OGC O04 The influence of age in oesophageal cancer treatment decisions: a machine-learning approach

Abstract

Abstract Background Although patient age plays a crucial role in determining curative treatment options for Oesophageal cancer (OC), the exact extent of its influence is not well defined. We used an interpretable machine learning (ML) approach to model for the first time the exact impact of age in combination with other key decision-drivers, such as tumour and patient characteristics, on the predicted probability of receiving different types of curative treatment as determined by a single centre OC multidisciplinary team (MDT). Methods Retrospective analysis of 399 OC patients undergoing curative treatment between 2010-2020 at a tertiary unit. A random forests (RF) classifier model was trained to predict curative treatment decisions for OC patients (neoadjuvant chemotherapy (NACT) + surgery, neoadjuvant chemoradiotherapy (NACRT) + surgery or surgery alone). Variable importance (VI) and Partial Dependence (PD) analyses were used to map the importance of age in the model, its influence on the predicted probability of each treatment decision, and how that relationship was affected when interacting with other decision-driver co-variates. Results Age was the most important variable for the model (26% of total importance). PD analysis demonstrated that patients over 70 years had a substantially higher predicted base probability of receiving curative surgery and lower probability of NACT. Moreover, while the probability of receiving surgery and NACT was primarily driven by disease characteristics (T and N staging) in patients < 70 years, age became increasingly important in predicting a surgery-only decision in those >70 years (P< 0.001). The base probability of surgery and NACRT decisions was additionally influenced by performance status and age, but age alone for NACT patients. Conclusions We have successfully combined ML modelling with PD analysis in a novel approach within the OC space to delineate the relationship between age and OC treatment decisions. Age heavily influences curative decisions for OC patients but plays a greater role in patients with specific tumour characteristics. This study provides the basis of a ML approach for examining subconscious decision-drivers in the management of OC. This in turn allows OC MDTs to examine areas of potential health inequality which may result from variability within that decision-making framework.