Surgeons Provide Clarity on Applications for Generative AI in Patient Care

Artificial intelligence (AI) has evolved into a technology capable of prognostication, communication, and decision-making at a level previously considered uniquely human.

The transformative potential of AI is poised to reshape numerous fields, including healthcare and, specifically, surgical care.

During the 2024 Clinical Congress Panel Session, “Generative AI Tools for Surgery: Will AI Change My Practice?,” Genevieve Melton-Meaux, MD, PhD, FACMI, FACS, professor of surgery at the University of Minnesota in Minneapolis, proclaimed that “AI is ushering in a new industrial revolution.” 

The evolution of AI in healthcare can be broadly categorized into three stages: standardization, automation, and adaptation.¹ The first epoch focused on creating standardized prediction scores, which laid the groundwork for data-driven insights. The second stage introduced automation, such as data summarization and report generation, streamlining routine tasks.

We are now entering the third period of adaptation, where AI tools collaborate with and augment human capabilities, which has fostered in a new era of human-AI partnership in medicine. Generative AI—a rapidly growing field in which large, complex models trained on massive amounts of data are able to generate new content and perform new tasks they haven’t been explicitly trained to do—will likely play a large role in enabling this new era.

AI in Surgical Practice Today

While generative AI tools specifically designed and approved for clinical use are still emerging, the precedent for using this technology to enhance patient outcomes is well-established. Tyler Loftus, MD, PhD, FACS, associate professor of surgery at the University of Florida in Gainesville, pointed out that several conventional AI-powered prediction tools already have been tested in clinical workflows, demonstrating tangible improvements in care. These tools often focus on specific tasks (i.e., predicting the likelihood of a specific outcome) and showcase the power of AI in targeted applications.

One notable example is a widely used commercial computer vision tool developed by a large medical technology company for real-time colonic polyp detection during colonoscopy. By highlighting potential polyps in real time, this tool assists endoscopists in identifying and removing precancerous lesions, ultimately increasing quality-adjusted life years for patients undergoing colonoscopies.^2,3 This tool exemplifies how AI can augment human capabilities in real time, improving diagnostic accuracy and patient outcomes.

Another compelling example comes from the Hypotension Prediction trial (also known as the HYPE trial)—a large study that demonstrated how an AI tool capable of predicting intraoperative hypotension can actually decrease intraoperative hypotensive events.⁴ By proactively identifying patients at risk of experiencing low blood pressure during surgery, clinicians may be able to intervene earlier, possibly mitigating complications after surgery and potentially improving overall surgical safety. These examples demonstrate the value of AI in enhancing clinical decision-making and improving patient care.

Generative AI Use Inside and Outside the OR

Generative AI represents a significant leap forward in capabilities compared to the traditional, yet already impactful, AI tools. Unlike traditional AI, which is trained for specific tasks, generative AI possesses greater flexibility, adapting to diverse inputs and performing tasks for which it hasn’t explicitly been trained. This adaptability opens up exciting possibilities for its application in surgery.

A crucial consideration for implementing generative AI in surgery is identifying areas where its potential impact is high while the associated risk is low. Achieving this balance means focusing on applications where a potential AI error would not directly lead to serious patient morbidity or mortality.

Several promising areas of research and early development highlight the potential of generative AI in surgery, including surgical registry curation, intraoperative guidance, and operative video analysis. Gabriel Brat, MD, MPH, FACS, an associate professor of surgery at Beth Israel Deaconess Medical Center in Boston, Massachusetts, and instructor of biomedical informatics at Harvard Medical School in Boston, highlighted some of these potential future applications.

One significant application lies in automating large-scale clinical data extraction. Large language models can automate the laborious process of extracting, curating, and harmonizing clinical data from electronic health records for inclusion in national registries like the ACS National Surgical Quality Improvement Program. This automation can significantly reduce the time and resources required for data collection, potentially accelerating research and quality improvement efforts. The feasibility of this concept has been demonstrated in vascular surgery.⁵

Another exciting area is leveraging generative AI to understand intraoperative events through operative video analysis. This functionality has numerous applications, including surgical education (providing coaching and feedback for surgeons and trainees); quality improvement (identifying critical steps during procedures, such as achieving the critical view of safety during cholecystectomy); and administrative tasks (automatically generating written operative reports from video footage).

Finally, generative AI holds potential for real-time intraoperative decision support through image-guided recommendations. In the near future, generative AI tools may be able to overlay a patient's preoperative imaging (such as their computed tomography scan) in real time during a surgery, providing surgeons with a dynamic understanding of the patient's anatomy and facilitating more precise surgical planning.

Barriers to Generative AI Implementation in Surgery

Despite the immense potential of generative AI in surgery, significant barriers to its widespread adoption remain, according to Daniel A. Hashimoto, MD, MTR, FACS, an assistant professor of surgery and computer and information science at the University of Pennsylvania in Philadelphia and director of the Penn Computer Assisted Surgery and Outcomes Laboratory.

One fundamental challenge is establishing robust methods for measuring the performance of these complex tools.⁶ There is no universally accepted standard, and each metric has its limitations.

For instance, evaluating a computer vision tool designed to detect polyps during colonoscopy requires distinguishing between pixel-level accuracy (correctly identifying individual pixels) and lesion-level accuracy (correctly identifying entire polyps). A tool might achieve high pixel-level accuracy while missing critical lesions, highlighting the importance of choosing appropriate performance metrics. Another challenge in performance measurement is that aggregated metrics can mask poor performance in specific cases or among marginalized populations. It is crucial to ensure these tools perform equitably across all patient demographics.

Regulating generative AI tools presents unique challenges due to their non-deterministic nature, as the same input can sometimes produce different outputs, making it difficult to guarantee consistent safety and effectiveness. Regulatory bodies, such as the US Food and Drug Administration (FDA), are still grappling with how to effectively oversee these rapidly evolving technologies, particularly regarding post-market surveillance.⁷

While there is tremendous excitement around using large language models in daily clinical tasks, they have largely been studied in the context of medical examination question-answering; not much evaluation of these tools with real-world data sets has been performed.⁸ More broadly, only about half of all FDA-approved AI applications in healthcare undergo clinical validation prior to approval. Surgeons will have an important role to play in the evaluation and validation of new AI tools as they become available.⁹

Caroline Park, MD, MPH, FACS, an associate professor of surgery at The University of Texas Southwestern Medical Center in Dallas, suggested that effective implementation is critical for leveraging this technology in healthcare. These tools must integrate seamlessly into existing clinical workflows without adding complexity or inefficiency. More importantly, they must be designed to enhance, not hinder, physician decision-making. Poorly designed AI tools, regardless of their accuracy, can have little to no positive impact or even negatively impact clinical decisions.

Public perception and trust are crucial for the successful integration of generative AI in healthcare. A recent study revealed that while approximately 30% of patients express distrust toward health information produced by generative AI itself, a majority (64%) trust their doctors to use it responsibly to improve their care.¹⁰ This highlights the vital role surgeons will undoubtedly play in evaluating and guiding the adoption and implementation of these powerful tools.

Patients trust their physicians to leverage these technologies safely and effectively, underscoring the responsibility of the medical community to ensure that these advancements are used ethically and in the best interests of patient care.

Dr. Jayson Marwaha is a general surgery resident at Georgetown University in Washington, DC, and an incoming minimally invasive surgery fellow at the University of Michigan in Ann Arbor.

References

1. Howell MD, Corrado GS, DeSalvo KB. Three epochs of artificial intelligence in health care. JAMA. 2024;331(3):242-244.
2. Barkun AN, von Renteln D, Sadri H. Cost-effectiveness of artificial intelligence-aided colonoscopy for adenoma detection in colon cancer screening. J Can Assoc Gastroenterol. 2023;6(3):97-105.
3. Hassan C, Povero M, Pradelli L, Spadaccini M, Repici A. Cost-utility analysis of real-time artificial intelligence-assisted colonoscopy in Italy. Endosc Int Open. 2023;11(11):E1046-E1055.
4. Wijnberge M, Geerts BF, Hol L, et al. Effect of a machine learning–derived early warning system for intraoperative hypotension vs standard care on depth and duration of intraoperative hypotension during elective noncardiac surgery: The HYPE randomized clinical trial. JAMA. 2020;323(11):1052–1060.
5. Flanagan CP, Trang K, Nacario J, et al. Large language models can accurately populate Vascular Quality Initiative procedural databases using narrative operative reports. J Vasc Surg. Published online December 16, 2024.
6. Reinke A, Tizabi MD, Baumgartner M, Eisenmann M, et al. Understanding metric-related pitfalls in image analysis validation. Nat Methods. 2024;21(2):182-194.
7. Warraich HJ, Tazbaz T, Califf RM. FDA perspective on the regulation of artificial intelligence in health care and biomedicine. JAMA. 2025;333(3):241-247.
8. Bedi S, Liu Y, Orr-Ewing L, Dash D, et al.  Testing and evaluation of health care applications of large language models: A systematic review. JAMA. 2025;333(4):319-328.
9. Chouffani El Fassi S, Abdullah A, Fang Y, Natarajan S, et al. Not all AI health tools with regulatory authorization are clinically validated. Nat Med. 2024;30(10):2718-2720.
10. Deloitte Cener for Health Solutions. Building and maintaining health care consumers’ trust in generative AI. Deloitte Insights. June 6, 2024. Available at: https://www2.deloitte.com/us/en/insights/industry/health-care/consumer-trust-in-health-care-generative-ai.html. Accessed March 3, 2025.