The lab focuses on Surgical Informatics and everything that entails. Our projects revolve around using Machine Learning to improve outcomes of various surgical procedures. Some of our current projects are listed below.
Opioid Prescribing
Surgical patients often require opioid pain prescriptions after discharge from the hospital to control their pain during recovery, but healthcare providers may overestimate the number of pills the patient needs - contributing to pill diversion and opioid misuse. We are leveraging machine learning and causal inference techniques to develop accurate opioid prescribing guidelines for each type of major surgery, to predict which patients need customized prescribing, to increase adoption of guideline-informed prescribing by healthcare providers, and to statistically adapt guidelines and patient-oriented models developed at one institution to the patient and procedure mix at other institutions.
CritUC: Surgical Risk Predictor for Ulcerative Colitis
Effectively managing uncertainty is one of the most common and challenging aspects of medicine. Medical standards of care can provide guidance, but are often not readily apparent or clearly defined for many clinical decisions. In this work, we describe a new method for identifying the risk level of a patient as perceived by a large population of physicians to deliver guidance at the bedside. Through analysis of large datasets of physician behaviors, perceived risk can be inferred for clinical decisions with no gold standard. To demonstrate this, we used administrative records of 6.8 million physician-patient encounters to guide one such decision: surgical timing in patients with ulcerative colitis. Our model was able to identify 78% of emergency surgeries an average of 381 days in advance (95CI: 232-530 days). Patients flagged in this way were 7.2 times more likely to proceed to surgery and cost an additional $200,000 compared to unflagged patients over six months. These patients also consumed twice the quantity of corticosteroids and were half as likely to experience steroid-free remission. A set of case-based scenarios presented to clinicians found stronger agreement with model predictions than actual outcomes. Finally, the primary signal identified by the model was found to be degrees of physician concern. For clinical decisions where there is no gold standard, the group behavior of physicians acting in their individual capacity in the best interest of their patients, is likely to represent a valuable baseline from which a choice can be made. Our approach has the potential to illuminate many individual clinical decisions currently limited by uncertainty in a practical manner.
COVID-19
The severity of a COVID-19 infection can range from mild or symptomless disease burden on one end to major physiological impacts requiring admission to intensive care. By applying unsupervised item response theory to patient medical record data, we are creating a continuous, objective measure of disease severity, ranging from 0-100, that does not rely on human-labeled data. This granular score can be used for patient care & management, aggregate-level severity tracking, and evaluation of public health measures or other policies.
Design-Thinking and Surgical Informed Consent
We are developing a Visual Consent Tool prototype in collaboration with the Gehlenborg Lab. Our prototype builds on existing surgical risk calculators by providing patient-facing risk visualizations and incorporating patient preference. Over the past year, we’ve used principles of design-thinking to continually iterate and improve our prototype. We conducted user interviews with patients and surgeons to obtain feedback regarding how our tool might be used in practice to support preoperative consultations, informed consent, and shared decision-making. In the process, we have gained valuable insight into surgeons’ and patients’ priorities during informed consent, the role of risk calculation in informed consent, and the limitations of traditional risk calculators.
Surgical Vision
Our goal is to leverage the vast amount of perioperative and intraoperative data to improve surgical care, training and delivery -- an area of medicine that remains to be highly subjective and lacks innovation. By using several data streams such as the electronic medical record (EMR), intraoperative data and video, we are able to not only characterize surgical procedures and the underlying techniques, but also objectively evaluate what is traditionally considered the art of surgery. To do this in a automated and efficient manner, we have built a multi-task architecture neural network and an analytic pipeline that allows for the processing of spatiotemporal data (i.e., surgeon behavior tracking and procedure segmentation) from video recordings, and textual data (i.e., operative notes and clinical data) from the EMR. This project was inspired by a need for objective evaluation of surgeons and their techniques, and is the first platform to leverage some of the applications of AI in minimally invasive surgery (i.e., laparoscopic, robotic) but within a much more complex and a much wider domain that is all open, non-minimally invasive surgical procedures.
"Un-Codeable" Diagnosis Identification Project
Several medical conditions exist that do not have their own unique diagnosis code in widely-used clinical terminologies. This is a significant limitation of real-world clinical data: conditions that do not have a unique identifier can be difficult to pick out from structured observational data sources such as administrative claims databases. These “un-codeable” conditions, as a result, cannot be easily tracked for clinical purposes or studied for research purposes, despite having physician-defined characteristics. Acute severe ulcerative colitis (ASUC) is one such condition. Accurate, automated, large-scale identification “un-codeable” conditions like ASUC may enable research into these poorly-understood conditions. In this project, we use machine learning techniques to identify "hidden" conditions in administrative claims databases that do not have their own code.
Surgical Intuition Project
Researchers have long believed that the purpose of AI in surgery is to meaningfully augment the surgeon's capabilities, not replace them. However, few practical implementations of "human-in-the-loop" AI in surgery have been proposed. Merging a surgeon's intuition with machine learning methods represents an important potential opportunity to address this challenge while also improving the performance of surgical predictive models. In this project, our goals are to develop a way to measure surgeon intuition regarding individual patients and combine this measure of intuition with machine learning models in a way that improves the ability to predict future patient outcomes.