ICU or no ICU? If that is the question, we may soon ask AI

December 17, 2019

A pilot study of 50 patients conducted at NYU Langone Medical Center and presented at the 2019 American College of Surgeons Clinical Congress tested the appropriateness of ICU triage by an artificial intelligence program compared to that of clinicians.  Using input from a number of sources and crunching large amounts of data, the computer created an algorithm with 87 clinical variables and 15 specific criteria related to the appropriateness of admission to the ICU within 48 hours of surgery.  For purposes of the study, criteria for appropriateness of post-operative ICU care were established. If any criterion was met, the ICU care was deemed appropriate, if not met, the admission was deemed an “over triage”.  Conversely, a patient who was not sent to the ICU but ultimately met a criterion was deemed an “under triage”.   

The clinicians and the computer algorithm were asked to prospectively assign a patient to receive ICU care or not.  Any over or under triage event was classified as failed. Appropriate triage determinations were made in 82% of cases by the AI program, while surgeons accurately assigned patients 70% of the time, followed by intensivists at 64% and anesthesiologists at 58%.  While the authors of the study concede that this study only represents a first step, they plan to continue to refine the program and to apply the concept to larger and more diverse patient populations.  

EHC NOTE: While the pilot study discussed was relatively small, the preliminary capability of an AI program to more accurately determine the need for ICU care is persuasive.  We assume that such programs can be further refined and improve their accuracy as they continue to crunch data, receive feedback and “learn”. Certainly, the ability to engage computer support technology across the perioperative continuum to improve the appropriateness, consistency and speed of decision making offers exciting possibilities to impact patient outcomes and perioperative efficiency.