AIMI Grand Rounds: AI-Driven Imaging Queue Prioritization in the Emergency Department - David Kim, MD, PhD

The AIMI Grand Rounds, sponsored by the Center for Artificial Intelligence in Medicine and Imaging (AIMI), is a virtual series held on every fourth Tuesday. The series is a crucial initiative for disseminating the latest AI advancements in medicine, aiming to drive transformative innovations in healthcare. Stanford participants of the live event can claim 1.0 CME credits: AMA PRA Category 1 Credits ™ or Non-Physician Participation Credit.

Speaker:

David Kim, MD, PhD: Assistant Professor of Emergency Medicine, Stanford University 

Bio: I am an NIH-funded emergency physician and researcher specializing in the application of artificial intelligence and patient monitoring to improve emergency care delivery. My research program focuses on developing and validating multimodal AI systems that integrate physiologic signals with clinical context to predict patient trajectories and support decision-making.

Abstract: Delays in emergency department (ED) imaging can compromise patient safety, yet queues typically operate first-in, first-out (FIFO) regardless of clinical urgency. While AI has improved post-imaging interpretation speeds, the critical “pre-acquisition” waiting period remains unoptimized. We developed a discrete-event simulation of 131,027 CT studies, calibrated to real-world operations, to evaluate AI-driven queue prioritization. A LightGBM model predicted clinical actionability at the time of order (AUROC 0.766), allowing dynamic re-ordering of the queue. In a held-out evaluation (n=20,795), the AI policy reduced median wait times for actionable studies by 10.8 minutes, increasing the proportion of actionable findings obtained within one hour from 48.3% to 57.3%. Crucially, the system effectively targeted the dangerous “tail” of the distribution: 90th percentile wait times for actionable patients dropped by 43.3 minutes (196.2 to 152.9 minutes). This acceleration came with minimal impact on lower-acuity patients. Our approach captures 80-87% of the theoretical maximum benefit, suggesting that pre-acquisition AI triage can significantly reduce time-to-diagnosis for the sickest patients without requiring additional hardware.