Author: Megan Rolfzen, MD
IARS and SOCCA 2026 Annual Meeting coverage
Artificial intelligence is moving from experimental research toward clinical applications in anesthesia and perioperative medicine. At the 2026 IARS and SOCCA Annual Meeting, experts discussed the data infrastructure, foundation models, predictive tools, and safeguards required to translate AI into routine patient care.
Meredith Adams, MD, MS, FASA, FAMI, emphasized that useful AI begins with standardized, high-quality data. Medical information is often stored differently across hospitals, making it difficult to combine datasets or develop models that work across institutions.
Her team validated a large language model-based system that automatically maps complex clinical variables into the Observational Medical Outcomes Partnership common data model. Standardization may support data sharing, multicenter research, collaborative biobanks, and the development of more generalizable foundation models.
Andrew H. Song, PhD, presented TITAN, a vision-language pathology foundation model trained using whole-slide pathology images and associated reports. The system can generate reports and retrieve similar slides to assist with difficult cases.
Broadly trained models such as TITAN may eventually support more individualized care by integrating pathology, molecular information, and other patient-specific biological features. However, using these tools to select anesthetic techniques or reduce cancer recurrence remains a potential future application rather than an established clinical practice.
Hannah Lonsdale, MBChB, addressed the gap between AI research and implementation in anesthesiology. Despite the specialty’s data-rich environment, anesthesiology ranks near the bottom of medical specialties in AI-related publications.
Many proposed technologies have not progressed beyond retrospective studies or experimental models. Before AI can be routinely incorporated into operating-room workflows, systems must demonstrate accuracy, clinical benefit, reliability, and generalizability across different institutions and patient populations.
Nima Aghaeepour, PhD, described machine-learning systems that combine electronic health record information with multi-omics data, including molecular and biological measurements. These models may identify risks that are not apparent from conventional clinical information alone.
One model identified patients undergoing relatively short procedures whose biological patterns predicted unexpectedly prolonged anesthetic courses. These patients also appeared to have a higher risk of postoperative complications.
Another AI model reportedly reproduced established drug-response patterns and outperformed practicing anesthesiologists in predicting intraoperative mean arterial pressure across institutions participating in the Multicenter Perioperative Outcomes Group.
Large language models still have difficulty reliably interpreting complex quantitative information. Additional development will be needed before they can safely generate individualized anesthetic plans from physiological, molecular, procedural, and electronic health record data.
The session also explored emerging technologies such as ambient documentation and digital twins. Ambient systems may automatically generate clinical notes from conversations and workflow data. Digital twins attempt to create detailed virtual representations of individual patients that can be used to simulate disease progression, treatment responses, or perioperative risk.
Key Takeaways
Clinical AI depends on standardized and representative data. Models trained on inconsistent or incomplete datasets may produce unreliable recommendations.
Foundation models may eventually integrate images, reports, physiological signals, electronic records, and biological data to support individualized perioperative decisions.
Predictive systems could assist with blood pressure management, complication forecasting, anesthetic planning, and postoperative recovery strategies. However, performance in retrospective studies does not guarantee improved patient outcomes in real clinical environments.
Anesthesiologists may increasingly become “human-in-the-loop” supervisors who oversee automated systems performing routine monitoring, documentation, and medication titration. Clinicians must remain responsible for interpreting recommendations and intervening when technology fails or encounters an unfamiliar situation.
Before widespread adoption, AI systems must be evaluated for:
- Clinical accuracy and patient benefit
- Bias and equitable performance
- Privacy and data security
- Transparency and accountability
- Reliability across different hospitals
- Environmental and computational costs
- Appropriate human oversight
Artificial intelligence may substantially change perioperative medicine, but its value will depend on whether it improves meaningful patient outcomes rather than simply adding sophisticated technology to existing workflows.
Thank you to IARS and SOCCA for allowing us to summarize this important coverage from the 2026 Annual Meeting.