Author(s): ROGGIO, F., TROVATO, B., MUSUMECI, G., Institution: UNIVERSITY OF PALERMO: UNIVERSITA DEGLI STUDI DI PALERMO, Country: ITALY, Abstract-ID: 1185

INTRODUCTION:
Various methods exist to evaluate human posture, each with unique benefits and drawbacks. Recently, markerless methods combined with machine learning techniques have become a promising alternative due to their efficiency. MediaPipe, a sophisticated machine learning algorithm created by Googles research team, is engineered for precise tracking of body posture, capable of identifying 33 landmarks throughout the human body [1]. Other studies validated its accuracy in joint tracking technique against the gold standard Qualisys motion capture system, providing Pearson’s correlation coefficients of 0.80 ± 0.1 for lower limbs and 0.91 ± 0.08 for upper limbs, indicating so, MediaPipes high reliability in estimating joint angles [2]. This study aimed to demonstrate the applicability and reliability of a machine learning method for the analysis of posture, provide normative data on the posture of healthy men and women, and also to investigate potential new posture patterns with the principal component and cluster analyses.
METHODS:
Healthy participants (n=192) with a mean age of 26.4 ± 1.2 years, were photographed in the frontal, dorsal, and lateral views. The Student’s t-test and Cohen’s effect size (d) were used to evaluate sex differences in the postural parameters assessed. The Intraclass Correlation Coefficient (3,k) was used to evaluate the reliability of the measures in a subgroup of participants. We also conducted multivariate statistical techniques to highlight any alternative pattern within the sample. Both Principal Component Analysis and five clustering methods categorized participants into two distinct groups (CG1, CG2).
RESULTS:
The data revealed sex differences, such as shoulder adduction angle (men = 16.12° ± 1.92°, women = 14.13° ± 1.53°, d= 1.14) and hip adduction angle (men = 9.90° ± 2.22°, women = 6.71° ± 1.53°, d= 1.67) while no sex differences were present for horizontal inclinations. The ICC results ranged from 0.67 to 0.95 supporting the reliability of the measures. The clustering analysis highlighted a shoulder-hip difference: 86.58 ± 10.73 normalized pixel distance for CG2 vs 57.51 ± 7.02 normalized pixel distance for CG1, with a significant effect size of d= 3.21.
CONCLUSION:
This study introduces a novel unsupervised machine learning approach for postural analysis, offering significant clinical benefits by identifying specific postural categories and reducing subjectivity in evaluations. This method, efficient and non-invasive, has potential applications in physical therapy, ergonomics, and sports, aiding in personalized treatment plans.

1. Lugaresi C, et al. Mediapipe: A framework for building perception pipelines. arXiv preprint arXiv:190608172. 2019;
2. Lafayette TBdG, et al. Validation of Angle Estimation Based on Body Tracking Data from RGB-D and RGB Cameras for Biomechanical Assessment. Sensors. 2023;23:3.