Purpose: To detect and quantify the association between cardiac substructure irradiation and the risk of developing radiation pneumonitis (RP) within a multivariate framework for patients treated with conventional external beam radiotherapy for non‐small‐cell lung cancer (NSCLC). Methods: All evaluable 3‐D conformal radiation therapy (3D‐CRT) treatment plans for patients with registered outcomes treated for NSCLC between 1991 and 2001 were eligible for this study (N=209). RP events occurred in 49 (23.4%) patients, with events defined as requiring steroid management or more intensive intervention (CTCAE v.4.0 Grade 2 or greater). Cardiac substructures including the left ventricle, right ventricle, left atrium, right atrium, ascending aorta, and descending aorta were contoured and individually reviewed by a single physician. Cross‐validation methods were used to build a multivariate model included clinical factors (age, gender, race, performance status, weight loss, smoking history, and histology); dosimetric parameters for cardiac substructures and normal lung [D5–D100, V10–V80, mean dose, maximum dose, and minimum dose ]; other treatment factors (chemotherapy, treatment time, fraction size); and the center of mass of the GTV within the lung, in the superior‐inferior dimension (GTV_COMSI). An optimal multivariate model was obtained by step‐wise variable selection and logistic regression. Results: Statistically significant variables (p less than 0.05) with the highest univariate Spearman rank correlations (Rs) included: Left atrium D5 (Rs,0.235), left atrium D10 (Rs,0.228), right ventricle Dmax (Rs,0.2), right atrium Dmax (Rs,0.197), and GTV_COMSI (Rs,0.22). The optimal logistic model used four variables, incorporating the left atrium D20 and V30, the lung D35, and the GTV_COMSI; (Rs=0.31) with an area under the receiver‐operator curve of 0.75. The best model using only whole‐heart variables had an Rs=0.26. Conclusion: Left atrium dose‐volume parameters had greater predictive power than other heart substructures and previously derived lung parameters to predict RP, and were incorporated into a robust prognostic multiparametric model.