ESTRO 2025 - Abstract Book

S3772

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

the heart base on OS was similar between both models McWilliam2017 – HR:1.012 (95% CI:1.003-1.021), minimal set – HR:1.009(95% CI:1.002-1.017), of note is that the CI of the result with the minimal adjustment set is narrower.

Conclusion: This work shows how causal inference techniques, and DAGs in particular, can be used to derive accurate measures of causal effects in radiotherapy research utilising VBA. This technique provides a systematised way of recognising and accounting for bias which enables greater confidence in results.

Keywords: Voxel-based analysis, causal inference,lung cancer

References: [1] McWilliam, A., Kennedy, J., Hodgson, C., Vasquez Osorio, E., Faivre-Finn, C., & van Herk, M. (2017). Radiation dose to heart base linked with poorer survival in lung cancer patients. European Journal of Cancer , 85 , 106–113. https://doi.org/10.1016/J.EJCA.2017.07.053

2238

Digital Poster Machine Learning Algorithms for Predicting Risk of Recurrence After Total Neoadjuvant Therapy in Locally Advanced Rectal Cancer Ricardo Andrés Oyarzun Silva 1 , María del Carmen Herrera Del Valle 1 , Laura Diaz Gomez 1 , Pablo Hernández Hernández 2 , Javier Jaén Olasolo 1,3 1 RADIATION ONCOLOGY, HUPM, Cádiz, Spain. 2 Anestesiae, HUPR, CÁDIZ, Spain. 3 INiBICA, HU JEREZ, CÁDIZ, Spain Purpose/Objective: To evaluate and compare the performance of three machine learning algorithms (Random Forest, XGBoost, and LightGBM) against traditional logistic regression for predicting recurrence risk in locally advanced rectal cancer treated with hypofractionated radiotherapy. Material/Methods: Retrospective single-center study (n=143, 2017-2022) including patients with locally advanced rectal cancer treated with hypofractionated radiotherapy (25Gy in 5 fractions using VMAT), followed or not by consolidation chemotherapy according to the RAPIDO protocol. Analyzed variables included clinicopathological characteristics (age, sex, TNM staging, tumor height) and molecular biomarkers (BRAF, RAS, EGFR, IMS). Four predictive algorithms (Random Forest, XGBoost, LightGBM, and Logistic Regression) were implemented using stratified cross-validation with Repeated K-Fold (k=10, repetitions=5). Results: Random Forest demonstrated superior performance (AUC=0.85, F1-score=0.67, Precision=0.53, Recall=0.87, MCC=0.59) compared to Logistic Regression (AUC=0.81), LightGBM (AUC=0.77), and XGBoost (AUC=0.71). In the validation set, Random Forest achieved precision=0.97 and recall=0.80 for non-recurrent patients, and precision=0.50 and recall=0.88 for recurrent patients (global accuracy=0.81). The most relevant predictive factors were age (23.2%), sex (19.0%), BRAF status (11.8%), and RAS status (10.7%).