Camera-Based Smart Mirror with Machine Learning for Postural Analysis: System Development and Reliability Evaluation

Early postural assessment using camera-based systems remains technically challenging due to variability in user positioning and limited evaluation of measurement repeatability. This study presents the development and repeatability evaluation of a smart mirror system for automated postural analysis using pretrained pose estimation and rule-based geometric classification. The system consists of a fixed camera mounted above a mirror and a connected computing device for real-time processing and visual feedback. Anatomical landmarks were detected from standardized anterior, posterior, and lateral views using an AI-based pose estimation model, and postural asymmetry was quantified using bilateral distance ratios and angular deviation thresholds derived from literature. Reliability was evaluated through repeated measurements to assess the consistency of landmark detection and postural classification outputs. Forty adolescents (age 12.8 ± 0.56 years; 28 males, 12 females) participated in present study. The system intra-rater reliability was evaluated by calculating Intraclass Correlation Coefficients (ICC) for the landmark data and Cohen's Kappa for posture classifications. The system demonstrated excellent reliability for key landmarks in scapula (ICC = 0.98, 95%CI 0.97-0.99) and hip-knee-ankle (ICC = 0.98, 95%CI 0.98-0.99). The classifications for scoliosis assessment also showed excellent agreement (κ = 0.90). These results indicate that the proposed system can produce repeatable posture measurements under controlled conditions; however, this study evaluates repeatability only and does not assess diagnostic accuracy or clinical validity. Further validation against clinical reference standards is required before broader application.