ESTRO 2025 - Abstract Book

S2348

Interdisciplinary – Other

ESTRO 2025

Kim, C., Lee, H., Seol, H., & Lee, C. (2011). Identifying core technologies based on technological cross-impacts: An association rule mining (ARM) and analytic network process (ANP) approach. Expert Systems with Applications, 38(10), 12559–12564. https://doi.org/10.1016/j.eswa.2011.04.042 Adel, M. E., & Harrison, C. (2024). Unravelling technology meta-landscapes: A patent analytics approach to assess trajectories and fragmentation. World Patent Information, 76, 102256. https://doi.org/10.1016/j.wpi.2023.102256

766

Digital Poster Impact of Missing Data on AI Fairness for Breast Cancer Radiotherapy: Insights from the PRE-ACT Project Francesco Cozzi 1,2 , André Panisson 1 , Alan Perotti 1 , Antonio Ferrara 1,3 , Guido Bologna 4 , Sofia Rivera 5 , Marie Bergeaud 6 , Catherine Gaudin 6 , Guillaume Auzac 5 , Thomas Sarrade 7 , Ines Vaz Luis 8 , Tim Rattay 9 , Alessio Romita 10 , Karolien Verhoeven 11 , Yuqin Liang 11 , Jordan Rainbird 12 , Maddalena Balia 13 , Bram Ramaekers 14,15 , Willem Witlox 14,15 , Cheryl Roumen 16 , Gabriella Cortellessa 17 , Francesca Fracasso 17 , Iordanis Koutsopoulos 18 , Chris Talbot 12 1 Responsible AI Team, CENTAI, Torino, Italy. 2 Department of Computer, Control and Management Engineering, Sapienza University of Rome, Rome, Italy. 3 Computer Science, Graz University of Technology, Graz, Austria. 4 University of Applied Sciences and Arts of Western Switzerland, HES-SO, Geneva, Switzerland. 5 Radiation Oncology Department, Gustave Roussy, Villejuif, France. 6 Research, UNICANCER, Paris, France. 7 Radiation Oncology Department, Tenon Hospital, Paris, France. 8 U1018, INSERM, Gustave Roussy, Villejuif, France. 9 Leicester Cancer Research Centre, University of Leicester, Leicester, United Kingdom. 10 Research, Medical Data Works B.V., Maastricht, Netherlands. 11 Department of Radiation Oncology, GROW School for Oncology and Reproduction, Maastricht University Medical Centre+, Maastricht, Netherlands. 12 Department of Genetics, Genomics & Cancer Sciences, University of Leicester, Leicester, United Kingdom. 13 Research, Therapanacea, Paris, France. 14 Department of Clinical Epidemiology and Medical Technology Assessment, Maastricht University, Maastricht, Netherlands. 15 School for Public Health and Primary Care, Care and Public Health Research Institute, Maastricht University, Maastricht, Netherlands. 16 Department of Health Services Research, Care and Public Health Research Institute CAPHRI, Faculty of Health Medicine and Life Sciences, Maastricht University, Maastricht, Netherlands. 17 Institute of Cognitive Sciences and Technologies, Consiglio Nazionale delle Ricerche, Rome, Italy. 18 Department of Informatics, Athens University of Economics and Business, Athens, Greece Purpose/Objective Artificial intelligence (AI) can transform clinical decision-making by improving diagnostics accuracy, reducing clinician workload, and enabling personalised treatments. However, AI models may not perform equally across patient groups, potentially causing unfair treatment recommendations. The PRE-ACT project [1] (Prediction of Radiotherapy side Effects using explainable AI for patient Communication and Treatment modification) aims to advance personalised radiotherapy (RT) by predicting the risk of developing side effects like arm lymphoedema in breast cancer (BC) patients. The aim of this study was to investigate fairness within PRE-ACT's AI models using datasets from multi-center BC cohorts (REQUITE, HypoG-01, CANTO, total n=6361). Material/Methods We trained a lymphoedema prediction model on all three cohorts, with ground-truth labels derived from follow-up lymphoedema assessments at 12, 36, and 60 months following RT. Data missing not-at-random (MNAR) may introduce biases into the model and raise fairness issues. Our analysis focuses on the CANTO-RT cohort (n=3080) [2], which includes socioeconomic variables. We evaluated model performance across different socioeconomic groups using ROC Area Under the Curve (AUC). To estimate uncertainty, we interpret AUC as a Wilcoxon statistic [3], enabling more accurate fairness assessments, especially in small subgroups where AUC scores are often imprecise. Results In CANTO, we show disparities in data availability, particularly in lower-income and lower-education groups, which have a higher proportion of patients that failed to attend the follow-up and have missing values (Figure 1). There is