From Postoperative Delirium to Language Models as Cognitive Aids: A Rapid-Fire Showcase that Spans the Research Spectrum

Connor Brenna, MD

A series of early-career investigators outlined their exciting contributions to academic anesthesiology on Saturday, March 22 at the 2025 Annual Meeting, presented by IARS and SOCCA. E. Railey White, MD, PhD, an assistant professor in Anesthesiology and Critical Care at the University of Pennsylvania, moderated a fast-paced spotlight on important research in anesthesiology during the “Early-Stage Anesthesiology Scholars (eSAS) Rapid Fire Showcase,” session during the Scholars’ Day.

Karam Atli, MD, chief resident physician at Washington University in St. Louis, opened the session with “Intraoperative Electroencephalogram Suppression and PACU Delirium in Patients Participating in the THRIVE Clinical Trial.” Postoperative delirium is an increasingly recognized source of cost and morbidity, including longer-lasting forms of cognitive dysfunction. However, relatively little work has characterized postoperative delirium at very early timepoints (e.g., in the postanesthesia care unit or PACU), where routine screening is seldom performed and the diagnosis is difficult to distinguish from lingering anesthetic effects. As a substudy of the ongoing Trajectories of Recovery after Intravenous propofol vs. inhaled VolatilE anesthesia (THRIVE) Trial, Dr. Atli will examine the relationship between suppression rate on a four-channel frontal electroencephalogram and delirium in the PACU (measured using the 3D-CAM tool) to understand whether the duration of intraoperative suppression associates with delirium during the first hour in the PACU, and whether anesthetic techniques have unique effects on this relationship.

In the second presentation, Richard Boyer, MD, PhD, an assistant professor of Anesthesiology, a Fun-Sun Frank and Baw-Chyr Peggy Yao Research Scholar in Anesthesiology and Director of Medical Accelerator and Digital Engineering (MADE) Lab at Weill Cornell Medical College, presented, “Digital and Wearable Perioperative Risk Stratification for Older Adults.” Traditional risk scores (e.g., American Society of Anesthesiologists classification, Revised Cardiac Risk Index, or metabolic equivalents) lack granularity, particularly for older adults, and failures of risk stratification can result in worse surgical outcomes, a higher rate of unplanned admission, and less effective shared decision-making. Dr. Boyer described a prospective observational cohort study of older adults undergoing major surgery, who are provided a smartwatch to wear for at least 7 days before, and 30 days after surgery. Data on parameters like heart rate, activity, and sleep are recorded throughout this period, and used to calculate advanced metrics like heart rate variability and estimated V̇O_2max. Participants also completed two gold-standard tests — a six-minute walk test on the day of surgery, and a Duke Activity Status Index (DASI) 30 days later — for comparison. Sixty-five patients with a mean age of 70 have completed the trial, and their data was modelled using interpretable machine-learning. Dr. Boyer’s team found that data easily measured from a wearable device, particularly heart rate recovery, baseline activity, and REM sleep, can accurately predict performance on the six-minute walk test and DASI, offering a potential alternative to traditional risk assessment.

Next, Emily Mackay, DO, MS, an assistant professor of Anesthesiology and Critical Care at the University of Pennsylvania, shared, “Targeted Intraoperative Echocardiography During Isolated CABG Surgery.” There is enormous practice pattern variation in transthoracic echocardiography (TTE) for isolated coronary artery bypass surgery (CABG): national statistics suggest that the use of TTE in these procedures could almost be predicted by a coin flip. That variation is likely borne out of risks associated with TTE, as well as limited evidence for the investigation in procedures which do not involve the cardiac valves. Consequently, practice guidelines have given TTE a Class IIb recommendation (“unknown usefulness”) for isolated CABG. Dr. Mackay’s work uses instrumental variable analysis to look more closely at this evidence, finding that TTE does have a survival benefit for higher-risk patients (e.g., those with congestive heart failure) — but possibly also an association with postoperative acute kidney injury. To better understand this relationship, Dr. Mackay’s team plans to undertake a randomized clinical trial studying TTE for isolated CABG.

The fourth talk, “Is Post-Intensive Care Syndrome a Nociplastic Syndrome?,” was presented by Lauriane Guichard, MD, an assistant professor of Anesthesiology at the University of North Carolina at Chapel Hill. Nociplastic pain, which arises from altered nociception in the absence of obvious tissue damage, is an independent predictor of opioid nonresponsiveness and worse postoperative outcomes, including pain chronification. It can also follow physiological stress or infection, and thus is itself a common consequence of critical illness. Using cohort data from the Michigan Genomics Initiative, Dr. Guichard’s team mailed questionnaires to patients who had previously undergone cardiac surgery followed by admission to the ICU (on average, 3.5 years earlier). Fifty-three patients mailed back completed questionnaires, sharing information relating to pain, mood, cognition, sleep, and postoperative satisfaction. Alarmingly, 15% of the study population reported new, widespread pain, which was associated with anxiety, cognitive dysfunction, poor sleep, and somatic symptoms. Next, Dr. Guichard plans to undertake a prospective observational cohort study of patients undergoing cardiac surgery with planned ICU admission; participants will complete preoperative questionnaires and bloodwork, and will be interviewed monthly for six months to evaluate nociplastic pain symptoms, as well as mood, cognition, sleep, and opioid use. This work may support the notion that nociplastic pain is a common element of “postintensive care syndrome,” the group of physical and psychological symptoms that as many as 80% of patients experience after receiving care for critical illness.

Finally, Nathan Hurley, MD, a resident in the department of anesthesiology and computer scientist at the University of Wisconsin-Madison, presented, “Language Models as Perioperative Cognitive Aids.” Dr. Hurley provided an overview of the recent explosion in Large Language Models (LLMs), which use deep learning to generate language. LLMs rely on statistics to evaluate a sequence of “tokens,” which can be thought of as similar to syllables (each one approximately four characters in English text), predict which token is most likely to come next in the sequence, and repeat. This “next-token prediction” can approximate thinking, but with the disadvantage that a strict reliance on statistical prediction can cause them to confabulate information. Faced with dozens of popular LLMs, most developed within the last five years, which ones can be trusted? To answer this question, Dr. Hurley’s team adapted clinical scenarios from the Stanford Emergency Manual, and prompted various LLMs to work through them. Experts assessed the answers and found that many models performed well, but newer and more complex models tended to perform the best. Dr. Hurley reported that models developed by OpenAI were among the top scorers overall, while evaluations limited to only open-source models (which tend to be more trustworthy) found that the Mistral LLM was best-suited to solving the team’s clinical problems.