Beyond the blank page: Frequentist and Bayesian perspectives on risk prediction algorithms

Abstract

Risk prediction has long been a cornerstone of surgical oncology, enabling surgeons to anticipate complications, tailor perioperative care, and improve outcomes. With the rise of artificial intelligence, machine learning (ML) models are increasingly being applied to predict outcomes, highlighting the growing significance of data-driven methods for clinical decision-making. Currently, frequentist approaches dominate prediction models, including most ML algorithms; these rely exclusively on observed datasets and risk overlooking the cumulative value of prior clinical knowledge. In contrast, Bayesian reasoning formally integrates existing evidence with new data. In this letter, we examine the strengths of frequentist-based prediction models, discuss how Bayesian methods may improve predictive accuracy, and argue that combining both approaches offers a promising path toward more robust, interpretable, and clinically useful prediction tools in surgery. This integration can yield robust, interpretable, and clinically relevant tools that advance personalized surgical care.

Keywords: Gastric cancer; Bayes theorem; Artificial intelligence; Probability learning; Prediction algorithms; Risk

Core Tip: Risk prediction in surgical oncology is evolving beyond traditional scoring systems. Frequentist methods, widely used in machine learning, offer transparency and reproducibility but rely solely on observed data. Bayesian reasoning, by contrast, integrates prior clinical knowledge with new information, mirroring real-world decision-making. The integration of both frameworks promises more reliable, interpretable, and personalized prediction tools, advancing the future of surgical care.