OGC P08 A machine-learning approach to predict Upper Gastrointestinal multidisciplinary team treatment decisions

Abstract

Abstract Background The complexity of the Upper Gastrointestinal (UGI) multidisciplinary team (MDT) is growing, leading to rising clinician workload, time pressures and demands. This increases heterogeneity or ‘noise’ within decision-making for patients with oesophageal cancer (OC) and may lead to inconsistent treatment decisions. In this context, machine learning (ML) approaches offer the potential to standardize, automate and produce more consistent, data-driven decisions. Such models benefit from incorporating information-rich, diverse datasets to increase predictive model accuracy. The aim of this study was to develop a model capable of predicting UGI MDT treatment decisions for OC patients (whether the patient should receive surgery alone (S), chemotherapy prior to surgery (C+S) or chemoradiotherapy prior to surgery (CRT+S)) using only variables available at the time of the first treatment decision. Methods We conducted a retrospective analysis of patients who underwent oesophageal cancer resections between 2010–2016. Twenty pre-operative clinical variables available to the MDT were used to develop a predictive machine learning model using a penalised Multinomial Logistic Regression (MLR) classifier with 10-fold cross-validation and 5 repeats. Results A total of 399 cases were identified with a male: female ratio of 3.6:1, and median age of 66.1yrs (range 32–83 yrs). The overall model accuracy was 61.4% (Kappa 0.399). A Receiver Operator Characteristic analysis produced an Area Under Curve of 87.7% for Surgery vs. CRT+S, 86.9% for Surgery vs C+S and 66.1% for C+S vs CRT+S when trialled on a validation set. Conclusions These results suggest even basic ML modelling techniques offer the potential to model and predict current UGI MDT treatment decisions when selecting patients for surgery or neoadjuvant therapy options. Such models may allow prioritization of caseload, improve efficiency, potentially reduce waiting times, and offer data-driven decisions where more complex cases pose challenges. While this study is the first step in such a process, future work will need to incorporate additional data modalities such as medical imaging and histopathology as well as expansion to other ML classifier algorithms to determine the best performing models in order to improve overall predictive accuracy.