TH‐AB‐304‐02: Fitting Grade>=2(2+) Acute Rectal Complication Rates in Prostate Cancer Patients to Lyman Kutcher Burman (LKB) and Logistic Regression NTCP Models Using Dosimetry and Patient Specific Characteristics

Abstract

Purpose:Models of rectal toxicity which include dosimetry only are known to have a relatively low predictive power, at least as measured by the area under the ROC curve (AUC). It has been suggested that the predictive power of models can be improved by including non‐dosimetric patient specific characteristics.Methods:We compiled a database of 79 prostate patients who were treated with an IMRT technique to a dose of 77.4 Gy, 1.8Gy/fx, with an integrated boost to 81–83Gy in a sub‐volume of a prostate which was identified on a pre‐treatment MRI study. Acute grade 2+ rectal toxicities were graded according to CTCAE v4 by a physician who retrospectively reviewed patient's medical records. We modified the LKB model to include one patient specific variable at a time, and we also used an NTCP model based on logistic regression to perform multi‐variate analysis. We used patient specific variables available to us in a retrospective study: age, diabetes, hormonal treatment, Gleason Score, PSA, Statin use, prostate volume, boost volume and rectal volume.Results:Grade 2+ acute rectal toxicity occurred in 20% (16/79 patients). The LKB model with dosimetry alone gives AUC=0.65. Four variables, age, diabetes, PSA, Statin use, increase the AUC in LKB model to a maximum of 0.79. The same four variables in the logistic regression model increase AUC to 0.87. The most significant correlations are with PSA and with Statin use.Conclusion:Including patient specific variables in toxicity models can significantly increase apparent predictive power of a model. Somewhat surprising finding of a strong correlation between rectal toxicity and PSA in our dataset suggests that conclusions from each individual study should be treated with caution, until independently confirmed. Larger databases from prospective studies or meta‐analysis of multiple studies may be needed to find patient characteristics that are truly predictive.