Artificial Intelligence for Augmenting Perioperative Surgical Decision-Making—Are We There Yet?
Interest in the role of artificial intelligence (AI) in medicine is growing exponentially. Machine learning models have been trained to perform diagnostic and decision support functions, including interpretation of radiographic abnormalities, detection of polyps during endoscopy, and prediction of melanoma from images of skin lesions. In surgery, AI also has the potential to inform intraoperative decisions.
Another concern is explainability. While deep learning is very powerful for making inferences through pattern recognition, it is less effective at explaining how it arrived at a particular decision. This is important, because current applications of AI are for augmenting surgeons’ judgment, not replacing them. Would a surgeon prefer a highly accurate model that does not explain itself or a less accurate model that does explain itself?
Finally, there are practicality issues for applying these models into everyday practice. For example, the authors6 describe feeding the model representative slices. What is considered representative? How many frames per patient are required for an accurate prediction? This highlights a critical need for methodological standards for reporting on the performance, validation, and deployment of AI models that perform computer vision tasks.
These questions for translating AI into surgical care remain. However, this proof-of-concept study will undoubtedly stimulate surgeons, innovators, and data scientists to continue investigating the role of machine-learning methods to improve care.
