P1.132. Assessing the Impact of Radiomics Features on Machine Learning Models for Predicting Multidisciplinary Treatment Decisions in Oesophageal Cancer

Abstract

Abstract Topic Esophageal Cancer: Other Background Medical Artificial Intelligence (MAI) is expanding within surgical oncology with new tools emerging for assisting cancer multidisciplinary teams (MDTs) in cancer types such as oesophageal cancer (OC). Radiomics has similarly gained popularity with proponents arguing that it provides comparable performance to human agents when assessing medical imaging however it’s value within OC remains unclear. This study evaluated machine learning (ML) models which combine radiomic and clinicopathological features even with missing or imbalanced classes for predicting treatment plans in curative OC patients to assess the added value of radiomic features over current clinical models in prediction quality and certainty. Methods Curative OC patients who underwent surgery with or without neoadjuvant therapy (chemotherapy or chemoradiotherapy) between 2010-2023 at a tertiary centre were identified from a prospectively maintained database and used to train extreme Gradient Boost (XGB) models. Clinicopathological data and computer tomography imaging was available as labelled or non-labelled data and stratified into training and validation sets for model training using 5x cross validation. Model performance was assessed on Area Under Curve (AUC) and degree of prediction uncertainty using Shannon Entropy. Results A total of 520 curative OC patients were identified. Clinicopathological data was available for all 520 patients while computer tomography imaging was available for 144 patients. Clinical model class-based AUCs ranged from 0.768-0.841 (SD range 0.026-0.037), while Clinico-Radiomic models only offered marginal increase in AUC (0.783-0.847, SD range: 0.018 – 0.044), Shannon Entropy for clinical models were of relatively uniform distribution whereas radiomic features peaked at high entropy. Conclusion Overall, the results suggest that incorporating radiomics features only marginally enhances predictive performance across treatment categories. Shannon entropy analysis additionally suggests the high dimensionality of radiomic features may potentially introduce noise into models making predictions more uncertain for a subset of patients. The gains in performance are consequently not sustainable enough, given the clinical workload and computation time required to extract radiomic features to advocate for introducing radiomic models for OC treatment allocation at present.