2640 - Proton-Specific Dosimetric Predictors of Radiation-Induced Side Effects in Locally Advanced Non-Small Cell Lung Cancer Treated with Chemoradiotherapy

Purpose/Objective(s):

Radiation pneumonitis remains clinically significant in LA-NSCLC; although protons reduce uninvolved lung dose, proton-specific planning yields different dose–volume distributions that may shift optimal DVH parameters. We sought to identify interpretable dosimetric predictors of grade = 2 pneumonitis in a proton-only LA-NSCLC cohort. Grade = 3 esophagitis and grade = 3 lymphopenia were evaluated as secondary endpoints.

Materials/Methods:

From 2010–2021, 225 consecutive LA-NSCLC patients received definitive proton radiotherapy (= 60 Gy) with concurrent chemotherapy +/- immunotherapy (after 2017). 72 treatments were delivered with pencil beam scanning, 153 with double scattering. Variables tested included clinical features (demographics and smoking history, comorbidity burden, staging, tumor histology and PD-L1 status, and baseline ALC) , dosimetric features (multi-organ DVH metrics for lung/heart/esophagus and thoracic cardiovascular substructures), and integral dose surrogates (EDIC, body mean).

Feature selection used LASSO with 5-fold stratified cross-validated regularization. XGBoost, SVM, RF, and GAM models were tuned and evaluated with 5-fold stratified CV and a held-out test set. Feature importance was quantified via impurity and permutation importance.

Results:

Crude incidence was 26.6% for pneumonitis, 7.5% for esophagitis, and 69.5% for lymphopenia. Pneumonitis models achieved test AUCs of 0.767 (GAM), 0.746 (RF), 0.718 (XGBoost), and 0.623 (SVM). Feature importance analyses identified pneumonitis risk signature dominated by right lung high dose DVH metrics (e.g., V55Gy, D80%, and lungs – IGTV D25%–D35%) and cardiovascular substructure dose (aorta, atria, ventricles, LAD), with additional contributions from lung/right-atrial volumes, EDIC/body mean dose, and baseline clinical factors. Partial dependence analyses indicated higher pneumonitis risk at right lung V55Gy > 10%, D80% > 1.5 Gy, and lungs – IGTV D25% > 40 Gy and D35% > 20 Gy, with additional modulation by EDIC < 5 and right lung volume < 2000 cc; secondary inflection points included esophagus D20% > 20 Gy, aorta D35% < 50 Gy, and atria D35% < 35 Gy.

Esophagitis reached a best AUC of 0.815 (RF; others mean 0.799) and was driven by esophageal dose-percentile and esophagus PTV overlap features, whereas lymphopenia reached a best AUC of 0.786 (GAM; others mean 0.748) and was dominated by baseline ALC and EDIC.

Conclusion:

In this proton-only LA-NSCLC cohort, models identified a consistent predictor profile linking adverse events to mid/high dose thoracic DVH exposure, anatomic scaling, and integral dose surrogates, with baseline vulnerability factors providing additional signal. These findings motivate prospective evaluation of planning objectives that limit right lung V55Gy to 10%, reduce moderate/high-dose lung exposure reflected by lungs – IGTV D25%/D35% (40/20 Gy), and monitor low-dose spread (D80% 1.5 Gy), while incorporating mediastinal substructure sparing where feasible.