Attention-Based Deep Sequential Models for Adhesive Capsulitis Classification Using a Single Azure Kinect-D Camera

Adhesive capsulitis (AC) is a debilitating condition characterized by significant shoulder pain and stiffness. This study proposes a novel attention-based deep sequential framework for classifying AC using a single Azure Kinect-D camera to capture shoulder abduction movements. The framework integrates attention mechanisms into state-of-the-art recurrent neural networks (LSTM, BiLSTM, and GRU), allowing the network to focus on clinically significant movement patterns. Among the models, The Att-GRU model achieved the highest performance metrics, with 98.51% accuracy, 98.28% precision, 100% specificity, 98.26% recall, 98.85% F1-score, and an AUC of 0.992, demonstrating exceptional diagnostic accuracy and reliability. Attention weight analysis revealed distinct movement dynamics in healthy and AC individuals, highlighting the model's ability to recognize disease-specific patterns. The proposed framework offers a non-invasive and objective tool for frozen shoulder diagnosis with significant clinical potential.