Machine Learning Applications for Objectively Assessing Surgical Skill and Instrument Dynamics

Abstract

The goals set forth in this dissertation are centered on advancing the current state-of-the-art literature for applying analytical approaches to determining surgical proficiency, testing new metrics for surgical movement efficacy, and understanding surgical instrument dynamics. Each of these three core research areas are focused on using modern, data-driven approaches to providing better patient care and improving the efficacy of surgical training programs. The amount of under-utilized data collected during surgical training and in live cases is astounding as simulation labs and operating rooms are equipped with countless types of sensors for monitoring patient data and spatio-temporal operating room dynamics, including surgical instrument motion.Initially, a series of regression models are trained and tested using a feature vector derived from the kinematics of two laparoscopic instrument while participants complete the Fundamentals of Laparoscopic Surgery peg transfer and suture with intracorporeal knot assessments. Models were trained to predict objective (time and error) and subjective scores. A recursive information gain feature selection method was applied for choosing the number of features to represent in each model. Model performances are compared and a discussion on the use of instrument kinematics as a measure of surgical competency is given.A second machine learning study is presented comparing deep neural network classification accuracies for determining performance at the peg transfer assessment. Two neural network architectures were tested: vanilla long short-term memory and convolutional neural network long short-term memory. The inputs to these models were the raw motion tracking points and the down-sampled video frames, respectively. Comparisons between the two neural network architectures were made and overall performances are discussed in the context of video-based surgical skill evaluation and conventional, objective metrics for assessing surgical proficiency.Instrument movement smoothness has served as a heuristic for suggesting surgical competency in several prior studies; however, the conventional metrics that have been used to quantify movement smoothness are susceptible to several pitfalls when applied to real-world data. Such metrics also consider movement patterns at a single time-scale. A new method for quantifying movement smoothness is presented using multiscale entropy of the velocity time series to calculate the complexity index over varying downsampling intervals. Tests using this metric are conducted on simulated point-to-point movement profiles with and without random noise and is validated using laparoscopic instrument tip motion data from a point-to-point movement experiment.Finally, a fundamental analysis of the dynamic response of a laparoscopic instrument under the presence of vibrotactile excitation and variable gripping pressures is presented. Conventional laparoscopic instruments lack the ability to transmit haptic feedback to surgeons’ hands, but future instrumentation could have this capability. To this end, it is essential that the impact of interfacial damping and stiffness, induced by contact forces exerted at the hand-handle interface, are understood. The connection between each of these studies is grounded in studying methods to improve inefficient training methods for laparoscopic surgery and improving laparoscopic instrument capabilities. The long-term vision for the foundational work in this dissertation is to develop a closed-loop system for intraoperative performance assessment and instrument feedback to recommend movements to surgeons in an effort to mitigate risk to patients and expedite learning for residents.