Modeling variable interactions using Bayesian networks to identify direct predictors of radiation-induced optic...

Introduction.Radiation-induced optic neuropathy (RION) is a rare but functionally-disabling complication of radiotherapy for brain and head/neck tumors. Even with modern proton therapy and adherence to dose constraints, about 6% of patients develop near-blindness from RION. Unlike regression models that estimate additive effects, a Bayesian network (BN) explicitly encodes conditional dependencies among input variables, and was used for more reliable assessment and individualized outcome prediction of RION.Methods.Demographic, ophthalmological and treatment data were prospectively collected from patients undergoing scanned proton beams between 2018 and 2024. Visual field deficit served as surrogate for grade ⩾2 clinically-relevant RION. Using clinical and treatment/dosimetry variables, under expert-informed constraints, BN with Markov-blanket identified direct RION predictors; its explanatory value was evaluated via log-likelihood ratio (LLR), integrated discrimination improvement (IDI), and net reclassification index (NRI).Results.Of 179 eyes (105 patients), 24% developed RION. The BN identified 15 directional associations, 12 were clinically valid, 2 conceivable and 1 weak. The Markov-blanket of the RION node identified baseline visual field deficit, hypertension, and hypercholesterolemia as key predictors. A logistic regression model based on these variables improved patient stratification (LLR = 63.34,p< 0.001, IDI = 0.35, NRI = 0.25); AUC reached 0.75 (95%CI: 0.66-0.84). Age and dosimetric factors did not improve RION risk prediction.Conclusion.BN analysis highlighted associations between vascular comorbidities and baseline visual function and occurrence of RION, while dosimetric factors were not direct predictors. This contrast with photon models is hypothesis-generating and favors proton-specific models for RION prediction.