Study Design and Patient and Tumor Characteristics
Tumor samples from 100 patients with disease recurrence were obtained retrospectively and matched to 200 patients without disease recurrence. Because the predefined cutpoint of H/I was previously tested with samples associated with patient outcome defined as local, regional, or distant metastasis, patients with contralateral and unknown recurrences were removed. Therefore our final cohort consisted of 83 patients with local, regional, or distant recurrence matched with 166 patients without recurrence (Figure 1B). A post hoc power analysis indicated that with this sample size, the power to detect an odds ratio of 3 for comparing risk of late recurrence without extended letrozole vs that with letrozole in the H/I-high group was 82% at a 5% significance level. Characteristics of the case-control population can be seen in Table 1. The majority of patients was aged greater than 50 years with T1/T2 stage disease, and a larger proportion was node-positive and did not receive adjuvant chemotherapy. Within unmatched variables, there was a statistically significant difference between the number of case subjects and control subjects for letrozole vs placebo and high vs low H/I groups (Table 1). Case subjects and control subjects are similar to the overall MA.17 study population with respect to key clinical parameters with the exception of nodal status and radiation therapy (SupplementaryTable 1, available online).
The observed clinical benefit from letrozole within different patient subgroups was similar between the case-control study and the entire study population reported in the MA.17 trial (Table 2). The benefit was statistically significant in both studies in all subgroups except within the lymph node–positive subgroup for the case-control, which may be because of small sample size for the relatively smaller size of the benefit (vs lymph node negative) (Table 2). For example, the odds ratio estimated by conditional logistic regression with an interaction between clinical factors and the treatment in this case-control study and the hazard ratio estimated from the overall MA.17 trial as in published literature were as follows: 0.20 and 0.45 for node negative, 0.63 and 0.61 for node positive, 0.44 and 0.73 for aged greater or equal to 50 years, 0.42 and 0.58 for hormone receptor positive, and 0.45 and 0.49 for ER+/PR+ (Table 2).
Prognostication of Late Disease Recurrence by H/I
In the placebo arm, H/I status was statistically significantly associated with prognosis in the unadjusted model, as women in the high H/I group experienced substantially worse prognosis than those in the low H/I group (OR = 2.24; 95% CI = 1.09 to 4.61; P = .03). H/I status remained marginally statistically significant in the adjusted model (OR = 2.15; 95% CI = 1.00 to 4.64; P = .05) after adjusting for all clinicopathological factors examined (age, tumor size, tumor grade, nodal status, ER, PR, HER2) (Supplementary Table 2). In the letrozole arm, however, H/I was not prognostic for late recurrence in neither the unadjusted nor the adjusted models (P = .72 and .63, respectively).
Predictive Performance of H/I for Benefit From Letrozole Treatment
In the unadjusted analysis, high H/I group was statistically significantly associated with patient benefit from letrozole (OR = 0.35; 95% CI = 0.16 to 0.75; P = .007) (Table 3). In the adjusted model, which included all clinicopathological factors as covariates, high H/I remained statistically significantly associated with patient benefit from letrozole (OR = 0.33; 95% CI = 0.15 to 0.73; P = .006) (Table 3; Figure 2), which represented a 67% reduction in the risk of recurrence with extended letrozole treatment as compared with placebo.
Forest plots showing odds ratio (OR) for recurrence associated with clinico-pathological factors and treatment effect for each of the HOXB13/IL17BR (H/I) groups. The odds ratio for recurrence was calculated from conditional logistic regression. Red color indicates H/I-high patients; black color indicates H/I-low patients. All statistical tests were two-sided. CI = confidence interval; ER = estrogen receptor; PR = progesterone receptor.
The absolute RFS at 4 years estimated from this case-control study was 86.6% (95% CI = 80.6% to 90.9%) and 93.4% (95% CI = 90.2% to 95.6%) for patients receiving placebo and letrozole, respectively. This is consistent with the 4-year disease-free survival (87% and 93% for placebo and letrozole, respectively) estimated from the parental MA.17 trial. For low H/I patients, the 5-year RFS was 87% (95% CI = 76.8% to 92.9%) and 91% (95% CI = 83.1% to 95.3%) in the placebo and letrozole group, respectively, showing a non-statistically significant reduction in the risk of recurrence of 4% (P = .35). However, for high H/I patients, the 5-year RFS was 73% (95% CI = 56.6% to 84.1%) and 89.5% (95% CI = 80.3% to 94.5%) in the placebo and letrozole group, demonstrating a 16.5% reduction in the absolute risk of recurrence at 5 years (P = .007) (Table 4).
To examine whether continuous ER and PR measurements would affect the predictive performance of H/I, gene expression values of ER and PR by RT-PCR were used in the adjusted model instead of binary immunohistochemistry ER and PR values. High H/I remained statistically significant with extended letrozole benefit (OR = 0.32; 95% CI = 0.14 to 0.72; P = .006), whereas low H/I still showed no statistically significant benefit (OR = 0.59; 95% CI = 0.25 to 1.39; P = .23).
Linearity testing with a conditional logistic regression model indicated that there was a significant nonlinear relationship of H/I index as a continuous variable with recurrence (P = .02); therefore H/I index was transformed by restricted cubic spline when assessing the statistical significance of the interaction between the treatment and continuous H/I index. Likelihood ratio test showed that the interaction between H/I and letrozole therapy was statistically significant (P = 0.03), adjusting for clinicopathological factors.
Predictive Performance of H/I Within Clinically Relevant Subgroups
Unplanned analyses were conducted to determine whether patient subgroups were differentially associated with the predictive performance of H/I by including a conditional logistic regression three-way interaction term between the clinical variable of interest, treatment, and H/I groups. High H/I was statistically significantly predictive of letrozole benefit in our overall study population. Analyses of subsets including nodal, HER2, ER, and PR status are shown in Supplementary Table 3 and Supplementary Figure 1 (available online).