Multiomic based Bayesian network toxicity modeling for simultaneous prediction of multiple toxicity outcomes in NSCLC.

Abstract

Radiation Pneumonitis (RP) and Radiation Esophagitis (RE) are two prominent dose limiting toxicities from NSCLC radiotherapy. This study aimed to develop a multi-objective Bayesian network (BN) model to predict multiple NSCLC outcomes simultaneously using dose-volume histograms (DVH), clinical, and multiomic features. Out of 179 NSCLC patients considered, 30.2 % patients were found to have toxicity ≥grade 2 RP and 15.08 % had ≥grade 3 RE per CTCAE v5.0. A total of 672 features were extracted. We built a parallel dimensionality reduction technique by introducing an L1-norm penalty to select the best omic features for predicting RP and RE. We then employed an additional penalized logistic regression approach for feature selection on reduced covariates. A score-based structure learning algorithm (Tabu search) was used to build a BN model for multi-outcome predictions of RP and RE. A cross-validation approach was used to help create the optimal structure with a train/test split of 70 %/30 % respectively. The model predictive performance was evaluated by area under the curve (AUC) and area under the precision-recall curve (AUPRC). Free-response area under the curve (FRAUC) was also used to evaluate joint prediction performance. The AUC and AUPRC values on the test set for RP and RE predictions were 0.86/0.76 and 0.81/0.50, respectively. The FRAUC value on the test set was 0.74. We developed a highly predictive multi-objective BN model using multiomic covariates that was capable of simultaneously predicting RP and RE, which are two of the most prominent toxicities in NSCLC radiotherapy.