Background
The over 400,000 individuals receiving hemodialysis in the United States (U.S.) have exceedingly high rates of cardiovascular morbidity and mortality, with 30 % of hospitalizations and nearly 50 % of deaths attributed to cardiovascular causes [
1]. Care of this complex population is expensive. In 2011, persons with end-stage kidney disease represented just 1.4 % of Medicare enrollees but consumed 6.3 % of the total Medicare budget [
2]. Inadequate volume control is associated with both adverse cardiovascular outcomes and substantial healthcare costs among hemodialysis patients [
3‐
5]. Volume-related hospital admissions are a significant driver of the cardiovascular hospitalization rate in the hemodialysis population, and estimated annual costs related to these encounters total over $250 million [
1,
6].
Some volume overload hospitalizations may be preventable with better dialysis facility fluid management practices. For example, close attention to prescribed target (“estimated dry”) weight achievement at the end of each dialysis treatment as well as delivery of effective dietary salt and fluid restriction counseling by dialysis unit personnel may prevent some volume-related complications [
4,
7,
8]. Tracking volume overload hospitalizations represents one potential strategy to measure and assess dialysis facility fluid management practices. The Medicare-based United States Renal Data System (USRDS), a national registry of end-stage kidney disease patients, is a readily available and cost effective data source often used to monitor and study cause-specific hospitalizations in the U.S. hemodialysis population.
Administrative claims data, such as that housed in the USRDS, are primarily generated for reimbursement and billing purposes. These data may not always capture clinical subtleties, potentially affecting the accuracy of claims-identified, cause-specific hospital admissions. For example, general population validation studies suggest that ~25 % of true heart failure hospitalizations are not captured by administrative claims data [
9]. Prior evaluations of volume overload hospitalizations among hemodialysis patients were performed using USRDS data, each relying upon distinct combinations of discharge diagnosis and/or procedure codes to define events [
6,
10,
11]. However, the validity of these claims-based definitions is unknown. In the medically complex hemodialysis population, restrictions on the number of diagnosis and procedure codes that can be billed per inpatient encounter, among other factors, may influence the ability of investigators to accurately identify cause-specific hospitalizations in administrative data. As such, when choosing claims-based volume overload definitions for observational studies, investigators must consider the implications of outcome misclassification and appropriately prioritize validity metrics (e.g. sensitivity and specificity) to optimize study accuracy. Study objectives and corresponding study design should guide claims-based outcome definition selection.
We undertook this study to evaluate the validity of several claims-based definitions for volume overload hospital admissions in the hemodialysis population using rigorous medical record reviews and medical center billing data.
Discussion
To our knowledge, this is the first study evaluating the accuracy of administrative claims definitions for volume overload hospitalizations in a hemodialysis population. Our study demonstrated that clinically adjudicated volume overload hospitalization prevalence differed from claims-derived prevalence estimates. In general, claims-based definitions had high specificity and low sensitivity. Our data suggest that certain claims-based definitions for volume overload hospitalizations could: 1) generate inaccurate estimates of temporal trends in disease surveillance programs; 2) misestimate the contribution of volume-related admissions to overall hemodialysis population health care utilization costs; or 3) render inaccurate estimates in observational studies seeking to understand how exposures impact rates of volume overload hospitalizations.
Existing data reveal that volume related-factors such as chronic volume expansion, interdialytic weight gain, and ultrafiltration rate contribute to the high hospitalization and mortality rates experienced by hemodialysis patients [
3,
5,
17‐
20]. Thus, there is growing interest in identifying, quantifying and monitoring associated outcomes such as volume overload hospitalizations. To detect cause-specific hospitalizations, investigators and regulators typically rely on diagnosis and procedure codes in administrative healthcare databases such as the USRDS. However, administrative healthcare data may be inaccurate or incomplete for a variety of reasons. First, available diagnosis and procedure codes may not accurately identify the clinical condition of interest [
21,
22]. Second, medical record documentation, coding and billing practices may vary across healthcare providers or institutions, creating data inconsistencies [
23,
24]. Third, only a limited number of discharge diagnosis codes per hospitalization can be billed to insurers, possibly reducing clinical event ascertainment. Fourth, patients could receive treatment at a hospital or clinic without insurance filing, rendering administrative data sources incomplete [
22,
23,
25]. While administrative databases are often the most accessible data sources, they may not be the most accurate. Potential data shortcomings must be considered when defining clinical outcomes.
In claims-based studies of hemodialysis patients, investigators have defined volume overload hospitalizations using a variety of fluid-related discharge diagnosis code combinations (e.g. fluid overload, pulmonary edema, pleural effusion, and heart failure) in varying billing positions (Additional file
2: Table S1) [
6,
10,
11]. Banerjee et al. defined volume overload hospitalizations as the presence of a fluid overload or pulmonary edema discharge diagnosis code (separately) in any billing position [
10]. Others have employed more restrictive definitions. Arneson and colleagues considered several fluid-related diagnosis codes (e.g. fluid overload, pulmonary edema, heart failure) present in the primary billing position only [
6]. Whereas Weinhandl et al. defined volume overload hospital admissions as the presence of a fluid overload or pleural effusion discharge diagnosis code in the primary position only, or in the primary or leading secondary positions (separately) [
11]. Not surprisingly, we found that broader (versus narrower) definitions identified more true positive volume overload admission events, but did so at the expense of capturing more false positive events. Most notably, we observed that claims-based definitions containing heart failure diagnosis codes (definitions 3, 6 and 7 with codes considered in any position) had the greatest tendency to identify false positive events. This finding may be attributable to the fact that some ICD-9 codes can be used to bill for both chronic stable heart failure and acute heart failure events.
Some investigators have identified volume overload admissions using discharge diagnosis codes in conjunction with dialysis procedure codes. For example, the claims-based definition used by Arneson et al. included fluid-related discharge diagnosis codes
and also required the presence of a dialysis procedure code billed on the day of admission or the following day [
6]. Inclusion of disease-specific procedure codes often increases definition specificity [
23]. As anticipated, when we added dialysis procedure codes to diagnosis code-based definitions, we observed gains in specificity paired with reductions in sensitivity. However, the overall impact on validity estimates was minimal. This finding may, in part, be attributable to a hospital’s tendency to adhere to a patient’s outpatient hemodialysis schedule. Based solely on schedule, regardless of clinical presentation, greater than a third of all patients would be expected to receive dialysis within 24 to 36 h of admission. Furthermore, Medicare billing rules may impact the accuracy of claims-based definitions relying on dialysis procedure codes. Hospitals cannot bill dialysis CPT codes for treatments provided without the physical presence of the attending physician during the dialysis session [
16]. In administrative data, this billing rule may lead to underestimation of dialysis procedures in academic environments where trainees supervise emergent overnight or weekend dialysis without in-hospital attending presence and in community hospitals where remote nephrology coverage is common. Thus, to maximize definition stability across clinical practice environments and to avoid outcome misclassification related to billing rules, it may be prudent to omit dialysis procedure CPT codes from claims-based definitions for volume overload hospital admissions.
Dialysis patient clinical complexity may also impact accuracy of claims-defined, cause-specific hospitalizations. A limited number of diagnosis codes can be billed for each hospital encounter. Most often, payers reimburse hospitals for inpatient services based upon billed Medicare Severity Diagnosis Related Groups (MS-DRGs). The patient’s primary (or principal) discharge diagnosis in combination with other factors such as patient sex, discharge status, complications and/or comorbidities documented as secondary discharge diagnoses, medical procedures performed, and length of stay, determine the assigned MS-DRG and corresponding level of reimbursement. Hemodialysis patients often have multiple comorbidities and are treated for numerous clinical conditions during hospitalizations, resulting in a wide range of potential discharge diagnoses from which to choose for coding and billing purposes. Medicare policies allow hospitals to preferentially select discharge diagnosis codes to maximize payment as long as they are supported by adequate medical record documentation [
26]. The tendency of heart failure-based definitions to identify false positive volume overload admissions may be explained, in part, by a health system’s preference for coding more resource intensive conditions or comorbidities such as heart failure. Such practices likely vary across healthcare and reimbursement settings.
The ideal claims-based definition for volume overload hospital admissions would have perfect sensitivity (i.e. it would not capture any false negative events) and perfect specificity (i.e. it would not capture any false positive events). As claims-based definitions for clinical event identification are often imperfect, investigators must weigh the advantages and disadvantages of employing more sensitive versus more specific outcome definitions. Study objectives should drive this decision. More
sensitive outcome definitions may be preferred in scenarios where enhanced inclusiveness is desired (e.g. epidemiologic surveillance studies) and when generalizability is important (e.g. quality assessment initiatives) [
27,
28]. We demonstrated that claims-based volume overload admission definitions with poor sensitivity (<50.0 %) led to systematic underestimation of the volume-related hospital admission burden. This finding suggests that existing national prevalence estimates and temporal trends of volume-related hospitalizations may be conservative. On the other hand, more
specific outcome definitions may be favored in observational studies examining exposure–outcome associations via relative effect measures. In the setting of non-differential outcome misclassification, implementation of claims-based volume overload admission outcome definitions with perfect specificity will generate unbiased risk ratio estimates [
29]. Consideration of definition PPV and NPV is also important. Positive predictive value, like specificity, is an indicator of false positive event ascertainment, whereas NPV, like sensitivity, is an indicator of false negative event capture. However, unlike specificity and sensitivity, generalizability of PPV and NPV to a population other than the validation cohort depends on the prevalence of the outcome of interest in that population.
Strengths of our study include random selection of hospital admissions enabling estimation of the full spectrum of validity metrics, rigorous data abstraction by two independent reviewers, and utilization of standardized procedures for volume overload admission adjudication. Our study also has limitations. First, we used data from a single academic medical center. Validity estimates may not generalize to administrative data from hospitals with different billing and coding practices. Reassuringly, our study had a similar frequency of billed volume-related discharge diagnosis codes to prior investigations using USRDS data [
6,
10,
11]. Second, an established, universal definition for the clinical diagnosis of volume overload does not exist. To address this limitation, we developed a standardized algorithm for clinical adjudication based on guideline body-accepted clinical and radiologic evidence of volume overload [
14,
15]. Third, we investigated inpatient volume overload hospital admissions. Our validity estimates may not generalize to other hospital-based encounters such as observation stays or emergency department visits. Given that reimbursement rules and billing mechanisms differ across hospital encounter type, optimal volume overload definitions may vary across inpatient admissions, observation stays and emergency department visits [
30,
31]. Future studies should assess the validity of claims-based definitions for volume-related observation and emergency department visits. Fourth, we evaluated inpatient admissions from January 2010 through June 2013. Our validity estimates may not generalize to periods outside of the study timeframe. Our modest sample size prevented evaluation of potential temporal coding trends on claims-based definition validity during the study period. Finally, we studied in-center hemodialysis patients. Results should not be extrapolated to excluded populations such as peritoneal dialysis or home hemodialysis patients or those with non-dialysis dependent chronic kidney disease.