O114: Enhancing palliative oesophageal cancer treatment pathway prediction and prognostication through machine learning models

Abstract Introduction Non-curative treatment plans comprise the majority of Oesophageal Cancer (OC) decision-making by Multidisciplinary Teams (MDTs), operating under significant caseload pressures. Machine Learning (ML) can semi-automate, streamline, and standardise these decisions to predict treatments and prognosticate for new patients. We present ML models which predict treatment decisions and prognostication in palliative OC patients. Methods Using clinicopathological data from 437 palliative OC cases treated at a single tertiary centre over 12 years, we trained several ML algorithms (Multinomial Logistic Regression [MLR], Random Forests [RF], Extreme Gradient Boost [XGB], Decision Tree [DT], and Random Survival Forests [RSF]) to predict treatment pathways (best supportive care, chemotherapy, radiotherapy, palliative stent, or stent plus oncological adjunct) and survival prognoses. Model performance was evaluated using Area Under the Curve (AUC) for classifier models and C-index (1-concordance) along with calibration plots for the survival model. Additionally, the models were integrated into a web application for clinician-friendly usability. Results Mean AUCs for the classifier models were: MLR 0.801±0.090, RF 0.806±0.078, XGB 0.817±0.079 and DT 0.762±0.108. Mean C-Index for the RSF survival model was 0.330±0.018 while calibration curves showed good calibration within the first 6-12months (median survival for the cohort was 6.31 months (range 0.1-105.8 months)). Conclusion This study is the first use of ML to predict palliative treatment plans and offer prognostication for OC patients, seamlessly integrated into a user-friendly interface for real-time treatment pathway prediction. This offers significant potential to streamline MDT caseload and provide data-driven decision-support for clinicians counselling these patients.