TH‐AB‐304‐07: A Two‐Stage Signature‐Based Data Fusion Mechanism to Predict Radiation Pneumonitis in Patients with Non‐Small‐Cell Lung Cancer (NSCLC)

Abstract

Purpose:For NSCLC radiotherapy, toxicity outcomes such as radiation pneumonitis ≥G2 (RP2) may depend on patients’ physical, clinical, biological and genomic characteristics, and on biomarkers measured during the course of radiotherapy. This can include 100s of predictors. To reduce complexity, a two‐step, signature‐based data fusion mechanism was developed to estimate a relationship between patient specific characteristics and the probability of RP2 in terms of a modifying effect on mean lung dose (MLD).Methods:Data came from 82 NSCLC patients, 15 with RP2. Besides MLD, each had 179 predictors including 10 clinical factors (eg, age, gender, KPS), cytokines before (30) and during (30) treatment, microRNAs (49), and single‐nucleotide polymorphisms (SNPs) (60). In stage1, cytokines, microRNAs, and SNPs were used to build separate “signatures” via ridge regression. In stage2, a logistic regression predictive model for RP2 was determined in terms of MLD, the other clinical factors, and the signatures using the least absolute shrinkage and selection operator (LASSO). Leave‐one‐out cross‐validation was conducted. This was all implemented via ‘glmnet’ in the R programming environment.Results:For stage1, signatures modifying the effect of MLD for cytokine_pre, cytokine_during, microRNA and SNP included 2, 19, 3, 12 important predictors, respectively. For stage2, only the cytokine_during and SNP signatures remained as important modifying effects to MLD. The cross‐validated area under curve (AUC) reaches 0.81 (95% CI 0.70–0.89 based on 2000 stratified bootstrap replicates); significantly better than a null value of 0.50 (p<0.01).Conclusions:As implemented here, the two‐stage, signature‐based data fusion mechanism approach includes many patient specific measurements in generation of the signatures (a characteristic of ridge regression), then only includes important signatures and other clinical factors for RP2 prediction (a characteristic of LASSO). This potentially more intuitive approach to handling high dimensional predictors could be an important component of decision support for personalized adaptive radiation treatment.