Transformer-Based Automated Skill Assessment and Interpretation in Robot-Assisted Surgery

Different artificial intelligence approaches have been made to automatically assess skills during robotic surgical training. However, limitations still exist in these studies, including issues related to feature engineering, cross-validation methods, complex model architectures, and interpretability. In response to these limitations, this study introduces a Transformer-based model that processes kinematic data and identifies surgical skill levels. The model performance was rigorously evaluated under the Leave-One-User-Out cross-validation method, resulting in a classification accuracy of 80%. Beyond skill level classification, this study also explores deeper into the interpretability aspect. It includes the extraction of global-attention from the model, providing insights into the significance of each part or gesture within a task during the classification decision-making process. This interpretability holds the potential to help surgeon improve their skill by offering a comprehensive and detailed understanding of their performance.