ESTRO 2025 - Abstract Book

S3800

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2025

2757

Proffered Paper Addressing Imaging and Biomarker-driven Uncertainty in Machine Learning-based Radiotherapy Outcome Prediction Alice Bondi 1 , Gregory Buti 2 , Antony Lomax 3 , Thomas Bortfeld 2 , Xinru Chen 4 , Ting Xu 4 , Zhongxing Liao 4 , Ali Ajdari 2 1 Biomedical Engineering, ETH Zurich, Zurich, Switzerland. 2 Radiation Oncology, Massachusetts General Hospital & Harvard Medical School, Boston, USA. 3 Physics, Paul Scherrer Institut (PSI), Villigen, Switzerland. 4 Radiation Oncology, University of Texas' MD Anderson Cancer Center, Houston, USA Purpose/Objective: With the explosion in and heightened interest in multi-omics data, advanced machine learning (ML) models are increasingly used to predict treatment outcomes and pave the way towards personalized radiotherapy (RT). However, inherent uncertainty in the input data can significantly impact the accuracy of models’ predictions, undermining their clinical utility. We aim to (i) systematically characterize the impact of these uncertainties on the performance of various ML-based outcome predictive models, and (ii) develop “upstream” data- and optimization driven solutions to improve models’ robustness to data perturbations. Material/Methods: A non-small cell lung cancer (NSCLC) dataset (n=219) was utilized to predict acute cardiac adverse events (CAE) using logistic regression (LR), support vector machines (SVM), random forest (RF), and neural network (NN). Model inputs included dosimetric indices of cardiac substructure (four chambers), clinicopathological factors, and high-sensitivity cardiac troponin T (hs-cTnT)—an established blood biomarker of cardiac injury [1]. Variability in cardiac (sub)structure segmentation was simulated by applying (Gaussian) random shifts within previously reported reasonable bounds [2]. Biomarker uncertainty was simulated by assuming 10% uncertainty in the mean value of hs cTnT [3]. The joint uncertainty impact on the model performance was characterized using Monte Carlo simulations. Four uncertainty mitigation strategies were tested to boost models’ robustness to data perturbation: Probabilistic random forest (PRF), Data Augmentation (DA), Adversarial Training (AT) [4], and a novel hybrid of DA and AT, termed Adversarially-Informed Data Augmentation (AI-DA). The algorithms’ performance was evaluated against the conventional (uncertainty-agnostic) ML model. Results: Segmentation uncertainties caused a wide variation in the dosimetric indices due to ( Fig 2 ), specifically in point wise metrics (i.e., max dose). The left atrium (LA) was the most sensitive to segmentation errors (up to 12Gy and 27Gy variations in mean- and max-dose, Fig 1-C,D ). SVM was the most robust model to data uncertainty but demonstrated the lowest AUC. Random Forest showed the best combination of robustness-performance (AUC=0.81±0.12 standard deviation, ( Fig 2-A,B ). All uncertainty mitigation techniques (applied to RF), except for PRF, improved the predictive uncertainty compared to the baseline model, with DA and AI-DA achieving the best performances. Regarding predictive uncertainty, PRF exhibited the greatest reduction in data uncertainty (~50%), while DA and AI-DA led to the largest decrease in model uncertainty (~40%)— Fig 2-C .