III-bradykinesia score (rs = 0.311, p = 0.010), the 39-item Parkinson’s Disease Questionnaire (PDQ-39) (rs = 0.272, p = 0.0027) and the Berg Balance Scale (BBS) score (rs = −0.350, p = 0.006). The optimal cut-off value of IPA for distinguishing PD from HCs was 12.96 with a sensitivity of 97.14%, specificity of 100.00%, area under the curve (AUC) of 0.999 (0.997–1.002, p < 0.001), and adjusted AUC of 0.998 (0.993–1.000, p < 0.001). The optimal cut-off value of IPA for distinguishing between PD with and without postural abnormalities was 20.14 with a sensitivity, specificity, AUC and adjusted AUC of 77.78%, 73.53%, 0.817 (0.720–0.914, p < 0.001), and 0.783 (0.631–0.900, p < 0.001), respectively. IPA was significantly correlated to the clinical manifestations of PD patients, and could reflect the global severity of postural abnormalities in PD with important value in distinguishing PD from HCs and distinguishing PD with postural abnormalities from those without.


INTRODUCTION
Postural abnormalities are disabling motor complications affecting patients with Parkinson's disease (PD) and become increasingly severe as the disease progresses [1][2][3] . The most recognized type of postural abnormalities in PD patients is the classic stooped posture, with flexion of the hips and knees, and rounding of the shoulders, which differs them from general population 2,4 . Moreover, an important subset of patients present with more severe abnormalities of spinal alignment including sagittal abnormalities: camptocormia and anterocollis 5,6 ; frontal abnormalities: Pisa syndrome and scoliosis 2,7 . Some patients even suffer from a combination of several types of postural abnormalities. Since postural abnormalities in PD subjects usually develop insidiously over months to years before they become obvious 8,9 , early recognition of them facilitates the diagnosis of the disease, as well as the prompt intervention to avoid worse outcomes.
In clinical practice, there are various methods for evaluating abnormal posture of PD, such as clinical scale, wall goniometer, and photo-based measurement [10][11][12] . The most commonly used clinical scale for evaluating abnormal posture of PD is the 13th item of the third part of Movement Disorder Society-Sponsored Revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS-III 3.13). However, the 5-class ordinal scale gives only a broad classification of postures and is not suitable for a more detailed description because of its insensitivity to small changes which may be clinically relevant 10,13 . In addition, simple geometric parameters, angles, are generally examined to evaluate postural abnormalities in PD. Conventionally, three common methods, including total camptocormia (TCC) angle, upper camptocormia (UCC) angle, and lower camptocormia (LCC) angle are used to assess the severity of camptocormia 6,12,14 . Similarly, the drop head angle (DHA) characterizes the severity of anterocollis while the Pisa angle or the lateral trunk bending (LTB) angle serves to assess PD patients with LTB like Pisa syndrome and scoliosis 4,[15][16][17] . Though the wall goniometer method and photo-based measurement method can provide accurate and quantitative measurement of the above angles, they only reflect the severity of postural abnormalities of a certain plane and do not offer a global assessment of postural orientation quality of the patients 4,12,15,16,18,19 . The global and quantitative assessment of posture is urgently required to monitor the progress of the disease and measure treatment effects.
In the last decade, three-dimensional (3D) stereophotogrammetry including Kinect depth camera has been widely used to provide objective information about main joint motions on the three planes of movement [20][21][22] . To further facilitate the interpretation of these large amount of kinematic data, indices such as Gait Profile Score (GPS) 23,24 , Trunk Profile Score (TPS) 25 , and Arm Profile Score (APS) 26 have been proposed and show promising clinical values. Inspired by these attempts, we recently developed an intelligent evaluation system to assess postural abnormalities in PD based on Kinect and machine learning 27 . The automated and accurate assessment of postural abnormalities for each PD patient was realized with only six selected features F1, F2, F3, F4, F5, and F7 ( Fig. 1). Their feature importance in the constructed decision tree model was 13.2%, 12.6%, 16.5%, 11.3%, 6.7%, and 40% severally 27 . Actually, F2, F4, and F5 here shared the same definitions of LTB, TCC, and LCC angles, respectively, which reflected the severity of trunk abnormality 12,16 . F1 and F3 reflected the severity of head and neck abnormality similar to DHA 4,16 . F7 was a normalized feature which offered a general assessment in patients with both lower and upper camptocormia 27 . In this study, we proposed a summary index, the index for postural abnormalities (IPA), which was a combination of all these features (F1, F2, F3, F4, F5, and F7). We explored the correlations between IPA and other clinical manifestations of PD patients and discriminated between participants with different severity of postural abnormalities. In our hypothesis, the IPA is a useful method to assess the clinical severity of postural abnormalities in PD globally with important value in distinguishing PD from healthy general population and distinguishing between PD with and without postural abnormalities.

DISCUSSION
In this paper, a summary index, the IPA was introduced for quantifying the postural abnormalities of PD patients and comparing them with HCs. This IPA was obtained based on objective kinematic data derived from Kinect depth camera and it had the merit to summarize the overall quality of an individual's trunk segment alignment during quiet standing. Notably, the IPA showed important value in distinguishing PD from HCs as well as distinguishing PD with postural abnormalities from those without. In a word, IPA is a useful synthetic index for evaluating postural abnormalities in PD. Specifically, in this study, the IPA performed surprisingly well in distinguishing PD from HCs with an AUC of 0.999 (0.997-1.002, p < 0.001) and adjusted AUC of 0.998 (0.993-1.000, p < 0.001). Meanwhile, it provided moderate to high accuracy in distinguishing PwPA from PwtPA with an AUC of 0.817 (0.720-0.914, p < 0.001) and adjusted AUC of 0.783 (0.631-0.900, p < 0.001). Some researchers have been trying to distinguish PD from healthy individuals or to distinguish PD of different severity stages with the help of kinematic data derived from patients' gait, posture, and fine movements of the limbs 22,28-30 . Mirelman et al collected multiple gait characteristics obtained with multiple wearable sensors to classify PD motor stages using machine learning methods and they found discriminatory values between motor disease stages with mean sensitivity in the range 72-83%, specificity 69-80%, and AUC 0.76-0.90 31 . A study used normalized stride length (SL) and gait velocity (GV) for recognizing PD using Microsoft Kinect and achieved a high accuracy rate of 97.2%. Though the sample size was relatively small with 18 PD patients and 18 HCs, it suggested the potential use of Microsoft Kinect image and depth sensors for these applications 29 . In this study, we focused on assessment of the global quality of postures and proposed the novel summary index of IPA. Our finding indicated that IPA has important value in distinguishing PD from healthy general population and distinguishing between PD with and without postural abnormalities, which may be helpful in early recognition of PD and early intervention of postural abnormalities for patients.
Previous studies have reported the characteristics of postural abnormalities in PD. For instance, Margraf et al. compared TCC, UCC, and Pisa angles of 192 PD patients and 78 HCs with the free NeuroPostureApp© (http://www.neuroimaging.unikiel.de/NeuroPostureApp) 19 . They found that PD patients had a worse posture than HCs in all three angles (p < 0.001). For the TCC angle, 39.1% of the patients had a normal posture (<17.4°), 47.9% a presumed stooped posture (>17.4°, <30.2°), and 6.3% had camptocormia (>30.2°). A large comparative study with general population (GPP) revealed that the dropped head angle (DHA), anterior flexion angle (AFA), and lateral flexion angle (LFA) of the thoracolumbar spine were 21.70 ± 14.40°, 13.13 ± 10.79°, and 5.98 ± 12.67°for PD patients and −3.82 ± 4.04°, 0.86 ± 4.25°, and 1.33 ± 2.16°for age-matched GPPs, respectively 4 . Our study disclosed significantly higher values of feature F1, F2, F3, F4, F5, and F7 in PD compared to HCs which was consistent with that study. However, all these features alone only reflect the severity of postural abnormalities of a certain plane. To summarize the global quality of an individual's body segment alignment during quiet standing, another study introduced the Postural Profile Score Index (PPS) 20 . Twelve joint angles of trunk and of lower limbs, considered representative of the whole-body posture were acquired. The root mean square difference between them and those of the unaffected participants (the Postural Variable Score, PVS) were computed. Then, the PPS was calculated as a combination of the selected PVSs. The authors found significant difference in PPS between PD and HCs (8.59°vs. 6.11°, p < 0.001) but did not reveal any correlation of PPS with respect to UPDRS-III or Hoehn-Yahr scale, nor did they propose the cut-off value to distinguish PD from HCs 20 . In this study, the IPA was a combination of the selected features (Fig. 1) and the corresponding feature importance. Therefore, it would be an ideal candidate index for well evaluating and reflecting the overall progress of postural abnormalities in PD from the very beginning.
Moreover, significant weak to moderate correlations were detected between IPA and clinical measurements such as many composite scores of MDS-UPDRS, PDQ-39 score and BBS score (shown in Table 2). Results of linear regression models further confirmed that IPA significantly contributed to MDS-UPDRS total score, MDS-UPDRS-III total score, MDS UPDRS-III 3.3, 3.7, 3.8, and 3.13 score, and PDQ-39 score. Among them, MDS UPDRS-III 3.13 score was mostly explained by the IPA (adjusted R 2 = 0.351, p < 0.001), indicating IPA is effective in judging postural abnormalities of PD patients. The MDS-UPDRS together with its composite scores is the most commonly used clinical scale for PD and has been recognized as a reliable and valuable tool to assess the severity of PD 13,[32][33][34][35] . The PDQ-39 can provide a summary score of the impact of the illness on functioning and well-being and will be useful in the evaluation of the overall effect of different treatments [36][37][38] . The BBS has been validated to be a valuable screening tool and ongoing assessment tool for patients with PD 39,40 . Results from an observational study of 283 PD patients with ≥5°of forward trunk bending (FTB), lateral trunk bending (LTB), or forward neck bending (FNB) revealed that degree of trunk bending was associated only with motor impairment in LTB (odds ratio [OR], 1.12; 95% confidence interval [CI], 1.03-1.22). ROC curves showed that patients with LTB of 10.5°might have moderate/severe motor impairment 16 . Another two studies demonstrated that camptocormia, antecollis, and Pisa syndrome were associated with severe impairment of neck and back functions, as well as pain in PD patients 41,42 . In our study, weak to moderate correlations were identified between IPA with common PD clinical scales, which indicates that IPA can reflect the clinical severity of postural abnormalities in PD to some extent.
The present study has some limitations. First, the sample enrolled was only composed by a total of 100 participants and it could not be representative of the general population. Second, we did not further verify the results of the study by carrying out prospective researches. In the future, more studies with a much larger sample size and well-controlled homogeneity of participants will be necessary in order to validate this index as a reliable tool in PD patients' evaluations.
Despite these noted limitations, there are several highlights of this study. First of all, we proposed cut-off values to distinguish PD from HCs and PD with postural abnormalities from those without via global assessment of postures. Moreover, the IPA was obtained based on objective kinematic features derived from Kinect depth camera and computer algorithm which was accurate and repeatable 30,43 . In addition,

Device and testing
There are many researchers assessing posture based on Kinect with satisfactory validity and reliability obtained [45][46][47] . In this study, a Kinectcentered motion analysis device integrating a Kinect v2.0 depth camera (RGB 1920 × 1080 pixels @30fps, depth camera 512 × 424 pixels @30fps, 4-microphone linear phased array, Microsoft) and an independent computer that ran a data capture program was developed by iFLYTEK Suzhou Research Institute 27 . The participants were asked to stand directly in front of the Kinect camera (at a distance of 2 meters) 48 at ease for 5 s and then actively correct their abnormal posture for 5 s. After that, they were asked to turn left for 90°, relax and stand for 5 s, and then actively correct their abnormal posture for another 5 s. Close-fitting clothing were required and long hair should be tied up 27,49,50 . After the recording, the values of the six selected features F1, F2, F3, F4, F5, and F7 were automatically obtained by computer algorithms. The definition of these features were illustrated in Fig. 1. IPA was calculated with an equation, which was set as follows: IPA ¼ F1 13:2% þ F2 12:6% þ F3 16:5% þ F4 11:3% þ F5 6:7% þ F7 40:0% In other words, IPA was defined as the sum of the products of each selected feature and the corresponding feature importance 27 . Demographic and clinical information such as gender, age at admission, onset age, first symptom (tremor or rigidity) and disease duration were collected. All the PD patients were assessed with the following scales: Hoehn-Yahr scale, the Movement Disorder Society-Sponsored Revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS), the Mini-Mental State Examination (MMSE), the Non-motor Symptoms Scale (NMSS), the Pittsburgh Sleep Quality Index (PSQI), the Constipation Severity Instrument (CSI), the Hospital Anxiety and Depression Scale (HADS), the Berg Balance Scale (BBS), and the 39-item Parkinson's Disease Questionnaire (PDQ-39). The sub-scores for tremor (items 3.15, 3.16, 3.17, and 3.18), bradykinesia (items 3.4, 3.5, 3.6, 3.7, 3.8, 3.9, and 3.14), and gait (items 3.10, 3.11) were obtained from the MDS-UPDRS-III which were referred to as MDS-UPDRS-III T, MDS-UPDRS-III B, and MDS-UPDRS-III G, respectively. All PD patients performed the assessments 30 min to 2 h after medication intake (based on the participant's feedback when they usually experience best ON) in ON medication condition 51 . The PD patients were further divided into group with postural abnormality (PwPA) and group without postural abnormality (PwtPA) based on whether the MDS-UPDRS-III 3.13 item was greater than or equal to 2 points 10 .

Statistical analysis
Quantitative data were shown as mean ± standard deviation (SD). The normality of distribution of demographic and clinical data was initially tested using the Kolmogorov-Smirnov test. The student's t test and the Mann-Whitney U test were used for comparison of normally and abnormally distributed data, respectively. The Chi-square test was used to evaluate the differences in categorical variables. Spearman's correlation analysis was conducted to explore correlations between IPA and other clinical data. A correlation coefficient of 0.00-0.10 indicates negligible correlation, 0.10-0.39 indicates weak correlation, 0.40-0.69 indicates moderate correlation, 0.70-0.89 indicates strong correlation, and 0.90-1.00 indicates very strong correlation 52 . Linear regression was used to determine the proportion of the variance in clinical measurements explained by the IPA. Receiver operating characteristic (ROC) curves were operated to estimate the optimal IPA cut-off values which were determined by maximizing the Youden's index on the ROC curve. An area under the curve (AUC) value <0.7 indicates a low diagnostic accuracy, 0.7-0.9 indicates moderate accuracy, and >0.9 indicates high accuracy 53 . Statistical analyses were performed using GraphPad Prism version 7 (Graph Pad Software Inc, San Diego, CA, USA) and Stata version 16 (StataCorp, College Station, TX, USA), and significance was set at a twotailed p value < 0.05.

DATA AVAILABILITY
The data that support the findings of this study and the algorithms to extract the features mentioned in the article are available from the corresponding author upon reasonable request.