(P39) Novel Use of Machine Learning for Predicting Radiation Pneumonitis in Locally Advanced Stage II-III Non-Small Cell Lung Cancer

Abstract

Radiation pneumonitis (RP) is a common and radiotherapy dose-limiting toxicity for locally-advanced non-small cell lung cancer (LA-NSCLC). Prior studies have proposed relevant dosimetric constraints in order to limit this toxicity. We incorporate machine learning techniques in order to perform additional in-depth analyses of contributing factors to the development of RP in order to uncover previously unidentified criteria, establish more discrete and robust dosimetric criteria, and the elucidate the relative importance of individual factors. Utilizing a cohort of 203 consecutive stage II-III LA-NSCLC patients treated with definitive chemoradiation at our institution from 2008-2016, we evaluated 32 continuous and categorical features per patient grouped into risk factors, comorbidities, pretreatment imaging, stage, histology, radiation treatment, chemotherapy, and dosimetry utilizing a novel machine learning algorithm to identify predictive features for the development of RP. Univariate analysis was performed using optimally trained decision stumps to determine statistically significant features and their corresponding RP thresholds. Multivariate analysis was carried out using Mediboost for feature selection. MediBoost is a novel patient stratification tool that builds decision trees with optimal accuracy comparable to less-interpretable ensemble methods commonly used in the literature. Patients were treated to a median dose of 66.6 Gy in 1.8 Gy fractions. 17.7% patients developed grade ≥3 RP. On univariate analysis, lung dosimetric factors of lung mean, lung V10, and lung V20 were found to be the significant RP predictors with the greatest balance of specificity and sensitivity. On multivariate analysis, using the MediBoost algorithm, esophagus max (61.5%) and lung V20 (30.5%) were the two most common primary differentiators of RP, with lung mean and esophagus mean important secondary RP differentiators. We highlight the utility of a novel machine learning algorithm and demonstrate its ability to both identify known and novel predictors for the development of symptomatic RP. We provide interpretable decision trees using MediBoost that can aid clinicians in decision optimization regarding mitigation of RP and demonstrate the potential utility of our machine learning techniques for future studies.