Causes of Death in End-Stage Kidney Disease: Comparison between the United States Renal Data System and a Large Integrated Health Care System

End-stage kidney disease (ESKD) burdens >600,000 individuals in the USA, with a 5-year survival rate of approximately 50% [1]. ESKD remains a condition with high mortality where the immediate period after transition to ESKD is associated with the highest death rates [2]. Mortality rates up to 30% have been described within the first year of transition to ESKD [3‒6]. Cardiovascular disease is reported as the leading cause of death (COD) among dialysis patients, followed by sepsis/infection [7, 8]. Among ESKD patients in the USA, nearly a quarter of deaths are listed as being from an unknown cause or missing cause [1]. Understanding the outcomes of treatment approaches among individuals with ESKD will continue to depend on an improved understanding and an accurate ascertainment of the COD in this population [8].

Accurately identifying COD in the ESKD population has been challenging. Several ESKD observational studies have attempted to compare the accuracy between death records and national databases, but overall, the findings have been inconsistent. Observational reports comparing death certificates with dialysis registry data from outside the USA have demonstrated poor concordance in CODs [9‒12]. In the USA, evaluations of the accuracy of CODs between death certificates and registry data have been mostly derived from smaller and more specialized populations whose results may not be generalizable to the overall US ESKD population [13, 14]. Overall, national and large databases/registries provide a valuable source of information for epidemiologic research. They help drive the focus of clinical research endeavors, public health allocations, and treatment guidelines.

The United States Renal Data System (USRDS) maintains one of the largest disease registries in the USA. It serves as the primary source of outcome data for the ESKD population in the USA [3]. We previously found that within an integrated health system of Kaiser Permanente Southern California (KPSC), 86–90% of death records had corresponding USRDS records, of which 87% had death dates that matched [15]. Furthermore, there was 98.3% concordance of death reported by both KPSC and the USRDS. The death date concordance achieved 99% from both sources when allowing a potential 1-week lag in reporting systems. These findings suggested that death records in the USRDS registry have high correlation with KPSC and accuracy with regards to the date of death. However, it bears mention that the accuracy of the COD listed on the mortality records within the USRDS registry was not assessed.

The state of California represents one of the most populated and largest economies in the world. Given that California has the largest absolute number of incident and prevalent ESKD patients across all USA states and territories, the KPSC database is a particularly important resource with respect to its capture of a large proportion of the broader incident ESKD population. The KPSC population is sex/gender-balanced and racially/ethnically diverse across different socioeconomic levels. Using an incident ESKD population derived from the large diverse real-world clinical environment of KPSC, we sought to evaluate the concordance of CODs between mortality records from KPSC with the CODs obtained from the USRDS registry.

A retrospective cohort study of KPSC members identified between January 1, 2007, and December 31, 2016, was performed. KPSC is an integrated health system comprised of 14 medical centers, over 200 clinics, and patients at over 300 dialysis facilities throughout Southern California [16]. The ESKD population is racially/ethnically diverse, reflective of the 4.7 million KPSC membership population. The membership and CKD patient population are racially, ethnically, and socioeconomically diverse, reflecting the general population of Southern California [17‒19]. All KPSC members have similar benefits and access to health care services, clinic visits, procedures, and copays for medications. Complete health care encounters are tracked using a common electronic health record (EHR) from which all study information was extracted. The study protocol was reviewed and approved by the KPSC Institutional Review Board (#10254) and was exempt from informed consent.

The study population has been previously described [2, 15]. In brief, the study population included individuals aged 18 years and older who transitioned to ESKD between January 1, 2007, and December 31, 2016, and had mortality records with a COD reported in both USRDS and KPSC databases. The KPSC ESKD population is monitored and followed through an internal registry overseen by the KPSC Renal Business Group [16]. ESKD was defined as treatment with hemodialysis, peritoneal dialysis, or transplantation. Individuals with a reported COD until December 31, 2016, were included in this study.

KPSC members with ESKD were matched to USRDS data by social security number, name, sex, and date of birth. Demographic characteristics for all matched ESKD patients were obtained from the KPSC EHR.

Mortality information was obtained from USRDS and KPSC databases to confirm the reported death. In the USRDS database, CODs are obtained from several sources including the Centers for Medicare and Medicaid Services (CMS) Enrollment Database, CMS Death Notification Form 2746, Medical Evidence Form 2728, Organ Procurement and Transplantation Network Transplant Follow-up Form, CROWNWEB database, and inpatient claims [20]. USRDS CODs are classified into 8 large group categories: cardiac, vascular, gastrointestinal, infection, liver disease, metabolic, endocrine, and others including drug overdose, seizures, and withdrawal from dialysis. As per the 2019 USRDS Researcher’s Guide, death dates among ESKD patients are determined from a hierarchy of several sources. Form 2746 is the highest priority, whereas form 2728 is fifth on the list.

KPSC maintains a comprehensive research data warehouse with data from its EHR and other sources. Mortality information, including the death date and COD, is primarily obtained from the linkage to California State Death Master Files, supplemented by 6 data sources: California State Multiple Cause of Death Master Files, Social Security Administration (SSA) Death Master Files, KPSC hospital and emergency room records, KPSC Perinatal Service System files, KPSC Membership System files, and claims submitted to KPSC from outside facilities or information reported to the health plan directly. The state death record CODs are classified in accordance with the International Classification of Disease (ICD) Tenth Revision (ICD-10) for deaths following 1999. Thus, KPSC CODs were categorized based on CDC-suggested categories: diseases of the circulatory system, endocrine/nutritional/metabolic diseases, diseases of the genitourinary system, neoplasms, diseases of the respiratory system, diseases of the digestive system, certain infectious and parasitic diseases, external causes of morbidity, diseases of the musculoskeletal system and connective tissue, diseases of the nervous system, mental/behavioral/neurodevelopmental disorders, congenital malformations/deformations/chromosomal abnormalities, diseases of the blood/blood-forming organs/certain disorders involving the immune mechanism, and diseases of the skin and subcutaneous tissue. The COD is selected from the conditions entered by the physician on the “cause of death” section of the death certificate. When more than one cause or condition is entered by the physician, the COD is determined by the sequence of conditions on the certificate classified in accordance with the ICD and associated selection rules and modifications [21, 22]. The underlying COD, cause 2, cause 3, and cause 4 in the cause of mortality records in the KPSC death file were assigned as the underlying COD and secondary, tertiary, and quaternary CODs separately. While not necessarily “the gold standard,” death certificates are most frequently for reporting CODs throughout the USA.

For those with death records in both USRDS and KPSC, demographic information was extracted. Mean (standard deviation) and frequency (percent) were provided for continuous or categorical variables separately.

The primary analysis compared the KPSC underlying CODs to the underlying CODs from USRDS. Underlying CODs in KPSC were reviewed separately by 2 physicians and categorized into the classification system according to the USRDS (Medicare Form 2746) categorizations. If the classification differed between the 2 physicians (nephrologists), then the inconsistent causes were reviewed by a third physician (nephrologist), and a final decision was made after discussion. The agreement between the 2 physicians was compared using Cohen’s weighted kappa statistic. Percent total agreement, both positive and negative agreement, and Cohen’s weighted kappa statistic were determined between KPSC underlying CODs and USRDS underlying CODs overall and for each COD subcategory [23].

The secondary analysis further compared the USRDS underlying CODs to the secondary, tertiary, and quaternary CODs in KPSC records. Frequencies and percentages were compared for each subcategory of USRDS underlying CODs except the cause listed as “Other.”

Overall and subcategory-specific agreement of the underlying COD were assessed using the Cohen’s weighted kappa statistic with 0.81–1.0 indicating almost perfect agreement, 0.61–0.80 indicating substantial agreement, 0.41–0.60 indicating moderate agreement, 0.21–0.40 indicating fair agreement, and ≤0.20 indicating poor agreement [24]. Given low kappa values may reflect low prevalence of a cause in the population, and no lack of agreement, the proportion of negative agreement, and proportion of positive agreement were also determined [25]. Positive agreement was defined as both databases recording the same COD. Negative agreement was determined as a COD that was not recorded in both USRDS and any of the KPSC reported CODs.

All statistical analyses were generated using the SAS Enterprise Guide (version 7.1; SAS Institute, Cary, NC, USA). Results with p < 0.05 were considered statistically significant.

A total of 4,188 death records were reported for all adult KPSC members who transitioned to ESKD between January 1, 2007, and December 31, 2016 (Fig. 1). Of these, 4,118 (98.3%) had the COD reported by both USRDS and KPSC. Overall, 87% of records had an exact match on the date of death, and another 12.0% matched within a week from both KPSC and USRDS databases. These results have been described in detail previously [15].

The mean age at death was 71 (SD 11.9) years, and 41.2% were women. The race/ethnicity composition of the population was 38.2% non-Hispanic White, 21.0% Black, 28.8% Hispanic, and 9.1% Asian (Table 1). Overall, 8.8% were on either hospice or palliative care at their time of death.

The most common CODs reported in KPSC as obtained from the California State Death File were diseases of the circulatory system (35.7%) followed by endocrine, nutritional, and metabolic diseases (24.2%); diseases of the genitourinary system (12.9%); and neoplasms (9.6%) (Table 2). The most common COD reported from the USRDS were cardiac disease (46.9%), followed by others (37.1%) including 12.6% reported as withdrawal from dialysis, followed by infection (10.1%) (Table 3).

Causes of mortality records in KPSC were reviewed and recategorized into the USRDS cause of mortality groups by 2 physicians. Categorization of causes between 2 initial raters was found to have a Cohen’s weighted kappa statistic (95% CI) of 0.74 (0.69–0.79), indicating substantial agreement among raters. Any inconsistent causes were reviewed by a third physician, and a final decision was made after discussion. Overall agreement between the underlying COD in KPSC with the underlying COD in the USRDS was 36.4% (1,500 records), with Cohen’s weighted kappa statistic of 0.20 (0.18–0.22), indicating slight agreement (Table 4). Cohen’s weighted kappa statistic (95% CI) for agreement between cardiac disease, vascular disease, infection, and liver disease and other diseases individually were between 0.21 (0.18–0.23) and 0.006 (−0.003 to 0.01) (Table 4). Negative agreement between underlying CODs in KPSC and underlying CODs in USRDS was highest for gastrointestinal causes (98.3%), followed by liver disease (97.9%). Negative agreement was lowest for cardiac causes (69.0%). Positive agreement between underlying CODs in KPSC and underlying CODs in USRDS was highest for cardiac causes (46.4%). Positive agreement was lowest for metabolic causes (1.4%), with the prevalence of metabolic-related death of only 0.4% listed in USRDS record.

When comparing the underlying CODs in USRDS to the underlying, secondary, tertiary, or quaternary CODs from the death certificate records in the KPSC database, it was found that 453 of 2,593 (17.4%) causes in categories other than “other causes” had no agreement. Specifically, 1,670 of 1,932 (86.4%) underlying CODs assigned as cardiac disease in USRDS were also listed as one of the causes in California death certificate records, while only 277 of 416 (66.6%) underlying CODs assigned as infection in USRDS were also listed as one of the causes in California death certificate records (online suppl. Table 1; see www.karger.com/doi/10.1159/000520466 for all online suppl. material).

Cardiac disease was listed as the most frequent underlying COD for ESKD patients in both sources with slight difference in the percentages. The cardiac COD was reported in 43.6% of females and 49.3% of males in the USRDS database but only 19.6% in females and 28.7% in males from the KPSC database. Among linked male records, positive agreement between KPSC and USRDS CODs was highest for cardiac disease (80.7%). Negative agreement between male records was highest for vascular disease (94.3%). For females, positive agreement between KPSC and USRDS CODs was highest for cardiac disease (63.7%), and negative agreement was highest for infection (92.3%) (Table 5).

Our study comparing CODs between death certificates from an integrated health system (KPSC) and the USRDS registry data found only fair to slight agreement. Our results suggest that CODs derived from the USRDS should be used with caution when describing and analyzing the ESKD population in the USA. Our study is one of the largest samples of death certificates evaluated against USRDS data to date and further supports the need to improve upon the accuracy and reliability of CODs for the ESKD population which would help improve the understanding of treatment-related outcomes and shape management of ESKD patients. Overall, improving outcomes and even reducing mortality in ESKD will continue to depend on a better understanding and an accurate ascertainment of the COD in this population.

We observed only slight positive agreement between underlying CODs recorded in KPSC and USRDS death records. In evaluating subcategories of each COD listed by the USRDS, we similarly found only slight to fair agreement. Positive agreement was highest for cardiac disease and was lowest for metabolic conditions. Overall, we observed higher agreement among male patients. We suspect this may be attributed to the increased prevalence of cardiac disease among the male population and is thus being considered a more common underlying COD by evaluators. Overall, accuracy in the ascertainment of the COD is also essential from a public health perspective. The appropriate classification of disease-specific mortality in epidemiologic and clinical studies can help allocate resources at a population-management level [26].

Comparison between CODs as reported by death certificates versus large registry data has been challenging. This is likely due to inherent limitations in data reporting and differences in categorizations of CODs between sources. A large study evaluating 28,675 patients with deaths reported in the National Death Index (NDI) and a large registry similar to the USRDS (the Australian and New Zealand Dialysis and Transplant Registry) also found low agreement (36%, Cohen’s weighted kappa statistic 0.22) in underlying CODs between the 2 sources. The lack of agreement was felt to be, in part, a result of the absence of certain CODs in each registry. The Australian and New Zealand Dialysis and Transplant registry does not list diabetes or renal failure as a COD, whereas the NDI does not list dialysis withdrawal as a cause [10]. Perneger et al. [14] evaluated death certificates of 335 patients linked to the Maryland Medicare End-Stage Renal Disease Program (now a part of the USRDS) and similarly found agreement to be 31% with a Cohen’s weighted kappa statistic of 0.20. One smaller study compared the COD from 221 Hemodialysis (HEMO) study patients as determined by trained HEMO study outcome review committee physicians to the Health Care Financing Administration (HCFA) Death Notification Form 2746. While they found that broad categories of death were similar between the HEMO committee (39% cardiac) and the HCFA form (45% cardiac), there remained significant differences. There were twice as many deaths from an unknown cause in the HCFA death classification as in the HEMO death classification [13]. Similarly, our study found only 1.7% to have an unknown COD among KPSC death records, whereas the USRDS records had 37.1% classified as other/unknown. This discrepancy likely reflects differences in processes of care and data collection between the 2 sources: death record completion by physicians who may be more meticulous in coding the underlying COD versus records completed for a large registry which may not undergo the same level of adjudication in confirming the accurate COD.

There are several key differences in cause-specific mortality worth highlighting. While the USRDS reports cardiovascular disease and infections as the leading CODs among dialysis patients [7], we found that KPSC death records indicated diseases of the circulatory system (35.7%), endocrine/nutritional/metabolic disease (24.2%), and diseases of the genitourinary system (12.9%) to be the three most common ESKD CODs. Registry reports traditionally have been limited in restricting the reportable underlying COD categories available to the evaluator. Most dialysis registries, like the USRDS, use an internal coding system to classify the COD instead of the ICD-10 [12]. Neither diabetes nor renal disease are listed as a potential underlying COD on the USRDS CMS Death Notification Form 2746, whereas genitourinary disease was reported to be the third leading underlying COD among our KPSC ESKD population (12.9%). In addition, the overall mortality burden attributed to kidney disease or diabetes may be underrepresented in the USRDS registry, which focuses more on the mechanism or event that directly led to the COD. Given that chronic diseases are present for a number of years, it is difficult to determine whether chronic disease was the underlying COD or contributed to the underlying mechanism of death (e.g., sudden cardiac death or myocardial infarction). Unfortunately, this becomes problematic in that it creates room for more individual interpretation in completing the death forms. While kidney failure is not listed as a category for the COD in the USRDS registry, the USRDS reports that withdrawal from dialysis accounts for 18.1% of deaths among the ESKD population [7]. In our study, we observed 12.6% deaths due to withdrawal from dialysis. Furthermore, there is variability in who completes the USRDS CMS Death Notification Form 2746. The form may be completed by the primary nephrologist, a covering physician, or a nonphysician who all may have differring perspectives and information on the patient.

Death certificates themselves have their own limitations in reliably reporting the underlying COD. The determination of underlying COD is determined by the person completing the death certificate and then coded based on the text. Therefore, errors can be introduced during assessment and coding [9, 21]. One study found that while there was high reliability (>90%) of COD statistics for major CODs such as cancers and acute myocardial infarction, there is actually low reliability (<70%) for COD statistics for chronic diseases such as diabetes and renal diseases. Coders overall had difficulty in selecting chronic diseases as an underlying COD [9]. There remains no gold standard for identifying or coding CODs for death certificates or registry data. A more standardized approach to the definition and classification of ESKD CODs may improve reliability and accuracy.

There are several potential limitations to our study that may influence the interpretation the findings. While the match between KPSC and USRDS death records may be accurate, KPSC death records are identified through various sources, both internal and external. External sources are matched to patients with a probabilistic algorithm thus false matching may occur. However, KPSC’s internal quality control data suggest a high degree of accuracy overall [15]. Similarly, USRDS data are obtained from multiple sources, and inaccuracies can occur within those sources as well. Second, there are inherent differences in the documentation patterns and differing perspectives between providers completing the death certificate. CMS Death Notification Form 2746 is often completed by administrative staff members at the dialysis unit who are not necessarily trained to complete these forms, adding another layer of variability. Overall, there is inherent heterogeneity between individuals in this categorization process which perhaps an automated coding system may be able to overcome. The review of death records by our 2 physician review led to one-third of cases having some discrepancy. To increase our capture of COD, we evaluated agreement between any CODs listed on the death certificate (secondary, tertiary, and quaternary) but still found concordance to be only fair. Another potential limitation of our study is that these findings may not be generalizable beyond KPSC where demographics and form completion patterns may also be different.

Among an incident ESKD population with complete mortality records, we found that there was only fair or slight agreement between COD reported between the national USRDS registry and KPSC, a large integrated health care system. The accuracy and agreement could potentially be improved by increasing physician awareness and by standardizing the coding of death certificates. Accurate death records within the ESKD population are necessary to properly evaluate risk and prognostic factors that will ultimately inform preventative and therapeutic considerations and guide resource allocations in this high-risk population.

The ethical review board of KPSC sites approved the protocol and related documents. The study adhered to Good Clinical Practice guidelines as well as local laws and regulations. The study protocol was reviewed and approved by the KPSC Institutional Review Board (#10254) and was exempt from informed consent.

The authors declare that they have no relevant financial interests relevant to this work and manuscript.

This study was supported by KPSC internal research funding. The funder/supporter had no role in study design, data collection, analysis, reporting, or the decision to submit for publication.

S.J.J. and J.J.S. contributed to the research idea and study design; H.Z., S.K.B., and S.F.S. contributed to data acquisition; H.Z., J.J.S., S.F.S., S.K.B., and N.S.T. contributed to analysis or interpretation of data; J.J.S. and S.J.J. contributed to study supervision; H.Z. performed statistical analysis; S.K.B., J.J.S., H.Z., S.F.S., C.M.R., and S.J.J. performed drafting and critical revision of the manuscript. Each author contributed important intellectual content during manuscript drafting or revision and agrees to be personally accountable for the individual’s own contributions and to ensure that questions pertaining to the accuracy or integrity of any portion of the work, even one in which the author was not directly involved, are appropriately investigated and resolved, including with documentation in the literature if appropriate.

All data generated or analyzed during this study are included in this article and its online supplementary material files. Further enquiries can be directed to the corresponding author.

Hui Zhou is the co-first author.