Simultaneous Prediction of Specific Radiotherapy Outcomes Using a Multi-Objective Bayesian Network (moBN) Approach

Abstract

Purpose/Objective(s): Models for predicting radiotherapy outcomes are typically developed individually per outcome, which overlooks potential intercrossing biophysical relationships when optimizing competing risks for treatment planning or clinical trial design. Here, we develop a joint outcomes model for simultaneous prediction of local control (LC) and multiple undesired complications in non-small cell lung cancer (NSCLC) patients that investigate hierarchical relationships among clinical, physical, biological, and imaging radiomics information using probabilistic graphical models. Purpose/Objective(s): One hundred eight patients treated on prospective protocols under institutional review board approval were considered. Sixty-eight patients selected for training had 48 cases of LC, 5 events of >= grade 3 (G3) radiation pneumonitis or fibrosis (RPFIB3), 11 events of >=G3 cardiac toxicity (Car3) and 7 events of >=G3 esophagitis (Eso3). Each patient had 289 features from different biophysical resources (dosimetric, SNPs, miRNAs, clinical factors, pre-and during-treatment cytokines, pre-and during-treatment PET data). A multi-objective Bayesian network (moBN) approach was developed to find the most relevant biophysical features for simultaneous prediction of the four clinical endpoints, and to explore the intercross-relationships of important features. Area under the free-response receiver operating characteristics curve (AU-FROC) measured the prediction performance of the integrated moBN internally on cross-validation and externally on the remaining testing set of 40 patients not used in training. Result(s): The resulting moBN for prediction of the four combined outcomes includes 4 dosimetric, 6 SNPs, 2 miRNAs, 4 cytokines (1 pre, 3 during) and 3 PET radiomics (1 pre, 2 during) features, and indicates biophysical inter-relationships and crosstalk among the variables predicting the different endpoints. Performance prediction analysis is summarized in Table 1. The combined moBN achieved an AU-FROC of 0.87 on internal cross-validation and 0.79 on the testing dataset; this is compared to AU-FROCs of 0.66 and 0.60, respectively, using only common dose-volume metrics for NSCLC. Conclusion(s): The moBN approach shows promise to predict an individual NSCLC patient's LC and radiation-induced toxicities simultaneously, indicating potential for incorporation into a radiotherapy decision support system. It also reveals biophysical relationships among variables and cross-talk in predicting outcomes. However, further independent testing on multi-institutional datasets may still be required for prediction, as would further in vitro and in vivo experimental validation of the biophysical relationships.