Association of Genetic Testing Results With Mortality Among Women With Breast Cancer or Ovarian Cancer

Allison W. Kurian, MD, MSc; Paul Abrahamse, MA; Irina Bondarenko, MS; Ann S. Hamilton, PhD; Dennis Deapen, DrPH; Scarlett L. Gomez, PhD; Monica Morrow, MD; Jonathan S. Berek, MD, MMSc; Timothy P. Hofer, MD, MSc; Steven J. Katz, MD, MPH; Kevin C. Ward, PhD, MPH

Disclosures

J Natl Cancer Inst. 2022;114(2):245-253. 

In This Article

Methods

Study Cohort and Dataset

All women diagnosed with breast cancer or ovarian cancer from January 1, 2013, to December 31, 2017, in California and Georgia and reported to SEER registries in California (the Los Angeles Cancer Surveillance Program, the Greater Bay Area Cancer Registry, and the Cancer Registry of Greater California) and in Georgia (the Georgia Cancer Registry) were linked to clinical germline genetic testing results from 4 laboratories (Ambry Genetics, Aliso Viejo, CA; GeneDx, Gaithersburg, MD; Invitae, San Francisco, CA; Myriad Genetics, Salt Lake City, UT) that performed the substantial majority of such testing as determined by genetic counselor and patient surveys.[3,4,6] Probabilistic methods were used to optimize ascertainment and linkage accuracy, as previously reported.[3,6] The analytic dataset combined genetic results from the 4 laboratories, from reports dated 2012 through the first quarter of 2019, with SEER variables.

Patients were included in the analytic cohort if they linked to a genetic result, had stages I-IV breast cancer or epithelial ovarian cancer, and received chemotherapy. Exclusion criteria included those ages younger than 20 years, more than 1 primary tumor, and diagnosed only on death certificate. Missingness was less than 5% for all variables except grade for ovarian cancer (20.6% missing). All observations with missing values were excluded except for ovarian cancer missing grade. Patients with nonepithelial ovarian cancer (eg, germ cell, sarcoma, and other histologies) were excluded because of their different epidemiology, genetics, and clinical course (Supplementary Figure 1, available online). The analytic file included both registry and laboratory information and was stripped of protected health information [as defined by the Health Information Portability and Accountability Act Privacy Rule].[26] The study was approved by institutional review boards associated with the SEER registries.

Test Results From Laboratories

Germline genetic testing results were provided by laboratories at the level of the affected gene and consisted of the interpretation according to American College of Medical Genetics criteria that was returned to the ordering clinician: PV or likely PV (analyzed together as PV), variant of uncertain significance (VUS), and benign or likely benign (analyzed together as negative). Results from all laboratories were combined to ensure anonymity, and gene-specific results were analyzed only for those genes tested by 2 or more laboratories (n = 86).

Measures

Demographic and clinical measures were selected that were conceptually appropriate based on previously demonstrated relationships to cancer-related mortality, including social determinants of health (eg, race and ethnicity, poverty), tumor biologic features (eg, grade, subtype), and treatments (Table 1 and Table 2). SEER registries provided diagnosis age, race and ethnicity (non-Hispanic White, Black, Asian, Native American and Alaskan Native, Hispanic), percent poverty at the census tract level (<10%, 10%-19%, ≥20%), marital status, tumor stage and grade, breast cancer subtype defined by expression of estrogen and/or progesterone receptors (ER/PR) and HER2 [ER/PR-positive, HER2-negative; HER2-positive with any ER/PR status, defined hereafter as HER2-positive; and ER/PR-negative and HER2-negative, defined hereafter as triple-negative breast cancer (TNBC)], and ovarian cancer histology (serous, mucinous, endometrioid, clear cell, or other adenocarcinoma). SEER registries provided information on breast cancer first-course treatment including surgery (breast-conserving surgery, unilateral or bilateral mastectomy), chemotherapy, radiotherapy, endocrine therapy, or HER2-directed therapy. First-course ovarian cancer treatment information included type of surgery, specifically debulking surgery; other surgery (SEER codes 17, 25–28, 35–37, 50–52, 55–57); or no surgery and radiotherapy. SEER registries provided date and cause of death. Overall, cancer-specific and other-cause mortality data were available through December 31, 2019, and patients alive then were coded as censored. Patients who died of other cancers or noncancer were coded as censored at date of death.

Statistical Analysis

Our question was whether PVs were associated with risk beyond that accounted for by known risk factors: thus, the base model included known correlates of breast and ovarian cancer-specific mortality (described in Measures). As covariates were selected based on known mortality associations, we did not refine the model further by excluding covariates based on P values or effect size. Genetic test results were then added, with the primary result being the magnitude and precision of resulting coefficients.

Separate models were specified for patients with each breast cancer subtype and ovarian cancer, because treatments and relationships of predictor variables to outcomes likely differ between these groups. We used multivariable Cox proportional hazard survival models to examine the association between genetic results, demographic and clinical factors with breast, and ovarian cancer-specific mortality. Competing risks of noncancer deaths were treated as censored. Date of chemotherapy initiation was used as the starting point for survival, and treatments that occurred after chemotherapy were coded as time-varying covariates to account for immortal time bias. Ovarian cancer grade was imputed using multiple imputation techniques.

Sensitivity Analysis

Proportional hazards assumptions were tested by including time-dependent covariates of all independent variables and testing for significance. All interactions between key covariates were tested. To assess generalizability to the nontested population, we examined a model with weights for test receipt. Weights were generated from a logistic regression model of genetic testing receipt across all patients (tested and not tested), using clinical and demographic measures as covariates. The inverse of the predicted probabilities of test receipt were used as weights. We examined respecification of competing mortality risks using a Fine and Gray analysis. To address potential effects of test timing and treatment selection, we excluded patients tested after treatment initiation. To account for potential error in reported cause of death, we evaluated overall rather than cancer-specific mortality.

processing....