ESTRO 2025 - Abstract Book

S2349

Interdisciplinary – Other

ESTRO 2025

significant variation in AUC between socioeconomic groups, with AUC differing by income and education. Confidence intervals around the estimate allow a more precise assessment of uncertainty, with the lowest AUC and largest confidence interval seen for patients with low educational and low income level (Figure 2).

Figure 1: Percentage of missing values in CANTO at each follow-up phase after radiotherapy treatment, for different income categories (left) and education categories (right).

Figure 2: Overall AI model AUC with confidence intervals (horizontal lines) and subgroup-level AUC scores (markers) with confidence intervals (vertical lines).

Conclusion Focusing on the prediction of RT side-effects, this study underscores the need for considering fairness in AI-driven clinical decision-making. Clinical datasets often contain missing values from specific subgroups that can lead to biased model performance for these populations with increased uncertainty around the final estimate. Missing data might compromise model accuracy, raising significant fairness concerns. Addressing these issues could involve proactive measures towards improving follow-up data collection or adjusting AI models to mitigate biases.

Keywords: AI Fairness, Breast cancer RTT, Socioeconomic bias

References [1] Charalampakos, F., Tsouparopoulos, T., Papageorgiou, Y., Bologna, G., Panisson, A., Perotti, A., & Koutsopoulos, I. (2023). Research Challenges in Trustworthy Artificial Intelligence and Computing for Health: The Case of the PRE-ACT