Multi-region Multiomic Random Forest Toxicity Modeling of Radiation Pneumonitis

Abstract

PurposeRadiation pneumonitis (RP) is a major dose-limiting toxicity resulting from non-small-cell lung cancer (NSCLC) radiotherapy. Multiomic features (radiomics and dosiomics) could provide additional predictive information as compared to traditionally used clinical and dose-volume histogram (DVH) parameters. We aimed to investigate the utility of multiomic features to improve RP toxicity models.MethodsOut of 329 NSCLC patients considered, 85 patients (25.84%) were found to have toxicity ≥ grade 2 RP per CTCAE v5.0. A total of 422 radiomic and dosiomic features were extracted. Four toxicity prediction model types were created using clinical factors together with respective features from one of the following groups: (a) DVH (base model), (b) whole lung radiomics and dosiomics (WL-RD), (c) multi-region radiomics and dosiomics (MR - RD) and (d) multi-region DVH, radiomics and dosiomics (MR-DVHRD). Toxicity models were created using a random forest classifier with a Monte Carlo cross-validation approach of 100 iterations, and a training/test split of 80%/20%, respectively. Model predictive performance was evaluated by area under the receiver operating characteristic curve (AUC) and area under the precision-recall curve (AUPRC).ResultsThe AUC and AUPRC values (mean ± standard deviation) for the 4 model types were 0.81±0.04/0.70±0.06 (base model), 0.82±0.05/0.73±0.08 (WL-RD, p<0.05), 0.83±0.06/0.75±0.08 (MR–RD, p<0.05), and 0.82±0.05/0.72±0.08 (MR-DVHRD, p<0.05), respectively, wherein a paired test compared the performance metrics of omic models with the base model built on each iteration of cross0020validation.ConclusionsAll multiomic model types outperformed the base DVH model. MR-RD model had the best performance among all model types.