Accuracy and Reliability of Artificial Intelligence in Surgical Decision-Making: A Literature Review

This narrative literature review synthesized evidence to address gaps in knowledge regarding AI performance and its integration into surgical operations. The purpose of the review was to assess AI accuracy and reliability, benchmark real-time guidance technologies, identify data and ethical issues, compare model performance across different specialties, and review the role of AI in improving surgical accuracy and safety. It reviewed 28 studies conducted across various geographic and disciplinary contexts and discussed machine learning (ML) and deep learning (DL) as applied to major surgeries. Results show that AI models' overall performance is substantial in intraoperative (IOP) decision-making, with five of six studies reporting AUC values of 0.85-0.95, indicating significant discriminatory power. Moreover, the accuracy performance metric across 22 studies showed high predictive performance of AI models in surgical settings, with accuracies ranging from 80% to 99%, except for one study, which reported an accuracy below 70%. These findings emphasized the practical feasibility of AI in IOP decision-making. Hence, AI's role in IOP is promising, assisting surgeons' decision-making in the operating room. Therefore, ML and DL are highly precise in anatomic detection, surgical-phase detection, complication prediction, and real-time event detection. Developments in DL algorithms, such as convolutional neural networks and generative adversarial networks, have enabled more accurate surgical guidance and the prediction of IOP events, thereby increasing surgical accuracy and potentially reducing errors. However, the model's performance needs to be validated through long-term computational and real-time clinical study designs, ensuring appropriate strategies for data validation and model performance assessment. The narrative review study design focused solely on the narrative synthesis, rather than on data validation (internal or external) or quality assessment of the included studies. Therefore, future researchers should conduct a systematic review to validate the findings. The readers must be cautious when interpreting the findings. Hence, AI use in surgery training and workflow optimization has the potential to improve surgical performance and patient outcomes, but scalability and long-term outcomes have yet to be demonstrated. Although AI technologies can improve the accuracy and reliability of decisions made in IOP settings, it is critical to address methodological, infrastructural, and ethical constraints to enable safe and effective clinical application in major surgeries.