S2284
Physics - Machine learning and AI algorithms
ESTRO 2026
Digital Poster 1283 Evaluation of a Pre-trained Transformer-based Foundation Model (TabPFN) for Predicting Radiotherapy Toxicity in Head and Neck Cancer Thomas Young 1,2 , Laia Humber Vidan 3 , Wahyu Wulaningsih 1 , Victoria Butterworth 1,4 , Andrew King 5 , Teresa Guerrero Urbano 1,4 1 Radiotherapy, Guy's and St Thomas' NHS Foundation Trust, London, United Kingdom. 2 Cancer Studies, King's College London, London, United Kingdom. 3 Radiation Oncology Group, Vall d'Hebron Institue of Oncology, Barcelona, Spain. 4 School of Cancer & Pharmaceutical Sciences, King's College London, London, United Kingdom. 5 School of Biomedical Engineering & Imaging Sciences, King's College London, London, United Kingdom Purpose/Objective: Toxicity following head and neck cancer (HNC) radiotherapy significantly affects survivors’ quality of life. Robust normal tissue complication prediction (NTCP) models are required for treatment personalisation but classical models overlook non- dosimetric predictors. Machine learning (ML) techniques have been applied to develop NTCP models incorporating additional features. While more complex ML methods such as Random Forest (RF) may provide flexibility to data modelling, consistent performance improvement has not yet been shown compared to multivariable logistic regression (MVLR). Tabular Prior-Data Fitted Network (TabPFN) foundation model has significantly outperformed other ML methods for classification tasks and requires no hyperparameter tuning (1). It utilises a transformer- based architecture pre-trained on millions of synthetic datasets to process tabular data. In this work, we evaluated TabPFN in the context of HNC toxicity prediction in one of the first applications of a transformer-based foundation model to structured oncological data. Material/Methods: Larynx, oropharynx and hypopharynx cancer patients treated with primary radical radiotherapy, with known 1-year post-radiotherapy CTCAE v4.0/v5.0 xerostomia grading, were identified from a pre-existing dataset (2). Toxicity cases (CTCAE grades 2-4) were matched with an equal number of randomly selected non- toxicity cases to avoid class imbalance issues. Relevant variables were utilised as input data (Figure 1).To counter multi-collinearity issues (3), feature selection (including univariate logistic regression with bootstrapping, Spearman correlation filtering, and Recursive Feature Elimination) was applied to MVLR and RF, but not TabPFN models. RF hyperparameters were tuned with nested cross-validation. Feature importance was calculated for models. Each model
generated six groups with statistically different survival curves. Pairwise feature distributions show a strong correlation between the risk group and the engineered features.
Conclusion: We present a novel pipeline for interpretable multimodal survival analysis. Our approach supports transparent risk prediction and clinically actionable patient stratification, representing a step toward clinical adoption of AI-based survival analysis. References: [1] L. Jiang, C. Xu, Y. Bai, A. Liu, Y. Gong, Y.-P. Wang, and H.-W. Deng. Autosurv: interpretable deep learning framework for cancer survival analysis incorporating clinical and multi-omics data. NPJ Precision Oncology, 8(1):4, 2024.[2] M. Malafaia, T. Schlender, T. Alderliesten, and P. A. N. Bosman. A Step towards Interpretable Multimodal AI Models with MultiFIX. In Proceedings of the Genetic and Evolutionary Computation Conference Companion, pages 2001– 2009, 2025.[3] M. L. Welch, S. Kim, A. J. Hope, S. H. Huang, Z. Lu, J. Marsilla, M. Kazmierski, K. Rey- McIntyre, T. Patel, B. O’Sullivan, et al. RADCURE: An open-source head and neck cancer CT dataset for clinical radiation therapy insights. Medical Physics, 51(4):3101–3109, 2024. Keywords: interpretability, multimodality, survival analysis
Made with FlippingBook - Share PDF online