External Validation of Medicare Claims Codes for Digital Mammography and Computer-Aided Detection

More than one half of incident breast cancers occur among Medicare-enrolled women (1), and Medicare enrollees receive approximately one third of all screening mammograms nationwide (∼13 million annual mammograms; refs. 2, 3). Thus, improving the quality of screening mammography received by Medicare enrollees remains a public health priority. In 2001, Congress extended Medicare coverage to digital mammography and the application of computer-aided detection (CAD) during screening and diagnostic mammography. Because Medicare pays supplemental fees for digital mammography and CAD, evaluation of the clinical and economic impact of the dissemination of these technologies within the Medicare population is needed. Although Medicare claims could be a fruitful data source for such evaluation, uncertainty remains about validity of the claims procedure codes signifying use of these technologies.

An internal validation study found substantial agreement for Medicare claims-based classification of mammograms as digital versus film and CAD versus non-CAD in the Carrier Claims and Outpatient Medicare claims files (4). Because accurate data must be internally consistent, the findings are consistent with accurate coding, but validation using an external reference standard is necessary for a high degree of confidence in data validity. The internal validation study was also conducted with claims from 2001 to 2003 and validation in more recent years is desirable. In addition, dual reimbursement for both digital mammography and CAD was not allowed by Medicare until 2004, so the validity of CAD codes on digital mammograms was not assessed.

We therefore capitalized on a recently developed data infrastructure—the linked Breast Cancer Surveillance Consortium (BCSC)-Medicare data—to externally validate more recent Medicare claims data for digital mammography and CAD against a BCSC-derived reference standard.

We used data from the Carrier Claims and Outpatient Medicare claims files, which were linked with BCSC mammography data derived from 4 regional mammography registries (Carolina Mammography Registry, New Hampshire Mammography Network, San Francisco Mammography Registry, and Vermont Breast Cancer Surveillance System). Data for this study were obtained from the BCSC Data Resource (5). The BCSC includes both community-based and academic mammography facilities that serve a geographically and ethnically diverse patient population. BCSC facilities transmit prospectively collected patient and mammography data to regional registries, which link the data to breast cancer outcomes ascertained from regional or Surveillance, Epidemiology, and End Results cancer registries.

The BCSC has established standard definitions for key variables and multiple levels of data quality control and monitoring (6). BCSC sites have received Institutional Review Board (IRB) approval for active or passive consenting processes or a waiver of consent to enroll participants, link data, and conduct research. All procedures are Health Insurance Portability and Accountability Act compliant, and BCSC sites have received a Federal Certificate of Confidentiality to protect the identities of patients, physicians, and facilities.

Claims and mammography data were matched using a deterministic algorithm derived by the National Center for Health Statistics that is based on social security numbers (when available), names, and birth dates. Among women aged 65 years and older with fee-for-service Medicare with a BCSC mammogram between 1998 and 2006, 82.8% had a matching Medicare mammography claim within 7 days (99.1% of which matched on the same date).

We identified a matched sample of bilateral mammograms captured in both Medicare claims and the BCSC among women who were aged 65 or older on mammography dates from January 1, 2003, to December 31, 2006. We identified bilateral mammograms based upon Medicare claims with Healthcare Common Procedural Coding System (HCPCS) codes 76091, 76092, G0202–G0205 (encompassing film screen and digital screening and diagnostic mammograms) and considered mammograms to have matching BCSC records if claims and BCSC records had the same date of service. We did not assess claims before 2003, because dual reimbursement for digital mammography and CAD was not allowed, leading to potential underascertainment of CAD codes on digital mammograms before 2003. We excluded mammograms with missing or uncertain BCSC data regarding film versus digital mammography, and for validation of CAD codes, missing or uncertain BCSC data on CAD status.

During the study period, BCSC facilities have reported the use of film versus digital mammography at the level of the individual mammogram. While most BCSC registries report CAD use at the mammogram level, one reported CAD use at the mammography facility level. To maximize the accuracy of the reference standard, we only included mammograms with mammogram level BCSC data on CAD use. We used these BCSC data to classify each mammogram as either film or digital and CAD versus non-CAD.

We used HCPCS codes on claims to classify mammograms as digital versus film and CAD versus non-CAD. Mammograms with codes G0202, G0204 were classified as digital and otherwise as film, whereas mammograms with codes 76082, 76083, 76085, G0203, G0205, and G0236 were classified as CAD and otherwise as non-CAD.

We quantified accuracy by computing: (i) the proportion of reference standard film, digital, non-CAD, and CAD mammograms that were correctly classified by claims as such; (ii) the predictive value (or the proportion of mammograms classified by claims codes as film, digital, CAD, or non-CAD that were likewise classified by the reference standard); and (iii) Cohen's kappa. We also evaluated accuracy over time by stratifying analyses by year, screening versus diagnostic/other purpose (according to the BCSC data), and for the CAD analysis, by film versus digital. All 95% confidence intervals around point estimates were negligibly small, so we report only point estimates. We conducted statistical analyses using R, version 2.12.0 (R Foundation for Statistical Computing). The study was approved by the IRBs at University of California, Davis (Sacramento, CA) and Group Health Research Institute (Seattle, WA).

We identified a sample of 253,727 mammograms with matched Medicare claims and BCSC records and complete digital versus film mammography data. The mammograms were obtained from 120,709 women who received an average of 2.1 mammograms during the study period (range, 1–7). On the date of mammography, women had a mean age of 73.5 years (SD, 6.1). Approximately 78% of the women were white, non-Hispanic; 8% were black, non-Hispanic; 4% were Asian; 3% other races; and 7% had missing race/ethnicity data. Of all mammograms, 87% were designated by the BCSC as having a screening purpose whereas the remainder had a diagnostic/other purpose. From this sample, 186,417 (73.5%) mammograms had BCSC data on CAD use for analyses of CAD coding.

In validation analyses of digital mammography classification, 97.2% of digital and 97.3% of film mammograms were correctly classified by claims and the κ statistic (0.90) indicated excellent agreement beyond chance (Table 1). A high proportion of non-CAD and CAD mammograms were also correctly classified by claims (86.7% and 96.6%, respectively) and claims-based classification of CAD use also had high agreement beyond chance (κ = 0.83). The predictive values of a digital or CAD classification (87.0% and 86.3%, respectively) were lower than the predictive values of a film or non-CAD designation (99.4% and 96.7%, respectively), because of the lower sample prevalences of digital mammography and CAD.

While correct classification of film and digital mammograms were similarly high across the study period, predictive value for digital and κ increased over the study period as the sample prevalence of digital mammography rose (Table 2). Because of relatively low prevalence of digital mammography in 2003 and 2004 (<15% in each year), predictive values during these years were each below 80%. Despite a sample prevalence of CAD of 27.8% in 2003, the predictive value of CAD designation in that year was 75.8%. As CAD prevalence increased from 2004 to 2006, the predictive value of a CAD designation exceeded 85% in each year.

The accuracy of claims-based classifications for digital mammography and CAD were similarly high in mammograms conducted for screening as compared with diagnostic/other purposes (data not shown). However, the predictive value of a CAD designation was lower among film mammograms (83.5%) than among digital mammograms (94.1%).

Using an external reference standard derived from high-quality mammography registry data, we validated Medicare claims codes for digital mammography and CAD. Analyses showed that Medicare claims codes can distinguish digital from film mammograms and mammograms conducted with and without CAD with a high overall accuracy from 2003 to 2006.

The predictive values of claims codes for digital mammography were below 80% in 2003 and 2004, when digital mammography prevalence was below 15%. It is possible that coding errors may have been more common early during digital mammography dissemination when some facilities may have used both film and digital machines simultaneously. While investigators should recognize the potential for greater misclassification during this time period, the predictive value of digital designation was 92.8% in 2005 when the sample prevalence of digital mammography was 19.0%, and digital mammography has since disseminated broadly in the United States. From 2007 to 2012, the proportion of Mammography Quality Standards Act certified mammography facilities with digital machines increased from 29% to 82% (7). Thus, the predictive value of a claims-based digital designation is likely to be very high during this period.

CAD is also more widely used now than during the study period. In 2008, CAD was used on approximately three quarters of Medicare mammograms (8). Thus, the high predictive values of a claims-based CAD designation (>80% from 2004–2006) are likely to persist in later claims. Nevertheless, the potential for misclassification of CAD status may be greater in claims before 2004. In addition, across the study period, the claims-based designation of CAD use was somewhat nonspecific; more than 13% of mammograms classified as non-CAD by the reference standard were classified as CAD by claims codes, leading to an overall predictive value of a CAD designation of 86.3%.

Strengths of this study include the linkage of high-quality mammography data to large samples of Medicare mammography claims, enabling well-powered external validation of the claims codes against a credible reference standard. Although study data derive from 2003 to 2006, study results likely generalize to more recent claims owing to the increasing prevalence of digital mammography and CAD in subsequent years.

Study data derive from Medicare enrollees receiving mammography within regional registries in 4 U.S. states, and results may not generalize to other U.S. regions or to mammography claims data for non-Medicare populations (e.g., private insurance plans). Although the BCSC data provide a credible reference standard, inaccuracies in BCSC classification of mammograms by digital or CAD status may have introduced error. Although a high percentage of BCSC women enrolled in fee-for-service Medicare could be matched to mammography claims (82.8%), the accuracy of mammography claims coding could be lower among women who could not be matched.

Medicare claims data are a potentially powerful resource for mammography research or quality monitoring. This study suggests that Medicare claims codes for digital mammography and CAD use are sufficient for accurate classification of mammograms by the use of these modalities, particularly in the context of broad uptake of each technology after 2005.

R. Hubbard has a commercial research grant for GE Healthcare. No potential conflicts of interest were disclosed by other authors. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute or the NIH.

Conception and design: J.J. Fenton, R. Smith-Bindman, R. Hubbard

Development of methodology: J.J. Fenton, R. Smith-Bindman, K. Lindfors

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): J.J. Fenton, W. Zhu, R. Smith-Bindman, K. Lindfors, R. Hubbard

Writing, review, and/or revision of the manuscript: J.J. Fenton, W. Zhu, R. Smith-Bindman, K. Lindfors, R. Hubbard

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J.J. Fenton, S. Balch

Study supervision: J.J. Fenton, R. Hubbard

The authors thank the participating women, mammography facilities, and radiologists for the data they have provided for this study. A list of the BCSC investigators and procedures for requesting BCSC data for research purposes are provided at http://breastscreening.cancer.gov/.

This work was supported by the National Cancer Institute–funded Breast Cancer Surveillance Consortium (U01CA63740, U01CA86076, U01CA86082, U01CA63736, U01CA70013, U01CA69976, U01CA63731, U01CA70040, HHSN261201100031C) and the National Cancer Institute grant R21CA158510.