Novel Data Linkages to Characterize Palliative and End-of-Life Care: Challenges and Considerations

Abstract

Context:

Working groups have called for linkages of existing and diverse databases to improve quality measurement in palliative and end-of-life care, but limited data are available on the challenges of using different data sources to measure such care.

Objectives:

To assess data concordance obtained from different sources in a novel linkage of death certificates, electronic health records (EHR), cancer registry data, and insurance claims for patients who died with cancer.

Methods:

We joined a database of Washington State death certificates and EHR to a data repository of commercial health plan enrollment and claims files linked to registry records from Puget Sound Cancer Surveillance System. We assessed care in the last month including hospitalizations, ICU admissions, emergency department (ED) visits, imaging scans, radiation, and hospice, plus chemotherapy in the last 14 days. We used a chi-square test to compare differences between healthcare in EHR and claims.

Results:

Records of hospitalization, ICU, and ED use were 33%, 15% and 33% lower in EHR versus claims. Radiation, hospice, and imaging were 6%, 14% and 28% lower in EHR, but chemotherapy was 4% higher compared to claims. These differences were statistically different for hospice (p<0.02), hospitalization, ICU, ER, and imaging (all p<0.01) but not radiation (p=0.12) or chemotherapy (p=0.29).

Conclusion:

We found substantial variation between EHR and claims for end-of-life healthcare use. Reliance on EHR will miss some healthcare use while claims will not capture the complex clinical details in EHR that can help define the quality of palliative care and EOL healthcare utilization.

Introduction

Optimizing end-of-life (EOL) care requires comprehensive analysis of the incidence of and reasons for unwanted or non-beneficial high-intensity healthcare. Many studies rely on administrative claims databases or single-site retrospective analyses to elucidate who is likely to be hospitalized or go to the emergency department at end of life (1–6) as these data sources are readily available. However, workgroups such as Measuring What Matters of the American Academy of Hospice and Palliative Medicine and Hospice and Palliative Nurses Association have called for unique data linkages to improve quality measurement in palliative and EOL care. (7) The utilization of widely available routine data, such as electronic health records (EHR), can facilitate palliative care and EOL research while avoiding the high costs and time intensity of primary data collection. (8)

Reliance on one data source may not provide a complete landscape of healthcare services provided and the reasons for such services. For example, Medicare claims underestimated the number of palliative care consultations delivered compared to findings from a phone survey of all Medicare-certified hospitals in Colorado to assess specialty palliative care services. (9) Ideally, researchers would be able to link insurance claims, which capture all billed healthcare use, to clinical registries such as the Surveillance, Epidemiology and End Results (SEER) cancer registry to confirm diagnosis, then to EHR to identify clinical and decision-making information not available in claims or registries.

In this brief report, we summarize challenges in creating unique data linkages to research palliative and EOL care. To illustrate some of these challenges, we report on our attempt to extend our previous research on EOL care for commercially insured adults. (10) We created a novel linkage of death certificates, a large health system EHR, cancer registry records and insurance enrollment and claims files. Our goal was to use the linked database to identify high-intensity EOL healthcare utilization from claims among people with a confirmed cancer diagnosis from SEER, then explore the reasons for such healthcare per the EHR. We describe our data linkage experience and offer considerations for researchers considering similar work.

Methods

First, we identified eligible subjects in a database of Washington State death certificates linked to UW Medicine, a health system spanning four hospitals, a comprehensive cancer center, and a network of outpatient primary and specialty care clinics. Of these subjects, we included those with a diagnosis of solid or hematologic malignancy who died in 2015-2016 and were age 18 or older. To ensure sufficient contact with UW Medicine to assess utilization, patients had to have at least one inpatient visit that was not for an elective procedure or two outpatient visits that were not for a second opinion in their last six months of life. We limited our years to 2015-2016 as this linkage was part of a mixed-methods study of EOL healthcare and we wanted proximal data to minimize recall bias in interviews with bereaved caregivers.

Next, we linked this cohort to a Hutchinson Institute for Cancer Outcomes Research (HICOR) database that includes registry data on cancer diagnoses from the Puget Sound Cancer Surveillance System (CSS), a part of the SEER program. These registry data were linked to enrollment and claims files from two commercial non-profit insurers, Premera Blue Cross and Regence Blue Shield, covering ~50% of commercially insured patients in the region. For inclusion in the database linkage (Figure 1), subjects had to be aged 18 or older, died in 2015-2016, continuously insured in their last 30 days of life and incurred at least one insurance claim in their last 90 days of life.

Figure 1,

Cohort diagram for population identification and subsequent linkage

We used Registry Plus Link Plus 2.0 to link between the HICOR and UW databases. Link Plus is a probabilistic record linkage program developed at the Centers for Disease Control and Prevention to facilitate record linkages between cancer registries. We matched patient records using Social Security Number, last name, first name, data of birth and death date. After receiving the linkage report, an analyst (LL) manually reviewed matches for quality. After linking the UW and HICOR databases, we extracted EHR data for all subjects in the linked cohort. The principal investigator (CLM) and a research assistant (CW) reviewed available EHR data for patients’ inpatient and outpatient visits, telephone encounters, and documentation from other health systems that had been uploaded into the UW Medicine EHR.

From the abstracted EHR data we created a binary variable (yes/no) as to whether we observed documentation of hospitalizations, intensive care unit (ICU) admission, emergency department (ED) visits, advanced imaging scans or radiation in the last 30 days of life, chemotherapy use in the last 14 days of life, and as a comparison of non-high intensity healthcare utilization, hospice use in the last 30 days. (11–12) We then compared our EHR findings to those from the claims database. We used a chi-square test to compare recorded healthcare use in each database and report statistically significant differences between databases. Finally, from the EHR we qualitatively examined documentation of: 1) goals of care discussions; 2) roles/needs of caregivers; 3) hospice discussions; and 4) palliative care needs and services.

Our research was in accordance with data use agreements between the insurers and HICOR to use claims for research purposes. We received a waiver of consent and authorization from UW Medicine to access subjects’ EHR per Washington state law regarding access to personal health information of deceased individuals. The University of Washington Institutional Review approved this study.

Results

We linked 142 subjects from the initial cohort of 2025 subjects from UW Medicine and 1336 subjects from HICOR that met eligibility criteria. After manual review we excluded six inaccurate matches, for a total of 136 subjects (7%). On average, subjects were 64.5 years old when they received a cancer diagnosis (range 35-90 years) and 40% were female. Most (89%) were white and married (87%) at the time of the cancer diagnosis. The most common cancer diagnosis was lung cancer (19%) followed by pancreatic cancer (12%).

Using the linked database, in the last month of life, 49% of patients (n=66) visited the ED, 57% (n=77) were hospitalized, 24% (n=33) had an ICU admission, 57% (n=77) received advanced imaging, 14% (n=19) underwent radiation, 7% (n=10) had chemotherapy, and 46% (n=63) enrolled in hospice. When using the EHR to identify healthcare utilization, we consistently underestimated all services except chemotherapy (Figure 2). Our estimates of hospitalization, ICU admission, and ED use were 33%, 15% and 33% lower in the EHR, respectively, compared to the claims database. We found 6% lower radiation receipt, 14% lower hospice use, 28% less imaging, but 4% higher chemotherapy use when comparing EHR to insurance claims. When comparing claims to EHR as sources of data, we found statistically significant differences for frequency of hospitalization, ICU, ED use, or imaging (all p<0.01) and hospice use (p<0.02). We found no significant difference between claims and EHR regarding radiation (p=0.12) or chemotherapy (p=0.29). While we attempted to locate advance care planning in claims using Current Procedural Terminology codes 99497 and 99498, <1% of patients had such a code, thus we did not evaluate these codes.

Figure 2,

Comparison of services identified by insurance claims versus EHR for the last 30 days of life for patients who died in 2015-2016 with cancer

From the EHR, we found goals of care documentation or a discussion about goals of care for 112 of 136 patients (82%). Caregiver availability, limits of available social/caregiving support, or patient caregiving needs were noted for 95 patients (70%). We identified documented discussions about hospice for 52 patients (38%). Other narrative data we found in the EHR that are not available in claims included a description of decision making around chemotherapy or radiation, including family preferences for palliative care consultations while patients were receiving chemotherapy or radiation (n=11, 8%).

Discussion

In this study, we created a novel linkage of death certificates, cancer registry records, insurance claims and EHR data. We found substantial variation between what was reported in the EHR and what we found in claims regarding EOL healthcare use. We note it is possible to link claims and EHR to provide a more complete picture of clinical details and decision-making, as we were able to find narrative data in the EHR for some patients regarding EOL decision making.

Unfortunately, when we linked our cohort to insurance claims, we lost over 90% of our sample, attrition similar to that noted by West and colleagues (13) when they attempted to link EHR and insurance claims. Similarly, Patorno and colleagues linked EHR data and claims of only 4% of their study population. (14) Part of this attrition is the result of our inclusion criteria. Although we assessed a commercially insured population that included Medicare Advantage patients covered by programs offered by Premera and Regence, we did not have access to claims for the large number of patients with cancer who are Medicare fee-for-service insured.

Additionally, commercially insured patients may change healthcare plans, which limits the ability to observe healthcare utilization. (15) While such switching is unlikely at end of life, we addressed this by requiring subjects be insured in their last month and incurred at least one paid claim in the last 90 days of life, which reduced our available population for linkage. While this criterion may exclude patients with limited EOL healthcare use, this criterion helps ensure we are only including patients enrolled in their last 30 days, which is the timeframe for our study outcomes. It is important to note attrition may occur as the result of differences in coding across databases, for example patient data in insurance claims may not match EHR data if a date of birth is entered incorrectly in one of the systems. (16) Finally, some commercially insured patients may seek care at another regional cancer center. UW Medicine is the largest publicly funded health system in the Puget Sound region; as such, this health system serves those who are uninsured or on Medicaid. Such patients would not be captured in the HICOR database.

In this analysis we treated claims as a gold standard for healthcare use. We acknowledge that insurance claims may be incomplete, but insurance claims often more accurately capture healthcare delivered if a patient utilizes multiple healthcare providers. A previous evaluation found claims captured 83-97% of end-of-life services for patients with cancer. (17) By relying on EHR data alone we under-estimated EOL healthcare utilization, because the EHR often does not contain information on healthcare delivered outside of the UW Medicine system unless a provider in the other health system or the patient provide documentation of such care to enter into the UW Medicine EHR.

Other studies have found high concordance between insurance claims and the EHR for services that are specific and generate a claim (e.g. influenza vaccination) compared to services that do not generate a claim (e.g. body mass index measurement) (18) or are performed as quality improvement initiatives. (19) As we relied on one data source for hospitalization data (e.g. EHR from one system or claims data from one insurer) we would have missed out-of-system hospitalizations. (20) It is also possible that miscoding of services contributes to the disconnect between EHR documentation and billed services in claims data. (21) Finally, the EHR database includes Seattle Cancer Care Alliance consultations. We found more notations indicating chemotherapy receipt in EHR versus claims; some were for patients receiving chemotherapy alongside clinical trial enrollment. In that case, the clinical trial sponsor may pay for chemotherapy rather than the insurer. If a patient comes to UW Medicine for a second opinion, the patient may return to another health system for care, which would not be recorded in UW Medicine EHR but would be paid for by the insurer.

We encountered several limitations during our data linkage. First, the database we used for this study resulted from a novel agreement between HICOR, local commercial insurers and the CSS, which allowed us to link to the UW Medicine EHR. This unique collaboration allowed us to address one of the challenges of using other claims-database linkages that may prevent data linkages using personal health information. (22) Unfortunately, as cancer is largely a disease of older adults and most older adults are enrolled in fee-for-service Medicare, many people did not have insurance claims in our dataset. Secondly, our EHR sample included patients receiving care at UW Medicine, using Dartmouth Atlas criteria for attribution of care to a healthcare system (i.e., one non-surgical hospitalization or two outpatient visits). (23) Patients referred for a single-visit second opinion would be included in the claims database but excluded from our attribution criteria for the EHR database. Similarly, patients evaluated at the cancer center for a clinical trial may be in the EHR database, but not the claims database since commercial insurance is generally not billed for such visits. Finally, we used Washington State death certificates to confirm death, so patients who died elsewhere were not included in this study.

We offer multiple considerations for researchers who want to triangulate data sources to capture multiple aspects of palliative and EOL care. First, as the EHR does not note eligibility for services and is not necessarily complete across all systems where a patient might receive care, studies that use EHR alone are subject to surveillance bias. (24) Second, the EHR is built for clinical care rather than research, so any study using the EHR must include careful consideration of data availability and completeness. (25) Given the significant expense associated with updating EHRs to incorporate algorithms and metrics for research, (26) it is difficult for most health systems to adapt their EHR to better suit research needs. Third, while routinely collected data such as EHRs offer a potentially cost-effective, easily available way to measure palliative and EOL care, having data from one healthcare system is not necessarily generalizable to larger populations. Fourth, availability of claims data may be a barrier for researchers, given the significant costs and lag time associated with accessing publicly-funded claims-linked databases. (27) However, even with exhaustive amounts of claims data, patients may still switch commercial insurers, Medicare Part D plans, or Medicare Advantage plans, making complete capture of healthcare use difficult. (28) Fifth, while the SEER registries are available for research on cancer, registries are not available for all diseases and conditions, which can make verifying a diagnosis difficult if using claims alone. Finally, if budget and time allow, larger population-based databases that use claims plus interviews (e.g. the Health and Retirement Study) should be considered for palliative care and EOL research, (29) to accommodate the large attrition typically seen when applying inclusion/exclusion criteria and linking between smaller databases.

Our findings suggest that while novel database linkages may be a rich source of data to analyze healthcare utilization, verify diagnoses, and confirm clinical outcomes (such as death) while incorporating rich clinical data from the EHR, such linkages can involve substantial loss of potentially eligible patients. When linking between different sources, researchers must consider the biases that linkage entails and the potential effect on sample size. Additionally, we found significant differences between EHR data and claims data in measuring EOL healthcare use. Researchers should note that reliance on a single source EHR will likely miss some healthcare use, while reliance on claims will fail to capture the complex clinical details available in EHR that affect palliative care provision and EOL healthcare utilization.