Automation of Surgical Workflow Recognition: Unveiling the Surgical Instrument Kinematics that Underly Robot‐Assisted Prostatectomy Procedures

Abstract

Robot‐assisted laparoscopic prostatectomy (RALP) is becoming the standard of care in prostate cancer surgery. Despite residing in the era of digital surgery, routine procedural evaluations still rely on manual video‐analysis. We, therefore, aim to expand the set of kinematic instrument metrics previously reported and use this multivariate data input to automatically separate the 10 steps in the workflow of RALP procedures. Using a mechanical instrument movement recorder, we digitize the x, y, and z trajectories of the tips of all 4 robotic arms and synchronize them with the endoscopic video feed for 2 standardized RALP procedures (with NeuroSAFE) and 2 accompanied by an extended pelvic lymph node dissection. We use trajectory derived kinematic metrics as an input for automated step separation. Extraction of instrument activities and kinematics successfully drove cluster formation at unsupervised t‐SNE, yielding silhouette separations scores between steps as high as 0.68. When these inputs are combined with manual scoring in supervised partial least squares discriminant analysis, we can isolate the main step discriminating metrics, namely intra‐instrument distances and position of the instruments within the patient. Combined instrument kinematics provide a direct reflection of the surgeon's highly complex patient interactions during different steps in the surgical workflow.