Including the Ductal Carcinoma-In-Situ (DCIS) Score in the Development of a Multivariable Prediction Model for Recurrence After Excision of DCIS

Abstract

IntroductionIndividual prediction of local recurrence (LR) risk after breast-conserving surgery (BCS) for ductal carcinoma-in-situ (DCIS) is needed to identify women at low risk, for whom radiotherapy may be omitted. Patients and MethodsThree predictive models of LR—clinicopathologic factors (CPF) alone; CPF + estrogen receptor (ER) + human epidermal growth factor receptor 2 (HER2); and CPF + DCIS score (DS)—were developed among 1102 cases of DCIS in patients with complete covariate and outcome data. Categorizations of discrete variables and transformations of continuous variables were examined in Cox models; 2-way interactions and interactions with time were assessed. Internal validation was performed by bootstrapping. Individual predicted 10-year LR risks were computed from covariate values, estimated regression parameters, and estimated baseline survival function. Accuracy was assessed by c statistics and calibration plots. ResultsThe strongest prediction model incorporated CPF + DS. The c statistics for CPF + DS, CPF + ER + HER2, or CPF-alone models were 0.7025, 0.6879, and 0.6825, respectively. The CPF + DS model was better calibrated at predicting low (≤ 10%) individual 10-year LR risks after BCS alone than those incorporating CPF + ER + HER2 or CPF alone, evidenced by c statistics and plots of observed by predicted risks. Among women aged ≥ 50 with no adverse CPF, the CPF + DS model identified the greatest proportion of women (62.3%) with predicted individual 10-year LR ≤ 10% without radiotherapy compared to the CPF + ER + HER2 (50.9%) or CPF alone (46.5%) models. ConclusionIndividual prediction of LR incorporating DS is more accurate and identifies a higher proportion of women with low predicted risk of LR after BCS alone, for whom radiotherapy may be omitted.