3416 - Multiomic Based Bayesian Network Toxicity Modeling for Simultaneous Prediction of Multiple Toxicity Outcomes in NSCLC

Materials/

Methods: We included 179 patients treated with intensity-modulated radiation therapy (IMRT, n = 91), passive-scatter proton therapy (PSPT, n= 55) or intensity-modulated proton therapy (IMPT, n = 33) for analysis. Patient toxicity was assessed using CTCAE v5.0 where RP = grade 2 (incidence=30.16%) and RE = grade 3 (incidence = 15.08%) were the endpoints of interest. Radiomic and dosiomic features were extracted from the total lung and esophagus volumes using the planning CT, pre-treatment PET, and the 3D radiation dose volume. A total of 672 features (14 clinical, 28 DVH-based dosimetric parameters, 420 radiomic, 210 dosiomic) were extracted. We built a parallel dimensionality reduction technique by introducing an L1-norm penalty to pick the best omic features for predicting RP and RE. We then employed an additional penalized logistic regression approach for feature selection on reduced covariates to find the optimal feature set to co-predict RP and RE simultaneously. A score-based structure learning algorithm (TABU) was used to build a Bayesian network (BN) model for multi-outcome predictions of RP and RE, using the final set of covariates as inputs to the BN structure. A cross-validation approach was used to help create the optimal structure with a train/test split of 60%/40% respectively.

Results: The model predictive performance was evaluated by accuracy and area under the curve (AUC). Area under the free-response receiver operating characteristic (AU-FROC) was also used to evaluate joint prediction performance. The accuracy and AUC values on the test set for RP and RE predictions were 0.83/0.84 and 0.87/0.72, respectively. The AU-FROC value on the test set was 0.74. The most important covariate predictors were one clinical, four DVH dosimetric, one PET radiomic, and four dosiomic for RP and one clinical, four DVH dosimetric, three CT radiomic, and two dosiomic for RE.

Conclusion: We constructed a highly predictive Bayesian multi-outcome toxicity model that can predict RP and RE simultaneously. This model has a high clinical significance as it can predict two of the most prominent toxicities in NSCLC with a high degree of reliability which can aid in clinical decision making.

Abstract 3416 - Table 1: Performance for RP and RE predictions for the test set