S3029
Invited Speaker
ESTRO 2026
reproducibility and generalizability across cohorts, centers, and treatment workflows, while also checking compatibility with plausible pathophysiological hypotheses. Clinical utility should finally ask whether VBA findings can be translated into simplified, reproducible structures, constraints, or hybrid prediction tools usable in treatment planning. The overall message is that, in radiation oncology, model validation often means validating not only a predictor, but the full inferential chain that generated it. 5408 Harnessing big data and predictive informatics for model validation: The potential of AI and data science Laia Humbert-Vidan Radiation Oncology, Vall d'Hebron Institute of Oncology (VHIO), Barcelona, Spain Radiotherapy predictive models have long been proposed as tools to support personalised treatment and toxicity risk stratification. However, their translation into routine clinical practice remains limited. A key barrier lies in how these models are developed and validated: many rely on relatively small, single-institution datasets and lack robust external validation, resulting in poor generalisability when applied in real-world clinical settings. Addressing this gap requires a shift towards large- scale, diverse data sources. Multi-centre datasets, disease-specific registries, and international data consortia offer the scale and variability needed to properly assess model performance across populations and clinical practices. At the same time, practical constraints around data sharing have driven the emergence of distributed approaches such as federated, swarm, and other decentralised learning frameworks, enabling multi-institutional validation without the need to centralise sensitive patient data. The use of such large and heterogeneous datasets introduces important challenges, including variability in imaging protocols, treatment techniques, outcome definitions, and data completeness. Overcoming these limitations requires robust data standardisation, interoperable infrastructures, and harmonised analytical pipelines. Within this context, AI and data science play a critical enabling role. Rather than focusing solely on model development, these approaches support the integration of high-dimensional data, such as imaging and dosimetric features, and facilitate scalable, reproducible validation through automated pipelines and cross-institution benchmarking. Finally, there is a need to move beyond static, cross- sectional validation. Radiotherapy is delivered within evolving clinical environments, where technologies,
and practical barriers. Understanding and addressing these barriers is essential if prediction models are to become truly useful in everyday radiotherapy practice.
5407 Statistical methods for model validation in radiation oncology: Best practices and challenges Giuseppe Palma Institute of Nanotechnology, National Research Council, Lecce, Italy Prediction models in radiation oncology are sometimes discussed as if validation were a single performance estimate on a hold-out test set. In practice, validation is a multi-layered process that must distinguish internal validation, internal-external validation, and true external validation, and must assess calibration and clinical consequences in addition to discrimination. The aim of this presentation is to clarify what must be statistically validated before model outputs can be considered credible for clinical translation. The first part of the talk will briefly revisit core principles of predictive modelling in radiation oncology: the limitations of random split-sample strategies; the advantages of bootstrapping for internal validation during model development, and of internal-external cross-validation to examine transportability across centers or cohorts; the inadequacy of relying on a single area under the curve; and the importance of calibration plots, calibration-in- the-large, calibration slope, uncertainty quantification, and decision-analytic measures to assess whether a model is reliable in new patients and new centers. The second part will focus on voxel-based analysis (VBA) as a paradigmatic case. As a high-dimensional methodology that can inform spatial biomarkers, constraints, and hybrid prediction tools, VBA does not only test a regression equation: it interrogates a spatial dose-response pipeline built on image acquisition, contouring, deformable registration, dose mapping, and voxel-wise statistical inference. This makes validation intrinsically more demanding. Key challenges include the choice of common coordinate system, the accuracy of spatial normalization, contouring variability, planned versus accumulated dose, spatial autocorrelation, multiple testing, and the dependence of results on smoothing, filtering, and error-control strategy. In this setting, a statistically significant spatial map is not yet equivalent to a clinically trustworthy finding. The talk will therefore propose a practical framework organized into three linked layers. Technical validation should test the robustness of the spatial signal to methodological choices and quantify mapping uncertainty. Clinical validation should examine
Made with FlippingBook - Share PDF online