Clinical Assessment of Adhesive Capsulitis Severity Using Transformer-Based Analysis of 3D Shoulder Motion

Author : ShiUk Lee, Konki Sravan Kumar, Kyuwon Lee

Abstract : Adhesive capsulitis (AC) is a musculoskeletal disorder causing pain, stiffness, and restricted shoulder motion. This study introduces Shoulder Geometric Pain Transformer (ShoulderGPT), a non-wearable, non-invasive deep learning framework for automatic AC severity classification using 3D shoulder kinematics from a single Azure Kinect depth camera. The proposed model combines Temporal Convolutional Networks (TCN) for fine-grained temporal dynamics and Transformer encodersfor long-range dependency modeling across multiple movement cycles. A dataset of 157 participants performing four shoulder movements was analyzed. ShoulderGPT model achieved the highest performance metrics, with 97.54% accuracy, 97.23% F1-score, and 100% specificity, outperforming attention-based LSTM and GRU models. Attention visualizations highlighted clinically relevant motion cycles, supporting interpretability. The proposed framework offers a robust, explainable, and patient friendly tool for automated AC severity assessment and clinical decision support.

Keywords : Adhesive capsulitis, kinematic analysis, temporal convolutional networks, transformer encoder, shoulder rehabilitation monitoring.

Conference Name : International Conference on Neuromuscular Re-education and Functional Recovery (ICNRFR-26)

Conference Place : Pattaya, Thailand

Conference Date : 4th May 2026