Background: The Functional Movement Screen (FMS) is widely used for movement assessment but suffers from subjective scoring that leads to inconsistent evaluations. While previous studies have focused on reliability, the validity of AI-supported assessment remains unexplored. Objectives: To evaluate the reliability and validity of an AI-based motion analysis system using MediaPipe for three FMS movements. Design: Prospective reliability and validity study with repeated measures. Methods: Thirty healthy adults (age 23.4±2.8 years) performed three FMS tests (Deep Squat, Hurdle Step, Inline Lunge) recorded on video. Three evaluators (two experienced physical therapists and one novice) assessed recordings in three phases: Phase 1 involved traditional assessment by experts only to establish criterion reference, Phase 2 had all evaluators using AI support, and Phase 3 consisted of repeated AI-supported assessment. The AI system provided real-time visual feedback of joint angles and alignment through MediaPipe skeletal tracking. Results: Criterion validity showed strong agreement between traditional expert assessment and AI-supported assessment (r=0.94, P<.05). Inter-rater reliability improved from good (ICC=0.89) to excellent (ICC=0.91) with AI support. The novice evaluator achieved immediate expert-level performance with only 0.05 points difference from experts. Intra-rater reliability was excellent for all evaluators (ICC=0.84-0.89). Conclusion: The AI-based system demonstrated strong validity and improved reliability for fundamental movement assessment. While AI support enabled novice evaluators to achieve expert-level performance immediately, it may increase sensitivity to subtle movement variations. This technology shows promise for standardizing movement screening, though current limitations restrict its application to standing movements.
Background: The increasing demand for real-time professional fitness coaching has led to a need for accurate exercise posture recognition using artificial intelligence. Objectives: To compare the performance of Feedforward Neural Network (FNN) and Stacked Long Short-Term Memory (LSTM) models in classifying fitness posture images using detailed joint coordinate labeling. Design: Comparative analysis of machine learning models using a labeled dataset of fitness posture images. Methods: A dataset from AI-hub containing images and data of 41 exercises was used. Five exercises were selected and processed using a custom program. Data was converted from JSON to CSV format, augmented with joint condition information, and analyzed using Google Colab. Results: The best FNN model achieved a training error of 1.21% and test error of 9.08%. The Stacked LSTM model demonstrated superior performance with a training error of 1.05% and test error of 6.09%. Conclusion: Both FNN and Stacked LSTM models effectively classified sequential fitness images, with Stacked LSTM showing superior performance. This indicates the potential of Stacked LSTM models for accurate fitness posture classification in real-time coaching scenarios.
Background: Artificial intelligence (AI) research on physical fitness posture estimation has been limited by a lack of comprehensive datasets and guidelines. This study analyzes the fitness image dataset provided by Korea's AIHub platform to advance posture estimation algorithms from exercise prescription and behavioral analysis perspectives. Objectives: To analyze fitness movements and guide correct exercise posture using AI-based visual and auditory feedback. Design: Descriptive analysis of a large-scale dataset. Methods: The study examined image and JSON labeling files from AI-Hub, analyzing 6.39 million fitness images across 41 exercise types. Data structure, exercise states, and annotation characteristics were analyzed in detail. Results: The dataset encompasses 816 distinct exercise states, captured from five camera angles with 24 key body points labeled per posture. Exercises were categorized into full-body workouts (17), barbell/dumbbell exercises (16), and furniture exercises (8). Gender distribution was 76% male and 24% female, with 41% in the 27-29 age group. The dataset allows for detailed analysis of correct and incorrect postures. Conclusion: This comprehensive analysis of the AI-Hub fitness dataset provides a robust foundation for developing AI models for fitness posture evaluation and feedback, benefiting exercise coach web/app service developers.