Making Label Noise Useful: Uncertainty-Aware Prediction of Radiation-Induced Toxicity from Spatial Dose Information
Abstract
Purpose
Accurate prediction of radiation-induced toxicity is crucial for optimizing radiotherapy outcomes, yet most existing models rely on supervised learning with clinician-graded toxicity scores that are susceptible to patient self-reporting errors and intra-observer variability. This study aims to develop an explainable, uncertainty-aware framework for toxicity prediction when labels are noisy.
Methods
We developed a probabilistic model for radiation toxicity prediction using a Bayesian neural network (BNN) that models label uncertainty. Spatial dose distributions were summarized into dose clusters and used as inputs to the BNN. Network weights were modeled using Gaussian variational distributions with learnable means and variances, assuming a standard normal prior and Bernoulli likelihood for binary outcomes. The model was trained by minimizing the evidence lower bound with β-annealing. During inference, Monte Carlo sampling (n=100) generated predictive distributions, from which prediction means and uncertainty estimates were derived. Performance was evaluated using discrimination metrics, calibration measures, and comparing to corresponding deterministic models. The approach was tested on RTOG 0617 trial data (n=457), with radiation pneumonitis dichotomized at grade ≥ 2.
Results
The probabilistic model achieved improved discrimination compared with the deterministic model (AUC-ROC: 0.725 vs. 0.707), with higher accuracy (0.783 vs. 0.551) and specificity (0.797 vs. 0.492) on the test set (n=69), but reduced sensitivity, reflecting a more conservative decision boundary. The model exhibited an expected calibration error of 0.231, indicating imperfect calibration and the need for post hoc calibration. Predictive uncertainty estimates ranged from 0.08 to 0.22, identifying a subset of cases with elevated uncertainty that may benefit from physician review.
Conclusion
This study demonstrates that an uncertainty-aware Bayesian network can predict radiation toxicity prediction while providing uncertainty information beyond deterministic models. Further validation with larger cohorts and enhanced feature representations is warranted to assess robustness and clinical utility.