Revisiting Tabular Data for Interventionalist's Action Recognition towards Improved Endovascular Robots

Abstract

Interventionalist (IC) action recognition and skill learning are essential for improving procedural quality and safety during endovascular catheterization interventions. Traditional trainee assessments are subjective and lack immediate feedback, limiting opportunities for rapid performance improvement. To address these limitations, this study introduces a deep-learning framework designed to systematically analyze ICs' action data, address inherent class imbalances, and enable real-time action recognition. First, the proposed framework leverages generative models to augment minority action classes, thus enhancing data representation and ensuring accurate recognition of catheterization actions. Synthetic data generated by six distinct generative models underwent rigorous evaluation, achieving high fidelity with average precision and F1scores exceeding 94% across all models except CTGAN. Secondly, we developed a convolutional neural network (IAR-Net) tailored to recognize seven distinct catheterization actions. Evaluated using the augmented dataset, IAR-Net achieved an accuracy of 98.9%, surpassing current benchmarks. Comparative analysis with state-ofthe-art machine learning and transformer-based models designed for tabular data confirms IAR-Net's performance and robustness in recognizing catheterization actions. Lastly, interpretability methods were incorporated to elucidate the model's decision-making process, improving understanding and increasing the trustworthiness of predictions. These outcomes offer a promising avenue for enhancing trainee assessment and training protocols, thereby accelerating broader acceptance and integration of robot-assisted endovascular systems into clinical practice.

Ähnliche Arbeiten