Cancer Registry Data Linkage of Electronic Health Record Data From ASCO's CancerLinQ: Evaluation of Advantages, Limitations, and Lessons Learned

PURPOSE

To evaluate the completeness of information for research and quality assessment through a linkage between cancer registry data and electronic health record (EHR) data refined by ASCO's health technology platform CancerLinQ.

METHODS

A probabilistic data linkage between Iowa Cancer Registry (ICR) and an Iowa oncology clinic through CancerLinQ data was conducted for cases diagnosed between 2009 and 2018. Demographic, cancer, and treatment variables were compared between data sources for the same patients, all of whom were diagnosed with one primary cancer. Treatment data and compliance with quality measures were compared among those with breast or prostate cancer; SEER-Medicare data served as a comparison. Variables captured only in CancerLinQ data (smoking, pain, and height/weight) were evaluated for completeness.

RESULTS

There were 6,175 patients whose data were linked between ICR and CancerLinQ data sets. Of those, 4,291 (70%) were diagnosed with one primary cancer and were included in analyses. Demographic variables were comparable between data sets. Proportions of people receiving hormone therapy (30% v 26%, P < .0001) or immunotherapy (22% v 12%, P < .0001) were significantly higher in CancerLinQ data compared with ICR data. ICR data contained more complete TNM stage, human epidermal growth factor receptor 2 testing, and Gleason score information. Compliance with quality measures was generally highest in SEER-Medicare data followed by the combined ICR-CancerLinQ data. CancerLinQ data contained smoking, pain, and height/weight information within one month of diagnosis for 88%, 52%, and 76% of patients, respectively.

CONCLUSION

Linking CancerLinQ EHR data with cancer registry data led to more complete data for each source respectively, as registry data provides definitive diagnosis and more complete stage information and laboratory results, whereas EHR data provide more detailed treatment data and additional variables not captured by registries.

INTRODUCTION

The National Cancer Institute's (NCI) SEER program is one of the premier cancer surveillance programs in the world. SEER registries collect data on patient demographics, primary tumor site/characteristics, stage at diagnosis, first course treatment, and follow-up for vital status. Goals of the SEER program include collecting complete and accurate data on cancers diagnosed among residents of geographic areas covered by SEER registries, conducting continual quality control and quality improvement studies to ensure high-quality data and providing research/data resources to the cancer prevention and control research community. The Iowa Cancer Registry (ICR) has been a member of the SEER Program since 1973.

CONTEXT

• Key Objective

• To determine if and how a linkage between the electronic health record information captured in CancerLinQ data with cancer information collected by the Iowa Cancer Registry could be used to improve the completeness and quality of the variables in either data source, and complement each other in providing a more complete picture of care for patients with cancer.

• Knowledge Generated

• Demographic data were comparable between data sets, whereas the proportion of patients receiving hormone therapy and immunotherapy data was higher in CancerLinQ data compared with cancer registry data. The Registry had more complete stage and laboratory testing information, which are important in assessing cancer-related quality measures. CancerLinQ data contained smoking, pain, and height/weight information for the majority of patients; these variables are not captured by the Registry.

• Relevance

• Linking electronic health record data with cancer registry data provides a more complete picture of a patient's cancer experience.

CancerLinQ is a health technology platform, developed and implemented by ASCO, to improve quality of cancer care and advance research. CancerLinQ aggregates data from electronic health records (EHRs) of participating oncology practices and cancer centers in the United States, harmonizes and standardizes it, and delivers it back to contributing centers through the embedded SmartLinQ application as clinical quality measure scores, dashboards, and reports. Following deidentification, research data sets are made available through the CancerLinQ Discovery program.¹ Some of the SmartLinQ quality measures are derived from ASCO's foundational Quality Oncology Practice Initiative (QOPI), which assesses nationally recognized quality measures among participating oncology practices to foster a culture of self-examination and improvement in cancer care.²

Leadership within SEER and CancerLinQ aimed to determine how data captured by SEER registries and EHR data compared (ie, could information captured in both data sources be used to improve the completeness and quality of the variables in either data source, and how might those improved variables affect assessment of quality measures?) and could complement each other in providing a more complete picture of care for patients with cancer (ie, could new variables be provided to users of either data source as a result of linking the data?). A data exchange was established to link ICR data with EHR data from an Iowa oncology practice participating in CancerLinQ. Primary objectives of the resulting analyses included the following: (1) evaluation and comparison of variables captured by both data sources; (2) identification of key variables within each data source that could supplement the other source; and (3) assessment of variation in compliance with selected breast and prostate cancer QOPI measures by data source. To serve as a comprehensive comparison for treatment and quality measures, SEER data linked to Medicare claims was used among patients with cancer enrolled in Medicare Fee-for-Service.³

METHODS

Data Sources and Study Population

In 2018, the ICR and CancerLinQ entered into a Cancer Surveillance Data Use Agreement. One of two community-based Iowa oncology clinics participating in CancerLinQ at the time of this study agreed to participate; permission was obtained to include their data in this analysis. Information Management Services (IMS), a contractor of the NCI SEER Program, served as the honest broker and conducted the linkage at a person level for cases diagnosed between 2009 and 2018 (according to ICR data). IMS conducted a probabilistic linkage using Match*Pro.⁴ Details on the linkage between the ICR and CancerLinQ data sets are provided in the Data Supplement.

The linked ICR-CancerLinQ data set was provided to the ICR for analyses. A retrospective secondary data analysis of the linked cohort was conducted. This study protocol was approved by the University of Iowa Institutional Review Board. We complied with the reporting recommendations of the Statement on the Reporting of Evaluation studies in Health Informatics (STARE-HI).

ICR data are collected by trained registrars following established abstracting rules and standardized manuals (eg, Facility Oncology Registry Standards, International Classification of Disease for Oncology [ICD-O] and American Joint Committee on Cancer [AJCC] Staging Manual). They use all available pathologic and clinical information to document date of diagnosis, cancer site, ICD-O-3 behavior (in situ v malignant), histology, stage, cancer site–specific data items (eg, human epidermal growth factor receptor 2 [HER2] status, prostate-specific antigen [PSA], and Gleason score values), first course treatment, and demographics. Data are consolidated across providers and structured as one record per cancer (ie, persons with multiple cancers have multiple records). The ICR only collects data on those who were residents of Iowa at the time of their cancer diagnosis.

CancerLinQ data were derived from EHR records containing the following structured data elements: International Classification of Diseases, Ninth (ICD-9) and Tenth (ICD-10) Revision diagnosis codes, representing the reason for clinic visit, cancer behavior, dates of service, cancer organ system, and anatomic site; SNOMED CT⁵ codes were used to identify diagnosis type, stage, and demographic data; RxNorm codes were used for medications ordered and administered; and the Logical Observation Identifiers, Names and Codes common terminology was used to standardize laboratory terms, smoking status, and pain assessments.

ICR data are routinely linked to Medicare claims data as part of the SEER-Medicare linkage conducted by the NCI and the Centers for Medicare and Medicaid Services (further details are provided in the Data Supplement). Patients with breast and prostate cancer in the ICR-CancerLinQ data set older than age 65 years were linked to the Iowa SEER-Medicare database by a common identifier. Those enrolled in Medicare Parts A, B, and D without Health Maintenance Organization (HMO) enrollment for at least 12 months after diagnosis were included in a comparison of treatment type (ie, chemotherapy, immunotherapy, and hormone therapy) among SEER-Medicare, ICR, CancerLinQ, and ICR plus CancerLinQ data sources.

Analysis

Person-level demographic information was compared between ICR and CancerLinQ linked data and to all cancer cases in the ICR to determine representativeness. Whereas the ICR has a single, static diagnosis date for each cancer, diagnosis dates in CancerLinQ data varied by visit; we therefore considered the earliest visit date with a cancer diagnosis code to be the diagnosis date. Similarly, the ICR contains a single, static ICD-O-3 site code, whereas ICD-9/ICD-10 codes used to categorize cancer site in CancerLinQ data could vary among visits depending on provider documentation. ICD-9/ICD-10 and ICD-O-3 codes were mapped to organ system for comparisons. Because diagnosis codes in CancerLinQ could vary between visits, we restricted analyses to a subset of patients with one primary cancer, according to ICR, to ensure linked records were for the same cancer. Additionally, the ICR has a single, static set of stage variables, whereas CancerLinQ stage could be recorded differently across visits depending on provider documentation. The highest stage documented in CancerLinQ within 6 months of the ICR-defined diagnosis date was used to compare with the ICR stage.

Treatment regimen was examined in people diagnosed in 2013 or later because SEER modified its categories in 2013 to reclassify certain biologic/immunotherapy medications more appropriately as biologic response modifiers; previously, these medications were categorized as chemotherapy. Analyses of treatment were limited to 2013+ to ensure definitions were consistent and comparable. RxNorm codes in CancerLinQ data were merged with the Cancer Research Network Chemotherapy Lookup Table,⁶ and drugs were categorized according to the SEER*Rx Interactive Antineoplastic Drugs Database⁷ as chemotherapy, hormone, or biologic/immunotherapy. Descriptive analyses of variables unique to the CancerLinQ data, including height/weight, smoking status, and pain assessments, were performed to assess completeness and timing.

Evaluations of three quality measures derived from ASCO's QOPI program or related QOPI Reporting Registry⁸ (detailed in the Data Supplement) were performed on patients with eligible primary breast-only or primary prostate-only cancers: (1) HER2(–) treatment measure: proportion of patients with HER2-negative or undocumented breast cancer who are spared treatment with HER2-targeted therapies⁹; (2) HER2(+) trastuzumab measure: proportion of patients with AJCC Stage I (T1c)—III and HER2-positive breast cancer receiving adjuvant chemotherapy who received trastuzumab¹⁰; and (3) high-risk prostate measure: proportion of high-risk or very high-risk (ie, T3a or T3b-T4 disease, PSA > 20, or Gleason score of 8+) prostate cancer patients who received adjuvant hormonal therapy combined with radiation therapy.¹¹ A subset of patients enrolled in Medicare Parts A and B without HMO enrollment for at least 12 months after diagnosis were included in SEER-Medicare data, which were linked to ICR data using a common identifier, for a comparison with the ICR and CancerLinQ data sources. Treatments in SEER-Medicare were identified using Current Procedural Terminology and ICD-9/-10 codes (Data Supplement). Compliance was calculated for each quality measure by dividing the number of patients who were concordant with the numerator criteria (ie, receipt/nonreceipt of the recommended treatment) by the total number of patients who met the denominator criteria.

Descriptive analyses were conducted to compare data sources; McNemar's chi-square tests of agreement for paired data were used to examine disagreement among the variables included in both CancerLinQ and ICR data sets. Statistical analyses were conducted using SAS version 9.4 (SAS Institute, Cary, NC).

RESULTS

There were 6,175 unique Iowa residents diagnosed with cancer between 2009 and 2018 (according to ICR data) who visited the CancerLinQ oncology practice, representing approximately 4% of the total Iowa cancer population. Seventy percent (n = 4,291) were diagnosed with one cancer; 30% had multiple cancers and were excluded from subsequent analyses. There was close agreement in overall organ system of the cancer; 70 (2%) people had discordant organ systems between data sources. Demographic variables were largely comparable between sources (Table 1). Race was more often unknown/missing in CancerLinQ data (2% v < 1%), whereas ethnicity data were more often unknown/missing in ICR data (6% v 3%). Median age was 68 years in both sources, and the population was predominantly White (97%-98%), non-Hispanic (92%-96%), and married (60%-64%). These characteristics were similar to the overall Iowa cancer population, except that those in the linked data set were more often female (58% v 51%) and had a higher proportion of breast cancer (27% v 15%).

TABLE 1.

Comparison of Demographic Characteristics in CancerLinQ and ICR Data Sources

Tumor and Treatment Characteristics

A comparison of cancer characteristics and treatments for all patients with one primary cancer is displayed in Table 2. Seventy-two percent did not have stage information captured in discrete variables provided by CancerLinQ compared with 15% in ICR data. The percent who received chemotherapy was the same (49%) in CancerLinQ and ICR data. The percent who received hormone therapy (30% v 26%, < .0001) and immunotherapy (22% v 12%, < .0001) was significantly greater in CancerLinQ data.

TABLE 2.

Comparison of Cancer Characteristics and Treatment Between CancerLinQ and ICR Data Sources for Those With One Cancer Only, and for Those With Breast or Prostate Cancers (one cancer only)

Cancer and treatment characteristics were compared specifically for breast and prostate cancers (Table 2) since they were the focus of the quality measure assessments. For patients with breast cancer, chemotherapy was comparable between data sources (30%), but rates of hormone therapy (76% v 64%, P < .0001) and immunotherapy (18% v 9%, P < .0001) were recorded more frequently in CancerLinQ data compared with ICR data. For prostate cancer, the percentage receiving chemotherapy (19% v 14%, P = .0143) and immunotherapy (20% v < 5%, P < .0001) were greater in CancerLinQ data. The proportion receiving hormone therapy was similar in CancerLinQ and ICR data (64% and 67%, respectively, P = .41).

For analyses of overall treatment data, SEER-Medicare claims data were limited to those with Part A, B, and D coverage without enrollment in an HMO in the year following diagnosis. SEER-Medicare had the highest rates of any treatment over all three data sources across chemotherapy, hormone therapy, and immunotherapy categories (considered together or separately) for both breast and prostate cancers (Fig 1). Indicators for these treatment categories were computed using Current Procedural Terminology and National Drug Code codes as classified by the NCI Observational Research in Oncology Toolbox.^12,13 Relative to ICR treatment rates, CancerLinQ data yielded treatment rates closer to those on the basis of SEER-Medicare for breast cancer (except chemotherapy). For prostate cancer, ICR and CancerLinQ treatment rates were similar for chemotherapy and immunotherapy, but hormone therapy was higher in ICR data. The combined CancerLinQ and ICR data set yielded treatment rates that were closest to SEER-Medicare rates.

FIG 1.

Comparison of treatment regimens within 12 months of diagnosis between CancerLinQ, ICR,^a and SEER-Medicare^b data among those with (A) breast or (B) prostate cancers who were enrolled in Medicare between 2009 and 2015. CLQ, CancerLinQ; HMO, Health Maintenance Organization; ICR, Iowa Cancer Registry.^aDue to changing coding guidelines, only 2013-2015 data was included for ICR.^bThose included in the SEER-Medicare population were required to have Parts A, B, and D coverage without HMO enrollment in the year following diagnosis.

Table 3 displays comparisons between laboratory tests documented in each data source. ICR captured HER2 results for 85% of women with breast cancer compared with 26% in CancerLinQ. Among those with documentation of testing, HER2 results (positive/negative/borderline) 8% were discordant between ICR and CancerLinQ data. Similarly, the ICR had Gleason scores for 76% of people with prostate cancer compared with 34% in CancerLinQ. PSA testing was documented for nearly all patients with prostate cancer in both data sets (93%-96%). PSA categories (≥ 20 v < 20) were discordant in 16% of people tested. Among those with documentation of Gleason scores, there were no discordant results categories (≥ 8 v < 8).

TABLE 3.

Comparison of Laboratory Testing^a Data Between CancerLinQ and ICR Data for People With First^b or Only Breast or Prostate Cancers

Quality Measures

Given differences in documentation of HER2 results between the CancerLinQ and ICR data sources, denominators for the HER2(–) therapy measure differed by over 100 people. However, the compliance rate between data sources for the overall population was similar, exceeding 93% (Table 4). For those included in the SEER-Medicare population, all data sources (CancerLinQ only, ICR only, CancerLinQ-ICR [combined], and SEER-Medicare) showed a compliance rate of nearly 100%.

TABLE 4.

Numerators^a, Denominators,^a and Compliance Rates of QOPI Measures by Data Source and Overall Versus SEER-Medicare Cohort^b

The proportion of people with documented HER2 testing also affected the denominators for the HER2(+) trastuzumab measure, as did missing stage information. The ICR denominator was three-fold larger than CancerLinQ, but the compliance rate was substantially lower (65% v 83%). The combined CancerLinQ-ICR data source yielded the highest compliance rate among data sources for the overall population (90%). The compliance rate was 100% in SEER-Medicare data, but that denominator was the smallest of all three data sources.

The denominator of the high-risk prostate measure was also affected by missing stage information or Gleason score values. The combined CancerLinQ-ICR data source had the largest denominator and a higher compliance rate (79%) than either data set alone. The compliance rate was 82% in SEER-Medicare data, which had the most complete treatment information; more instances of androgen deprivation therapy were found in the claims data compared with CancerLinQ or ICR data.

Additional Variables

Smoking, height/weight, and pain assessments were examined for timing and completeness in the CancerLinQ data set alone, since these variables are not collected in ICR (Table 5). Nearly all patients with one primary cancer had a smoking assessment, and 89% had it recorded within one month of diagnosis. Seventy-six percent had pain assessments, and 52% had them within one month of diagnosis. Finally, 97% had height/weight assessments, and 76% had them within one month of diagnosis.

TABLE 5.

Timing and Completeness of Smoking, Pain and Height/Weight Assessments Collected by CancerLinQ

DISCUSSION

Our comparison of data elements in the linked CancerLinQ and ICR data sets showed that demographic variables were comparable and that CancerLinQ data captured more occurrences of hormone therapy and immunotherapy, whereas ICR data contained more complete TNM stage, HER2 testing, and Gleason scores. These variables were important components of the QOPI measures and caused substantial variation in numerators, denominators, and compliance rates. CancerLinQ data contained smoking, pain, and height/weight information for the majority of patients.

CancerLinQ data generally contained more detailed information on chemotherapy, hormone therapy, and immunotherapy than the ICR. Registries obtain the majority of their data from hospital medical records and have challenges capturing therapies delivered outside of hospital settings (eg, ambulatory infusion centers, independent practices or clinics, or community pharmacies).^14-16 CancerLinQ data contain information on all medications ordered by oncologists in participating clinics irrespective of where they were administered; so, hormone therapy and immunotherapy treatments were captured more completely. However, this could result in overestimation without validation that the treatment was received. Additionally, for patients who receive hormone therapy or immunotherapy outside of an oncology practice (eg, androgen deprivation therapy administered in a urology practice, a common situation for patients with prostate cancer), CancerLinQ may not have a record of these.

Conversely, HER2 and Gleason score laboratory testing and stage information were more completely captured in ICR data. It is likely this information was documented in the EHR because it is necessary for treatment decisions. However, it was likely captured though free text contained in clinical notes or pathology reports rather than through discrete fields. Registry abstractors are trained to manually review clinical notes and pathology reports for this key information that is coded into discrete data fields.

Because of missing information on stage, HER2 status, and Gleason score in CancerLinQ data, the number of people eligible for inclusion in the breast and prostate QOPI measure analyses varied widely between data sources. Also, numerators in both breast measures varied because of differential capture of HER2-targeted therapies. When limited to the Medicare population, the combined CancerLinQ-ICR data yielded compliance rates for breast cancer measures that were more similar to SEER-Medicare compliance rates than either source alone. SEER-Medicare was expected to have the most complete treatment data for the fee-for-service enrollees included in the comparisons.

The high-risk prostate measure also required laboratory (PSA and/or Gleason score) and stage data, which were often missing in CancerLinQ data; therefore, ICR data had a larger denominator of people eligible for inclusion. However, the compliance rate for recorded receipt of androgen deprivation therapy was higher in the CancerLinQ data, and closer to the rate derived from SEER-Medicare data.

CancerLinQ contained information about smoking status, pain, and height/weight for the majority of people in the linked data set and were often available within one month of diagnosis. These variables represent the type of rich clinical detail commonly missing from analyses of SEER or SEER-Medicare data, which is useful for researchers.

The linkage between CancerLinQ and ICR data demonstrated strengths and limitations of each source. Registries collect comprehensive data on staging and treatment across multiple settings and providers including complete and definitive stage, site, and diagnosis date information, all on the basis of established coding guidelines. They also capture chemotherapy, surgery, radiation, and selected laboratory tests documented in discrete data fields. However, treatment data are limited to first course therapy only, are not longitudinal, and can miss therapies administered outside of hospital settings. CancerLinQ data, by contrast, are limited to treatments ordered or administered by providers within the participating network of clinics and have missing values for some key variables. However, they contain detailed, longitudinal information about chemotherapy, hormone therapy, and immunotherapy, and capture several variables not collected by cancer registries (eg, smoking status, pain, and height/weight).

The main limitation of the linkage and the additional information it could yield for cancer registries is the relatively small number of oncology practices potentially eligible to participate in this linkage. However, since the first practice was integrated into the CancerLinQ database in 2016, the network has grown to now include more than 2.2 million patients with an invasive malignancy, representing all US states treated in more than 100 cancer centers and oncology practices.¹⁷ Although this still represents a fraction of all patients with cancer in the United States, CancerLinQ's continued expansion, as well as the emergence of other data aggregation networks, demonstrates the potential to establish similar linkages with cancer registries representing other geographic areas.

In conclusion, linking EHR data with cancer registry data has several benefits for CancerLinQ practices, including more complete data to assess quality measures. For researchers, the linkage would provide enhanced data on treatment and risk factors such as smoking and body mass index Furthermore, EHR data (beyond just CancerLinQ data) could add valuable information to registry data. However, there is an unrealized opportunity for additional key information that may only be accessible in clinical notes/reports and is not yet structured in a manner that allows for automatic population of registry variables. These longstanding challenges with data utility and integration can be reduced with the more widespread use of oncology-specific data formats such as the Minimal Common Oncology Data Elements (mCODE), a standard on the basis of FHIR (Fast Healthcare Interoperability Resources) approved by Health Level 7 International (HL7),¹⁸ as well as new Federal requirements for data sharing.¹⁹ Developing additional methods to efficiently incorporate EHR data continues to be a primary focus of the NCI SEER Program to enhance the longitudinal evaluation of treatment and outcomes for patients with cancer.

AUTHOR CONTRIBUTIONS

Conception and design: Mary E. Charlton, Amanda R. Kahl, Robert S. Miller, Donna R. Rivera, Kathleen Cronin

Financial support: Mary E. Charlton

Administrative support: Mary E. Charlton, Donna R. Rivera

Provision of study materials or patients: Mary E. Charlton

Collection and assembly of data: Mary E. Charlton, Amanda R. Kahl, Bradley D. McDowell, Robert S. Miller, George Komatsoulis, Jacob Koskimaki, Donna R. Rivera

Data analysis and interpretation: Mary E. Charlton, Amanda R. Kahl, Bradley D. McDowell, Robert S. Miller, Jacob Koskimaki, Donna R. Rivera

Manuscript writing: All authors

Final approval of manuscript: All authors

Accountable for all aspects of the work: All authors

AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST

The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/cci/author-center.

Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments).