2024 FSA Podium and Poster Abstracts

P006: APPLICATIONS OF CLOSED-LOOP SYSTEMS, MACHINE LEARNING, AND NEURAL NETWORKS IN ANESTHESIA
Hunter Barcena; St. George's University

Introduction: Historically, anesthesiologists have been early pioneers of closed-loop devices. As early as the 1950s, Bickford and others used electroencephalogram measurements to develop an automated delivery of volatile anesthetic.1 A thermostat is an example of a closed-loop device outside the medical field.

Closed-loop control devices are fully automated systems in which a sensor(s) provides feedback to an algorithm (controller) that determines the action to achieve a desired target. Algorithms underly closed-loop systems and are similarly foundational to machine learning (ML). Supervised machine learning involves the creation of an algorithm that is “trained” to predict an outcome. A neural network is an example of a supervised machine learning algorithm utilized in medicine.

Artificial intelligence applications in anesthesia generally strive to accomplish one of three goals. Analyzing extensive data sets to identify novel patterns, generating models or algorithms to predict an event (i.e., intraoperative hypotension), and leveraging complex data sets (i.e., medical images) over time to provide real-time data to help anesthesiologists make decisions and respond to changes in patient conditions.2

Methods: A literature review was performed using the PubMed database, utilizing keywords such as “machine learning,” “artificial intelligence,” “anesthesia,” and “neural networks.” A thorough review of the results was conducted.

Results: There are approximately 20 studies investigating the application of closed-loop systems in controlling anesthetic depth in adults during the perioperative period.2 Most of these studies utilized the bispectral index (BIS) as the target variable for anesthetic depth. A meta-analysis by Brogi and colleagues of 15 studies showed that automated systems increased the percentage of time the outcome variable (depth of anesthesia) was maintained in the desired range by 17.4%.3 A subset of these studies also examined the proportion of time that the controlled variable was above or below the targeted set point, with a meta-analysis showing 12.3% more undershooting or overshooting in the manual groups compared with the closed-loop groups.3

Shalbaf and colleagues applied EEG features within a neural network model to discriminate different states of anesthetic depth and demonstrated 93% accuracy compared with the BIS index’s 87% accuracy.4 These results show that although BIS is a commonly used modality for measuring the depth of anesthesia, better depth measures may be achievable.

Discussion: Recent studies such as the HYPE RCT demonstrate the potential for ML to predict intraoperative hypotension before it occurs.5 Further research with more extensive study populations in diverse settings is needed to assess the effect on patient outcomes, safety, and generalizability of future ML studies.

The progression of Artificial intelligence raises essential questions in medicine. Should we allow artificial intelligence to access personal health data? Where does the responsibility land for machine learning output? How can we protect against clinical skill atrophy?6 Regulations for machine learning in healthcare are emerging slower than the developing technology. An understanding of the landscape of machine learning is critical for anesthesiologists to advocate for patient safety, collaborate with data scientists and policymakers, and utilize new technologies to provide first-class patient care.