A Predictive Model of Rectal Tumor Response to Preoperative Radiotherapy Using Classification and Regression Tree Methods

Purpose: The ability to predict rectal tumor response to preoperative radiotherapy before treatment would significantly impact patient selection. In this study, classification and regression tree (CART) methods were used to model tumor response to preoperative conformal high-dose rate brachytherapy by assessing the predictive value of vascular endothelial growth factor (VEGF), Bcl-2, p21, p53, and APAF-1. Experimental Design: Immunohistochemistry was used to detect VEGF, Bcl-2, p21, p53, and APAF-1 from 62 pretreatment rectal tumor biopsies. Scores were assigned as percentages of positive tumor cell staining and were used in CART analysis to identify the proteins that best predicted response to radiotherapy. Ten-fold cross-validation was used to prevent overfitting and multiple cross-validation experiments were run to estimate the prediction error. Results: Postoperative pathologic evaluation of the irradiated tumor bed revealed 43 responsive tumors [20 with complete response (T 0 ) and 23 with partial response] and 19 nonresponsive tumors. The optimal tree resulting from CART analysis had five terminal nodes with a misclassification rate of 18%. Of the five proteins selected for their predictive value, VEGF and Bcl-2 contributed most to the classification of responsive and nonresponsive tumors. All 10 tumors with no VEGF were completely responsive (T 0 ) to radiotherapy; 85% of those with VEGF and negative for Bcl-2 were responsive to therapy. Conclusions: VEGF and Bcl-2 status in pretreatment rectal tumor biopsies may be predictive of response to preoperative high-dose rate brachytherapy.