518. ARTIFICIAL INTELLIGENCE MODELS VS CLINICIAN PERSPECTIVES OF IMPORTANT FACTORS IN OESOPHAGEAL CANCER DECISION-MAKING

Abstract Background Decision-making for oesophageal cancer (OC) management remains multifaceted and nuanced in the face of a traditionally complex patient demographic. The evolving movement of artificial intelligence (AI) offers the potential of an imminent paradigm shift in traditional multidisciplinary team (MDT) workflow with Machine Learning models proving capable of encapsulating much of that decision-making framework. Clinician buy-in however stems from trust that such models accurately represent the perceptions and considerations of the MDT in its current form. Methods A mixed methodology study combining a national survey of United Kingdom OC MDT clinicians (sent to the British Society of Gastroenterologists, Association of Upper Gastrointestinal Surgery and UK& Ireland Oesophagogastric Cancer group) was conducted and compared with a Machine Learning (ML) based Random Forests model of treatment-decisions from a leading UK tertiary referral centre. The factors influencing OC treatment decisions from a human perspective were contrasted with variable importance from the AI approach. Human perceptions of barriers and attractions of AI-driven support tools were also explored to understand how human-AI collaboration could be facilitated in future. Results The National Survey which generated 67 responses found age and gender to play a more significant role in the ML model than they did in the clinicians' conscious decision-making process. Additionally, thematic analysis identified a variety of important factors, including molecular tests, symptoms, social circumstances, and nutritional status. The prospect of utilising AI-based decision support in the future received generally positive feedback, although opinions varied widely. Challenges to its adoption include perceptions of clinician superiority, the uniqueness of patient cases, the need for transparency and safeguards in the model, the requirements for data input, and the demand for proven effectiveness. Conclusions Comparison with the model highlighted demographic factors which appear to heavily influence treatment but were consciously deemed less important by clinicians. While overall, the response was positive towards an ML tool, some respondents felt very unwilling to use one. This could limit its adoption by MDTs unless decision-tools are developed to address such concerns around the barriers identified.