Artificial Intelligence for Predicting Adverse Surgical Outcomes: Challenges, Limitations and Implications for Clinical Translation - A Narrative Review.

The rise in the number of surgeries per year has led to the development of many artificial intelligence models for predicting surgical complications. Despite their ever-growing use in healthcare, artificial intelligence is not up to the mark yet. We need to search and critically overcome the hurdles preventing their safe and reliable use in surgical care. This narrative review aims to find and analyze the main limitations and challenges of artificial intelligence in predicting surgical outcomes. Across the reviewed literature, key limitations were identified in four domains: data-related, methodological limitations, performance and generalizability, and barriers to clinical implementation. Common issues included missing and imbalanced datasets, small sample sizes, retrospective single-center designs, high risk of bias, and inadequate external validation. Although several studies reported high predictive performance, these findings were often derived from non-representative datasets and lacked prospective validation. Additional concerns included limited interpretability, ethical and privacy risks, workflow integration difficulties, and potential amplification of healthcare disparities. Despite their potential, AI models for surgical outcome prediction remain constrained by multiple challenges. Substantial improvements in data quality, transparency, fairness, and robust multicenter prospective validation are required before AI can be safely and reliably integrated into routine surgical decision-making.