Materials/Methods: For this retrospective treatment planning study, we used the planning CT images of 23 patients treatedwith PORT for locally advanced NSCLC with a 2-dimensional (2D) technique. Intentionally, this study included a mix ofpatients: 12 male, 16 right-sided tumors, 8 with subcarinal node disease, 5 with lower lobe involvement, and 3 post-pneumonectomy.

The clinical target volume (CTV) encompassed 8 nodal regions, including bilateral supraclavicular (S/C) fossa, mediastinum,and ipsilateral hilum. The planning target volume (PTV) included the CTV with a 0.5 cm margin.

2D and 3D-CRT plans consisted of AP/PA fields to 39.6 Gy followed by off-cord oblique fields to deliver an additional 10.8Gy. 4-field IMRT plans used the same beam directions as the 2D and 3D-CRT plans, and 5-field IMRT plans included 5equally-spaced beams directions. A first phase treating the entire PTV to 39.6 Gy was followed by a cone-down phase to 10.8Gy.

To evaluate the heart and lung doses, we analyzed the mean dose to each structure, the percent volume of heart receivingat least 30 Gy (V30), and the percent volume of lung receiving at least 20 Gy (V20).

Results: The patient-averaged mean heart doses for the 2D, 3D-CRT, 4-field IMRT, and 5-field IMRT plans were 20.6 Gy, 13.4Gy, 13.4 Gy, and 12.0 Gy, respectively (2D vs. 3D-CRT, p�0.001; 3D-CRT vs. 5-field IMRT, p�0.01; 4- vs. 5-field IMRT,p�0.002). In all patients, 2D plans demonstrated higher heart doses than the other 3 plans. 5-field IMRT plans had much lowerheart doses than 2D plans, especially in the 10 patients with subcarinal or lower lobe disease. 3D-CRT and 4-field IMRT plansalso had lower heart doses than the 2D plans, but not as low as the 5-field plans.

For all patients, the average lung V20 and mean lung dose were significantly higher (p�0.001) in the 2D plans. The lungdoses in the 3D-CRT, 4-field IMRT, and 5-field IMRT plans were fairly similar. Dose-volume statistics are summarized in thetable.

Conclusions: For patients with locally advanced NSCLC receiving post-operative radiotherapy, 3D-CRT and IMRT plansresulted in significantly lower radiation doses to the heart and lungs than 2D plans. IMRT substantially reduced radiation doseto the heart in patients with subcarinal disease or lower lobe involvement, but was comparable to 3D-CRT for patients withoutsubcarinal or lower lobe disease. We prefer IMRT in patients with subcarinal or lower lobe disease and 3D-CRT for otherpatients.

1136 Improved Prediction of Radiation Pneumonitis using Multiple Additive Regression Trees

S.K. Das, S. Zhou, Z. Kocak, F. Yin, L. Marks

Dept of Radiation Oncology, Duke University, Durham, NC

Purpose/Objective: The ability to accurately predict the incidence of radiation-induced pneumonitis (RP) is essential to beingable to tailor radiotherapy treatment (RT) plans to reduce this toxicity. Classification and regression trees (CART) are routinelyused in RT toxicity modeling to extract the most significant parameters. In this work we implement and evaluate a recent

S228 I. J. Radiation Oncology ● Biology ● Physics Volume 63, Number 2, Supplement, 2005

statistical technique that combines weighted individual CARTs into a multiple additive regression tree (MART), to determineif it is a superior predictor of radiation pneumonitis.

Materials/Methods: Data for this study comes from 148 lung cancer patients enrolled on an IRB-approved prospective studyto assess lung toxicity resulting from RT (40–86 Gy at 1.6 Gy BID/1.8 - 2 Gy/fraction). Of these patients, 23 developed Grade�2 RP necessitating steroids and/or oxygen. Patient three-dimensional RT plans and dose-volume histograms were generatedusing the PLUNC treatment planning system. To predict RP, a program was written to generate individual CARTs and combinethem into a single MART. The goal of each successive component CART is to rectify the misclassifications from the previouslygenerated CARTs. The MART used as input parameters the lung dose-volume histograms, age, sex, race, tumor site (central,peripheral, upper, middle, lower, right, left), chemotherapy, surgery, pre-RT forced expiratory volume in 1 second (FEV1), andpre-RT Carbon Monoxide diffusion capacity (DLCO). Each component CART of the MART was grown until 10-fold crossvalidation indicated overfitting. Each successive CART component was added to the MART with diminishing weight. Again,to prevent overfitting, the optimal number of successive CARTs and their respective weights were selected using 10-fold crossvalidation. Two metrics were used in the 10-fold cross validation process: average of sensitivity and specificity, and the areaunder the Receiver Operating Characteristics curve. Sensitivity is the percentage of cases with RP that are correctly classified,and specificity is the percentage of cases without RP that are correctly classified. The higher the 10-fold cross validated averageof sensitivity and specificity (ASPS), the higher the predictive capability of the MART. Receiver operating characteristics plotsthe sensitivity vs. 1-specificity. The higher the area under the receiver operating characteristics curve (AUC), the more effectiveis the model (a model that is only as good as chance has an AUC � 0.5, and a model that is always correct an AUC � 1.0).

Results: Ten-fold cross validated average of sensitivity and specificity (ASPS) varied with increasing number of CARTs as:54% (1 CART), 83% (2), 87% (3), 88% (4), 86% (5), 85% (6), 85%(7). The corresponding AUCs were 0.54 (1), 0.92 (2), 0.90(3), 0.96 (4), 0.91 (5), 0.93 (6), 0.95 (7). Thus, the optimal MART consisted of 4 component CARTs (87% ASPS, 0.96 AUC),with each CART made up of between 1–4 levels. Using only a single CART, which is the first MART component, yielded anASPS of only 54%, and AUC of 0.54. The optimal MART extracted 9 significant parameters: lung volumes above 24 Gy (31%influence), above 36 Gy (16%), lower pre-RT FEV1 (9%), lung volumes above 90 Gy (9%), above 32 Gy (5%), above 70 Gy(4%), above 60 Gy (4%), tumor located in the upper lobe (3%), female sex (3%).

Conclusions: Greatly improved radiation penumonitis predictive capability can be achieved using multiple additive regressiontrees (87% ASPS, 0.96 AUC) when compared to a single classification and regression tree (54% ASPS, 0.54 AUC). Theimproved predictive capability of multiple additive regression trees may be effectively employed to alter radiotherapy plans todecrease the predicted incidence of radiation pneumonitis.

Funded in part by a grant from Varian Medical Systems and NIH R01 Grant CA69579.

1137 Radiation Pneumonitis in Lung Cancer Patients - The Neglected Patient-Related Variables

A. Bezjak,1 V. Soyfer,2 Q. Yi,1 A. Sun,1 G. Kane,1 J. Waldron,1 J. Cho,1 W. Wells,1 D. Payne1

1Radiation Oncology, Princess Margaret Hospital, University of Toronto, Toronto, ON, Canada, 2University of Tel Aviv,Tel Aviv, Israel

Purpose/Objective: Limited radiation tolerance of lung tissue remains a challenge in radiation (RT) treatment of patients (pts)with lung cancer. Most attention has been given to treatment-related variables that influence the development of RTpneumonitis, particularly RT dose and volume of lung receiving a certain dose. A number of pt-related factors are known tobe relevant to the development and clinical severity of RT pneumonitis, most notably pts pre-existing lung function. There isevidence from animal experiments, and limited information from humans, that some frequently prescribed medications maymodule the risk of developing RT pneumonitis.

To determine the potential protective role of concurrent medications, specifically glucocorticoids, angiotensin convertingenzyme (ACE) inhibitors and/or NSAIDs (non-steroidal anti-inflammatory drugs) including aspirin, on development of RTpneumonitis in pts treated by radical RT for non-small cell lung cancer (NSCLC).

Materials/Methods: Incidence and severity of RT pneumonitis were retrospectively analyzed for 142 NSCLC pts treated byRT to doses � 60 Gy (with or without chemotherapy) at Princess Margaret Hospital between 1999 and 2003. One of thevariables of interest was administration of concurrent medications, specifically glucocorticoids, ACE inhibitors and/or NSAIDs.Pneumonitis was retrospectively scored according to modified RTOG/EORTC pulmonary toxicity grading.

Results: Overall incidence of clinically significant (grade 2–5) RT pneumonitis was 26%. Of 77 pts who received concurrentmedications in any combinations, 14 (18.2%) developed Grade �2 RT pneumonitis, as compared to 23/65 (35.4%) of pts whowere not on these medications (p�0.02). The strongest protective effect was observed with ACE inhibitors, with 3/24 (12.5%)of pts on ACE inhibitors showing evidence of � grade 2 RT pneumonitis, as compared to 34/118 (28.8%) of pts not taking thatmedication (p�0.097). The difference was present but less marked for steroids (5/23 (21.7%) of pts on steroids showing �grade2 RT pneumonitis, vs 32/119 (26.9%) of pts not on steroids) and NSAIDS (10/51 (19.6%) vs 27/91 (29.7%). On univariateanalysis, presence of chronic lung disease, concurrent chemotherapy and smoking status were not correlated with clinicallysignificant RT pneumonitis, and neither was fraction size (� 2 Gy/ fr vs �2Gy/fr). On multivariate analysis, the effect ofconcurrent medications persisted after adjustment for smoking, fractionation size, preexisting lung disease and concurrentchemotherapy.

Conclusions: Pts who are on medications with potentially protective mechanism on lung tissue may have lower incidence ofRT lung toxicity. If these findings are confirmed, they may be a reason why the current methods of trying to predict the riskof RT pneumonitis from a given RT plan are inaccurate, as they focus on treatment-related variables and ignore potentiallyimportant patient-related factors, which may be easily modifiable.

It is important both to document concurrent medications, and to pursue these findings with confirmatory studies, in order totry to module the risk of RT and improve the therapeutic ratio.

S229Proceedings of the 47th Annual ASTRO Meeting