This study was approved by the Institutional Review Board of Gyeongsang National University Hospital, Jinju, Korea (approval no. 2020-09-007 and 2022-04-027), and written informed consent were obtained from all patients. Patients were enrolled from December 2020 to May 2023 among those who visited the hospital. We enrolled patients with Parkinson’s Disease at Hoehn and Yahr (HY) stage ≤ 2 from the Movement Disorder Clinic of Gyeongsang National University Hospital, Jinju, Korea. The diagnosis of Parkinson’s Disease was based on the Movement Disorder Society (MDS) Clinical Diagnostic Criteria for Parkinson’s Disease [14]. Patients with Parkinson’s Disease suffering from other neurological diseases such as stroke or neuromuscular disorders, and having a history of spine surgery or spinal cord injury were excluded. Those who were accompanied by hyperthyroidism, cancer, tuberculosis, liver cirrhosis, a renal disease requiring hemodialysis, and obesity (body mass index BMI ≥ 30 kg/m²) that could lead to muscle wasting were also excluded.

Clinical information and sarcopenia profiles of patients with Parkinson’s Disease were collected as follows: sex; age at Parkinson’s Disease onset; disease duration; MDS-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS) III and HY stage during “on” state [15, 16]; Tremor and PIGD scores, calculated by the sum of MDS-UPDRS items 2.10, 3.15, 3.16, 3.17, and 3.18, and the sum of MDS-UPDRS items 2.12, 2.13, 3.10, 3.11, and 3.12, respectively [17]; levodopa equivalent daily dose (LEDD); right and left asymmetry, present if right-left differences in the MDS-UPDRS items 3.3, 3.4, 3.6, 3.8, 3.15, 3.16, and 3.17b-e were ≥ 5 points [18].

To classify patients with sarcopenia among those with Parkinson’s disease, we applied the criteria set forth by the Asian Working Group for Sarcopenia (AWGS 2019). This entailed the utilization of the following standards: low appendicular skeletal muscle mass index (ASMI) for males (< 7.0 kg/m2) and females (< 5.7 kg/m2), coupled with low handgrip strength (< 28 kg for males, < 18 kg for females), or a low 6-meter walking velocity (< 1.0 m/s) [19]. The ASMI was calculated as the sum of skeletal muscle mass in the upper and lower extremities divided by the square of height, which was assessed by bioelectrical impedance analysis (BIA, InBody770, InBody Co., Ltd, Seoul Korea). Handgrip strength was measured by a Smedley-type dynamometer (Takei, Japan).

Participants were asked to walk a distance of 6-meters twice without any set velocity limit in order to acquire videos for pose estimation analysis. Due to the absence of a standardized protocol for recording angles, measurement methods and tools, a height of 90cm corresponding to the abdominal region was selected as the designated recording protocol. Subsequently, to ensure sufficient inclusion of patients’ movements within the video frame, cameras were positioned at 1-meter intervals along the central axis of the 6-meter linear path, both for lateral views and frontal perspectives. For frontal views, cameras were also situated at 1-meter intervals starting from the destination point. Recording was conducted using a smartphone (LM-V510N, LG Electronics Inc., Republic of Korea).

The gait pattern was analyzed using an application called ’Deevo.gait,’ which was integrated into the Dr.Log App. The Deevo.gait application utilizes Google’s Blazepose model, known for its accurate estimation of joint positions within the body. With the Blazepose model, keypoints representing shoulder, elbow, wrist, hip, knee, and ankle joints on both sides are identified from the video after the body is detected. Videos in which the patient’s body couldn’t be reliably detected, such as those containing two or more individuals, were excluded from the study. The keypoint coordinates consist of 3D data encompassing x, y, and z values, derived from the fusion of coordinates captured by the two smartphone cameras. Finally, clinical-related variables and gait function-related variables generated from pose estimation coordinate data were selected, and detailed information about them is presented in S1 Data.

When comparing the groups with and without sarcopenia, continuous clinical-related and physical function-related variables were expressed as ’mean ± standard deviation,’ while categorical variables were presented as ’number of patients (%).’ For the analysis of continuous variables between groups, the Shapiro-Wilk test was conducted to assess normality. In cases where variables met the assumption of normal distribution, Student’s t-test was employed; when normality assumptions were not met, the Mann-Whitney U test was utilized to determine the significance of inter-group differences.

To investigate correlations between variables, the Pearson’s chi-square test was applied when distribution assumptions were met. If not, Fisher’s exact test was used. A significance level of P < 0.05 was considered statistically significant.

Ultimately, a random forest model was employed to construct a predictive model for sarcopenia in patients with Parkinson’s Disease. The entire patient cohort was randomly divided into training (75.0%) and testing (25.0%) sets. In the process of dividing the data, allocation was randomized, and the training was iterated 10 times. Subsequently, the results for training and testing from these 10 iterations were aggregated to compute the confusion matrix, from which we calculated precision, recall, specificity, accuracy, and f1-score. Additionally, Gini importance, which calculates variable importance, was utilized to identify variables that discriminate sarcopenia using the testing set.

All figures were created and statistical analyses were conducted using R version 4.1.2 (R Core Team, R Foundation for Statistical Computing, Vienna, Austria, 2021). The ggplot2 package was used for visualizations, the random forest package for analysis, and the dplyr package for data preparation. Furthermore, the variables used in the random forest analysis and their corresponding descriptions are as presented in Table 1.

Overall, 38 patients with Parkinson’s Disease were enrolled in this study, among which, 9 (23.7%) were with sarcopenia whereas 29 (76.3%) were without sarcopenia (sarcopenia versus non-sarcopenia group, respectively). Sarcopenia profiles, baseline demographic, and clinical information are presented in Table 2. Patients with Parkinson’s Disease and sarcopenia demonstrated higher frequencies of low ASMI and low handgrip strength whereas 6-meter walking velocity and clinical characteristics were comparable between the groups.

After dividing the data into 75% for training and 25% for testing, the confusion matrix results from the random forest model, obtained after 10 iterations, are as displayed in Fig 1. Additionally, the performance metrics derived from this confusion matrix are presented in Table 3. The performance metrics delineated in Table 3 encapsulate the random forest model’s proficiency. For the training set, precision is exceptionally high at 0.981, indicating that the model accurately identifies the positive class, while recall at 0.680 suggests that some positive cases are missed. Specificity is nearly perfect at 0.995, demonstrating the model’s effectiveness in ruling out negative cases. The accuracy of the training set stands at 0.911, showing a high overall rate of correct classifications, and the F1-Score at 0.803 indicates a strong balance between precision and recall. In the test set, the model achieves perfect precision and specificity, signifying no false positives or negatives were identified. The recall is slightly lower at 0.667, pointing to the potential for missed positive cases. The accuracy climbs to 0.949, denoting a very high level of correct predictions across the board, and the F1-Score remains high at 0.800, reflecting the model’s consistent performance in predicting sarcopenia.

Building upon the previous analysis, factors significantly correlated with sarcopenia in Parkinson’s Disease patients were identified. The average hip angle (hip_mean), a measure of the mean angle at the hip joint during movement, showed a notable negative correlation (r = -0.429), suggesting that a smaller mean hip angle is associated with sarcopenia. Similarly, the ankle_dif, which captures the change in ankle angle from a stationary position to its maximum during walking, had a positive correlation (r = 0.429). This indicates that greater changes in ankle movement correlate with the presence of sarcopenia. The all_max_dif, representing the overall maximum difference in joint angles during walking, had a correlation (r = 0.424) close to that of the ankle, signifying its relevance. Moreover, the kneeankle_dif, denoting the combined difference in knee and ankle angles from standing still to maximum movement, was also positively correlated (r = 0.407). Expanding on other variables, a low Appendicular Skeletal Muscle Mass Index (low_ASMI) showed a very strong positive correlation (r = 0.932), indicating that lower ASMI values are highly indicative of sarcopenia. This index is critical as it benchmarks muscle mass against height squared, providing thresholds for sarcopenia diagnosis in males and females. Weight and Height were inversely correlated with sarcopenia, with coefficients of -0.651 and -0.649, respectively, implying that lower body weight and stature are associated with higher sarcopenia prevalence. Handgrip strength, an assessment of muscle strength, was negatively correlated (r = -0.543) with sarcopenia, highlighting the condition’s impact on muscular function. Lastly, the absolute value of ASM displayed a negative correlation (r = -0.498), further underlining the relationship between muscle mass and sarcopenia. These findings are visually synthesized in Fig 2.

The importance of variables for distinguishing sarcopenia in Parkinson’s Disease, as determined through the random forest analysis, was ranked with Low_ASMI and Handgrip_strength at the forefront, followed by Kneeankle_dif, Ankle_dif, Hip_min, among others. Excluding Low_ASMI and Handgrip_strength, which are directly related to sarcopenia criteria, the significance of the remaining variables suggests that the differences in ankle, hip, and knee angles from a standing to walking onset are of notable importance (Fig 3).