AW-TPANet: A Clinical-Prior-Driven Dual-Task Network with Adaptive Windows for Upper Limb Rehabilitation Assessment

This capsule contains the source code for the paper "AW-TPANet: A Clinical-Prior-Driven Dual-Task Network with Adaptive Windows for Upper Limb Rehabilitation Assessment", submitted to the IEEE Journal of Biomedical and Health Informatics (JBHI). AW-TPANet is a cascaded dual-task framework that integrates Anatomically Constrained Spatial Encoding with Temporal Point Attention (TPA) and Adaptive Window (AW) mechanisms for simultaneous action classification and continuous standardness scoring. Important Note on Data & Reproducibility: Due to strict patient privacy and ethical restrictions, the original clinical dataset (RehabU-15) cannot be publicly shared. To demonstrate the reproducibility and execution pipeline of our code, we have provided a simulated dummy dataset () containing randomly generated tensors.fixed_30frame_dataset.pt The purpose of this Code Ocean capsule is to prove that the proposed AW-TPANet architecture is correctly implemented, and that the complete pipeline (data partitioning, model training) runs successfully without errors. Please note that the accuracy and loss metrics outputted in this run are based on the dummy data and do not reflect the actual state-of-the-art performance reported in our manuscript.