Multi-modal Features Analysis and Performance Assessment for Endovascular Surgery Skills

Purpose: Interventional skill and assessment are essential for vascular interventionalists prior to performing clinical procedures. However, traditional surgical assessments often fail to deliver objective and accurate results. To address these issues, our study proposes a simulator and an analytical framework for evaluating endovascular procedural performance. Methods: A custom interventional operation simulator included an in vitro silicone phantom, mock circulation loop, visual module, force-sensing module, and custom software for kinematic and force data post-processing. An in vitro study was conducted to deliver medical guidewire to four target vessels by thirty interventionalists and ten medical novices. Four types of features (i.e., vascular morphological, device-vascular model interaction, kinematic and force) were obtained from the interventional tasks. An operational force prediction model was developed and trained using long short-term memory (LSTM) networks based on multi-modal feature datasets. Furthermore, features with significant differences were applied for qualitative and quantitative evaluation via support vector machine and Mahalanobis distance. Results: The well-trained LSTM model could predict the operation force, force and torque between instruments and vascular model. The qualitative assessment achieved an overall accuracy of 84.17% for two-class classification to distinguish the novice and expert attempts, and the quantitative assessment could assign effective scores for all attempts. Conclusions: This work provides an approach for the prediction and assessment of intervention performance behaviors, and results demonstrate the effectiveness of this method for promoting the development of interventional skill assessment.