By: Brielle Ochoa
On PubMed, the number of publications for “artificial intelligence” (AI) and “surgery” ten years ago was less than 100. In 2023, this number was up to nearly 3,500. But what basics should you know about AI in general in healthcare and surgery?
- AI is an umbrella term for a wide range of technologies, which share the goal of mimicking human intelligence and cognitive function and processing. AI includes large language models, robots, machine learning, deep learning, neural networks, and artificial general intelligence; while beyond the scope of this post, these terms are interrelated but distinct from one another.
- AI can be used in a variety of general healthcare and surgical applications, including algorithms for clinical decision making, as part of technology in medical devices, for pattern recognition in images and videos, and for preparation of written documents, such as research manuscripts and patient educational information.
- AI must be trained on a dataset. Key phases of AI model creation are that it must be developed based on a nonbiased, heterogeneous, and relevant data set for a focused clinical problem and then must be validated on a separate data set. The model should then be evaluated, adjusted, and/or re-created as needed prior to clinical use.
- AI is only as effective, safe, and relevant as it is to the dataset it was developed from. AI trained on data that is biased, false, or irrelevant for a task will yield inaccurate and potentially harmful outputs. This may sometimes lead to errors known as AI hallucinations.
- Not all aspects of AI are currently regulated by the U.S. Food and Drug Administration (FDA). AI related to medical devices falls under FDA authorization, and currently approved devices utilize machine learning, but none use generative AI or artificial general intelligence (as of October 2023). For clinical decision support tools, recent publications have emphasized the lack of transparency, standards, and evaluation of bias for these tools, which undermines their clinical utility.
The output of AI should be critically questioned in a similar manner to research findings and use of devices. A few recent issues regarding AI illustrate the importance of the above points:
- A risk-prediction algorithm used for patient care inappropriately underestimated risk for Black patients compared to White patients because the algorithm was developed based on health care costs instead of another measure of illness severity and health care needs.
- Several reporting guidelines have been proposed to evaluate AI development, analyses, and associated processes. A recent JAMA publication focuses on AI-related reporting guidelines, including for manuscript preparation and for AI used in research.
- A basic science publication included figures and labels that were AI-generated and nonsensical, likely due to poor training for the image generator driven by AI, and was retracted. This issue was not identified during the peer review or publication process.
The use of AI poses additional questions regarding its use in research and patient care: Should patients be informed of clinical decisions driven by AI, and if so, to what extent? Who is liable when AI is wrong or fails to identify a problem? How should surgeons be approaching AI and robotics? As AI technologies continue to develop, surgeons, as well as other healthcare providers, should be aware of how AI is being used in clinical care and research.
Brielle Ochoa attended the University of Delaware and the University of Virginia School of Medicine. She is a general surgery resident at the Medical University of South Carolina in Charleston, SC and a current pediatric surgery research fellow at Phoenix Children’s in Phoenix, AZ. Her interests are in research and clinical ethics.
As someone really interested in this topic, I really enjoyed reading this article. Hope to hear more from you about any advancements in healthcare tech specifically AI models or a review of certain interesting publications.