Author: Ali Ajdari, Alice Bondi, Thomas R. Bortfeld, Gregory Buti, Xinru Chen, Zhongxing Liao, Antony John Lomax, Ting Xu 👨🔬
Affiliation: The University of Texas MD Anderson Cancer Center, Department Of Radiation Oncology, Massachusetts General Hospital (MGH), Massachusetts General Hospital & Harvard Medical School, Paul Scherrer Institut, ETH Zurich 🌍
Title: Addressing Imaging and Biomarker-driven Uncertainty in Machine Learning-based Radiotherapy Outcome Prediction
Alice Bondi, Gregory Buti, Antony Lomax, Thomas Bortfeld, Xinru Chen, Ting Xu, Zhongxing Liao, Ali Ajdari
[Affiliation]
Purpose: To (i) systematically characterize the impact data- and model-based uncertainties on the performance of various machine learning (ML) outcome predictive models in radiotherapy (RT), and (ii) develop uncertainty mitigation strategies to improve models’ robustness.
Methods: A non-small cell lung cancer (NSCLC) dataset (n=219) was retrospectively analyzed. Four predictive models (logistic regression, support vector machines, random forest, and neural network) were trained to predict Cardiac Adverse Event (CAE) following RT using cardiac substructure's dosimetric indices, clinicopathological factors, and cardiac troponin T (cTnT)—an established blood biomarker of cardiovascular risk. A Monte Carlo approach was used to simulate the joint impact of variability in cardiac (sub)structure segmentation and biomarker uncertainty on CAE predictions. Four Uncertainty-Aware training strategies were tested to boost models’ robustness: Data Augmentation (DA), Probabilistic machine learning (PML), and two novel variants of Adversarial Training (AT). All models were tested under various Adversarial Attacks to quantify their sensitivity to data perturbation.
Results: Among cardiac substructures, left atrium (LA) showed the most sensitive to segmentation errors (up to 12Gy and 27Gy variations in mean- and max-dose). Random Forest model showed the best combination of robustness and performance (AUC of 0.81±0.12 standard deviation). Applying the uncertainty mitigation to the RF model, we found that all uncertainty mitigation techniques dramatically improved the predictive uncertainty (mean=20.2%, range=[8.6-39.7%]) compared to the baseline (uncertainty-agnostic) model. Except for PRF, the other three strategies improved AUC (from 0.80 to 0.83) while reducing models’ uncertainty.
Conclusion: Data uncertainties can cause predictions to cross decision boundaries, either correcting misclassifications or revealing fragility in the model’s predictions, highlighting the need for uncertainty-aware methods in clinical practice.