3447 - Physicians Views on Explainable Artificial Intelligence Models to Predict the Risk of Toxicity Following Breast Radiotherapy

Y. Liang¹, J. Rainbird², G. Cortellessa³, M. Balia⁴, G. Bologna⁵, I. Koutsopoulos⁶, A. Pannison⁷, A. Perotti⁷, B. Ramaekers⁸, T. Rattay⁹, S. Rivera¹⁰, A. Romita¹¹, C. Roumen¹², C. Talbot¹³, K. Verhoeven¹, M. Bergeaud¹⁴, and F. Fracasso³; ¹Department of Radiation Oncology, Maastricht University Medical Centre - GROW (MAASTRO), Maastricht, Netherlands, ²Leicester Cancer Research Centre, University of Leicester, Leicester,, United Kingdom, ³Consiglio Nazionale delle Ricerche, Instituto di Scienze e Technologie della Cognizione (CNR-IST), Roma, Italy, ⁴TheraPanacea, Paris, France, ⁵University of Applied Sciences and Arts of Western Switzerland, HES-SO, Geneva, Switzerland, ⁶Athens University of Economics and Business, Athens, Greece, ⁷CENTAI Institute, Turin, Italy, ⁸Department of Clinical Epidemiology and Medical Technology Assessment, Maastricht University, Medical Center+, Maastricht, Netherlands, ⁹Leicester Cancer Research Centre, University of Leicester, Leicester, United Kingdom, ¹⁰Gustave Roussy, Cancer Campus, Villejuif, France, ¹¹Medical Data Works B.V., Maastricht, Netherlands, ¹²Department of Health Services Research, Care and Public Health Research Institute (CAPHRI), Faculty of Health Medicine and Life Sciences, Maastricht University, Maastricht, Netherlands, ¹³Department of Genetics and Genome Biology, University of Leicester, Leicester, United Kingdom, ¹⁴Unitrad, Unicancer, Paris, France

Purpose/Objective(s): In breast cancer (BC) treatments, arm lymphedema can represent a side effect with potential major impact on patient’s quality of life. The possibility to obtain a personalized prediction of the risk occurrence for lymphedema can inform a more tailored treatment plan and suggest lifestyle advice to minimize the side effect and its severity. In the ongoing Horizon Europe project PRE-ACT (Prediction of Radiotherapy side Effects using explainable AI for patient Communication and Treatment modification), explainable Artificial Intelligence (XAI) algorithms are developed to predict and interpret lymphedema risk and other side effects in BC patients who underwent surgery, with or without systemic treatment and loco-regional radiotherapy (RT). A dedicated communication tool (PRE-ACT tool) for shared decision-making (SDM) is being designed to support radiation oncologists (ROs) and will be evaluated in a multi-centre randomized clinical trial (PRE-ACT-01) for which clinicians’ needs and recommendations are crucial. Materials/

Methods: From May to July 2023, interviews with ROs were carried out to gather feedback from potential stakeholders of the PRE-ACT tool and incorporate their point of view in the design process. Participants were recruited in the Netherlands, the United Kingdom and France (5, 5 and 8 ROs respectively). Questions regarding the kind of information that is needed about patient’s personalized lymphedema risk, how the information should be given to them and when the information is needed, were asked. The interviews were audio-recorded, transcribed verbatim and analysed qualitatively following a thematic analysis.
Results: ROs considered the PRE-ACT tool primarily as a support during joint consultation with patients before the start of RT. A global overview including the lymphedema risk accompanied with oncological benefits of RT should be provided. Explanations should cover a description of risk factors, highlighting their cumulative effect, and the risk estimation over different time periods after RT. Information should be provided at different levels of complexity (e.g. through pictograms, graphs, statistics, etc.) to embrace inclusivity together with personalized suggestions to minimize the risk. General sections on lymphedema, radiotherapy and prevention should also be incorporated as additional material. Trust in the predicted risk should be fostered ensuring reliable validation of the XAI model. This model should be trained on a good-quality dataset and supported by literature.
Conclusion: ROs have provided clear recommendations on what kind of information is needed as well as on how and when to receive the information of XAI prediction models, highlighting the need of accessibility and trustworthiness. Focus groups with former BC patients were held in parallel with the interviews with the ROs. These recommendations have fed into the design of a digital communication package, which will be tested in a clinical trial.