Accuracy of artificial intelligence for staging and decision-making in esophageal and gastric cancer

Abstract

Background: Artificial intelligence (AI) has demonstrated increasing potential in oncology, including in imaging analysis and as a tool for supporting clinical decision-making.
Objective: The aim of this study was to evaluate an AI conversational model, trained using up-to-date clinical guidelines, to assist with staging and initial treatment planning in patients with gastroesophageal cancer.
Materials and methods: We conducted a retrospective study of patients with esophageal, gastric, and gastroesophageal junction cancer treated at a tertiary care hospital between December 2023 and May 2025. A virtual clinical assistant or chatbot was developed using ChatGPT-4.5® (OpenAI), configured to interpret test results in accordance with the NCCN and ESMO guidelines. The chatbot was designed to determine the clinical stage (cTNM), suggest a treatment plan, and identify relevant findings. The answers were compared with the decisions made by the treating physician.
Results: Of the 53 patients evaluated, 34 met the inclusion criteria. The accuracy between the cTNM staging determined by the medical team and that of the chatbot was 94%. The concordance in the suggested therapeutic approach was 85%. The discrepancies were due to differences in the interpretation of the stage, clinical considerations not available to the AI, and social or logistical factors that influenced the medical decision.
Conclusion: AI, implemented through a chatbot trained using clinical guidelines, proved useful in supporting staging and initial treatment planning for gastroesophageal cancer, with high concordance with clinical decisions.