Privacy-preserving Algorithms Can Facilitate Surgical Video Databases for Quality Improvement, Education, and Research

Abstract

I n the early 1900s, surgeon Ernest Amory Codman resigned from Massachusetts General Hospital to found a private institution called the End Results Hospital. 1 He was motivated by the "end result idea," which he described as "the common-sense notion that every hospital should follow every patient it treats long enough to determine whether or not the treatment has been successful and then to inquire, 'if not, why not' with a view to preventing similar failures in the future." 2 Given the technology (or lack thereof) of his time, Codman was limited to recording elements such as a patient's diagnosis, comorbidities, treatment plans, complications, and short and long-term results in follow-up. He also cataloged his opinions on the nature of the cause of suboptimal results. 1,2 A century later, the "House of Surgery" has embraced quality improvement and datadriven decisions in the care we deliver to our patients. Cases are discussed in morbidity and mortality conferences, and technology has improved our ability to capture, store, and analyze data. However, we are still mainly limited to data that is manually extracted and reviewed (eg, research or quality improvement registries) or optimized for billing purposes (eg, International Classification of Diseases 10th Revision and Current Procedural Terminology codes). In surgery, much of these data are captured from perioperative phases of care, whereas data from the crucial intraoperative phase of care are limited to surgeon-dictated operative notes, which may be incomplete or even inaccurate. 3,4 Video has been demonstrated to be a better modality for capturing details about intraoperative events versus operative notes. 3,4 Still, the incorporation of video into quality improvement and research databases or clinical records has been limited. While video of surgery has been demonstrated to have value in assessing surgical performance and correlating with patient outcomes, 5,6 the minority of surgeons record video of their cases, with many citing hospital policies and medico-legal concerns as top reasons for not recording cases 7 as violations of patient or surgical team privacy may have significant consequences. Thus, preserving privacy in surgical videos is a crucial concern.De Backer et al 8 report their machine learning algorithms that leverage computer vision techniques to de-identify out-of-body frames in robotic surgical videos. Utilizing video from 6 different procedures on four unique surgical robotic systems across 6 different hospitals on 2 continents, the authors trained machine learning algorithms to detect entry and exit from robotic trocars to mark when the camera was entering or exiting a patient's body and enable classification of video frames as in or out-of-body. Out-of-body frames were blurred or redacted to prevent patient or surgical team identification. Their results demonstrated accuracy and negative predictive value of over 99% in robotic and laparoscopic videos. While there were instances of false positives and negatives, no frames containing pati...