Development of an Administrative Data-Based Frailty Index for Older Adults Receiving Dialysis

Abstract

Visual Abstract

Background 

Frailty is present in ≥50% of older adults receiving dialysis. Our objective was to a develop an administrative data–based frailty index and assess the frailty index’s predictive validity for mortality and future hospitalizations.

Methods 

We used United States Renal Data System data to establish two cohorts of adults aged ≥65 years, initiating dialysis in 2013 and in 2017. Using the 2013 cohort (development dataset), we applied the deficit accumulation index approach to develop a frailty index. Adjusting for age and sex, we assessed the extent to which the frailty index predicts the hazard of time until death and time until first hospitalization over 12 months. We assessed the Harrell’s C-statistic of the frailty index, a comorbidity index, and jointly. The 2017 cohort was used as a validation dataset.

Results 

Using the 2013 cohort (n=20,974), we identified 53 deficits for the frailty index across seven domains: disabilities, diseases, equipment, procedures, signs, tests, and unclassified. Among those with ≥1 deficit, the mean (SD) frailty index was 0.30 (0.13), range 0.02–0.72. Over 12 months, 18% (n=3842) died, and 55% (n=11,493) experienced a hospitalization. Adjusted hazard ratios for each 0.1-point increase in frailty index in models of time to death and time to first hospitalization were 1.41 (95% confidence interval, 1.37 to 1.44) and 1.33 (95% confidence interval, 1.31 to 1.35), respectively. For mortality, C-statistics for frailty index, comorbidity index, and both indices were 0.65, 0.65, and 0.66, respectively. For hospitalization, C-statistics for frailty index, comorbidity index, and both indices were 0.61, 0.60, and 0.61, respectively. Data from the 2017 cohort were similar.

Conclusions 

We developed a novel frailty index for older adults receiving dialysis. Further studies are needed to improve on this frailty index and validate its use for clinical and research applications.

In Brief

Key Points

• Accounting for comorbidity, this frailty index demonstrates an independent association with mortality and hospitalizations.
• This frailty index has potential application for both observational studies and population health management.

Introduction

Frailty, or decreased physiologic reserve and increased vulnerability to adverse outcomes, is often used for mortality risk stratification in older adults (¹,²). When examined in older adults receiving dialysis, frailty is present in <70% and is an independent risk factor for mortality when accounting for comorbidities (³,⁴). In these studies, frailty was operationalized as phenotypic frailty, the clinical syndrome characterized by presence of three or more of: exhaustion, slow walking speed, unintentional weight loss, weakness, and low physical activity (⁵). Although phenotypic frailty is grounded in a biologic hypothesis of vulnerability (⁵), it relies on in-person assessment (⁶). Another frailty assessment tool, a frailty index derived from a deficit accumulation index approach, has the potential to overcome implementation barriers of screening for phenotypic frailty by using previously collected data. However, this approach to frailty assessment has not been widely adopted for clinical application in the dialysis population.

Conceptually, the deficit accumulation index approach to frailty acknowledges the human body as a system with high redundancy of interrelated parts. With aging, individuals tend to have greater risk of adverse outcomes due to loss of this redundancy (i.e., accumulation of deficits) (⁷). Deficits are comprised of symptoms, signs, disability, disease, and laboratory measurements, can range in severity (e.g., liver failure, hearing loss), and carry equal weights (⁷,⁸). A deficit accumulation index quantifies frailty on a continuum from 0 to 1 as the proportion of deficits present yielding a frailty index. Although a frailty index is conceptually different from phenotypic frailty, studies have demonstrated that as scores on phenotypic frailty scales increase, the severity of frailty as measured by the frailty index also increases (⁹). Deficit accumulation indices, like phenotypic frailty, have been shown to reliably predict mortality and other adverse outcomes (⁷,⁸). Because of this predictive value, frailty indices have been utilized for mortality risk stratification in primary care, inpatient, and specialty clinics for conditions associated with accelerated aging (e.g., cirrhosis, HIV, oncology) (^10–13). Recently, a frailty index developed from Medicare datasets improved the prediction of adverse outcomes, creating a method for frailty adjustment to be incorporated in comparative effectiveness studies where frailty may confound the association between a treatment and outcomes (¹⁴). On the basis of this evidence, a frailty index could also be useful for risk stratification and comparative effectiveness research in older dialysis populations. However, it is not known if a frailty index derived from a dataset of older adults receiving dialysis can be achieved.

Our objective was to apply the deficit accumulation index approach to develop an administrative data–based frailty index using data from the United States Renal Data System (USRDS) (¹⁵), a national dialysis data registry. We evaluated its predictive validity for time to death and time to first hospitalization over 1 year.

Methods

Study Design

This is a retrospective cohort study using USRDS Standard Analytic Files (SAF) of older adults initiating dialysis for development of a frailty index. We determined each individual’s frailty index at 6 months after dialysis initiation, and conducted a longitudinal study to evaluate the predictive validity of the frailty index on death and future hospitalizations over the subsequent 12 months (Figure 1). This study was reviewed and approved by the Duke University Institutional Review Board under Pro00102857 and the USRDS. Given the retrospective nature of this study and the deidentified data, informed consent was waived by the Duke University Institutional Review Board.

Figure 1.: Study diagram for longitudinal data analysis.

Data Source

USRDS provides health system encounters for the majority of the dialysis population through linkage to medical claims (including hospitalizations, physician services, and prescription costs) from Medicare, the US federal health insurance program (¹⁶). The USRDS SAFs included claim records for Institutional (inpatient, outpatient, home health agency, and hospice) (Medicare Part A claims) and Physician supplier (Physician/supplier and Durable Medical Equipment) (Medicare Part B claims) services. We used the USRDS data, including Medicare claims, for cohort selection, frailty index development, and outcome ascertainment. We ascertained death dates from the USRDS Patients SAF file and date of first hospitalization from USRDS Hospitalization SAF.

Study Population

Similar to other frailty index studies (¹⁰,¹¹,¹⁴), we identified a study population of older adults aged 65–90 years. We included individuals who initiated dialysis between January 1 and December 31, 2013 with continuous Medicare coverage (Parts A and B) during the first 6 months after dialysis initiation. We excluded individuals missing a Center for Medicare and Medicaid Services (CMS) Medical Evidence (CMS 2728) form (which is mandated for completion by dialysis clinicians, and includes sociodemographics, comorbidities, and kidney disease items) within the 2 years from the start of the cohort (e.g., 2013 or 2014), and individuals who received a transplant, discontinued dialysis, died, or had evidence of living in a nursing home or hospice care within the first 6 months of dialysis, as ascertained from CMS 2728 form and Medicare Part B claims.

Data from this incident 2013 cohort were used for frailty index development. To evaluate the predictive performance of the frailty index in an unbiased cohort, we established a second cohort, a validation cohort, with data from individuals who initiated dialysis between July 1, 2016 and June 30, 2017 and met the inclusion criteria of the 2013 cohort described above. Henceforth, this validation cohort is referred to as the 2017 cohort. We selected the 2017 validation cohort date range to ensure the cohort was distinctly separated in time from the development cohort, which would help us assess the frailty index’s temporal generalizability (¹⁷).

Frailty Index Development

The frailty index was developed using the CMS 2728 form (comorbid conditions and patient status variables) and Medicare claims data. For a deficit accumulation index approach, we identified potential health deficits through a documented International Classification of Diseases ninth revision (ICD-9) diagnosis code, current procedural terminology (CPT-4) code, or health care common procedure coding system (HCPCS) code, found in the Medicare claims records (¹⁴). There were 216 total deficits considered for inclusion in the index on the basis of groups of ICD-9, CPT, and HCPCS codes, and selected CMS-2728 form variables. Taking an agnostic approach, we considered all plausible ICD-9, CPT, and HCPCS codes for inclusion in the frailty index. We utilized ICD-9 codes by 102 previously determined clinically meaningful categories (¹⁴). Similarly, we collapsed individual CPT and HCPCS codes into the 104 Berenson–Eggers Type of Service (BETOS) codes . The BETOS system is primarily used for cost analyses and provides clinical groupings of CPT and HCPCS codes. Of note, some CPT and HCPCS codes indicate diagnostic tests irrespective of abnormal findings (e.g., x-ray). Several CMS 2728 variables are comorbidities found in claims, so we selected the CMS 2728 variables representing types of deficits (e.g., signs, social vulnerability, disability) not often captured by claims. Those CMS 2728 variables considered for inclusion were employment status, body mass index (dichotomized into two variables: ≥30 kg/m² and <20 kg/m²), Medicare and Medicaid dual eligibility, inability to ambulate, inability to transfer, needing assistance with daily activities, and institutionalization (referring to residence in nursing home, assisted living center, or any other facility).

For deficit accumulation index development, we followed principles identified from prior studies to establish criteria for deficits: (1) the deficit must be positively correlated with age, identified as a positive slope and R²>0.3 via simple linear regressions of the condition on age, (2) the deficit must not be present in nearly all individuals in the youngest age group in the study, defined as ≤80% prevalence among those aged 65–69 years, and (3) the deficit must have a minimum lower prevalence threshold of ≥1% in the study population (¹⁰,¹⁴,¹⁸). We applied these criteria to claims during the first 6 months of dialysis for each individual in the 2013 cohort and to CMS 2728 variables assessed only once for each individual (at dialysis initiation). Within the available administrative data in the 2013 cohort, 53 variables met criteria to be deficits: 30 ICD-9 diagnosis code groups, 20 BETOS code groups, and three CMS 2728 variables (body mass index <20, inability to transfer, and needs assistance with daily activities) (Figure 2, Supplemental Table 1 for full list of codes and descriptions). For the 2017 cohort, we identified ICD-10 code groups that correspond to the ICD-9 code groups that met deficit criteria. We calculated the frailty index as the proportion of deficits identified in the first 6 months of dialysis. Absence of an ICD-9, CPT, HCPCS, or CMS 2728 variable in the first 6 months was interpreted as absence of the deficit.

Figure 2.: The 53 ICD-9, BETOS, and CMS 2728 form variables identified as frailty index deficits. BETOS, Berenson–Eggers type of service; BMI, body mass index; CMS, Centers for Medicare and Medical Services; COPD, chronic obstructive pulmonary disease; EKG, electrocardiogram; ICD-9, International Classification of Diseases 9th revision. Unclassified domain refers to claims that have not been further categorized by Medicare.

Statistical Analyses

We calculated summary statistics of cohort characteristics (age at index calculation [which is used in all analyses], sex, race/ethnicity [as reported in USRDS files], ESKD cause, modality, and comorbidity) and the frailty index both overall and stratified by sex, age group (5-year increments), and modality. We include race and ethnicity to characterize the cohort with respect to the overall US dialysis population. We used Wilcoxon rank-sum test to assess differences in frailty index by dialysis modality (in-center hemodialysis vs other).

In Cox proportional hazards regression models, we assessed the extent to which the frailty index predicted the hazard of time until death and time to first hospitalization, during a 12-month follow-up period (i.e., 6–18 months after dialysis initiation) (Figure 1). Model covariates included age and sex. Kidney transplant and loss of Medicare coverage were considered censoring events for both outcomes, whereas death before hospitalization was considered a censoring event for the time to first hospitalization outcome. We assessed for violation of the proportional hazard assumption using Schoenfeld’s residuals. To compare the predictive power of the frailty index with a claims-based, dialysis-specific comorbidity index developed by Liu et al. (¹⁹), we repeated these Cox models with the comorbidity index as the independent variable. Harrell’s C-statistic from Cox regression models were computed with the frailty index and comorbidity index individually and jointly (²⁰). We assessed the correlation between the two indices using the Spearman correlation coefficient.

We assessed the association between length of first hospital stay and the frailty index among those in the cohort who experienced at least one hospitalization during follow-up using negative binominal regressions with age and sex as covariates.

We conducted exploratory analyses to identify frailty index categories and assess for hazards of death and first hospitalization associated with each frailty index category. Using a visual inspection of a locally estimated scatterplot smoothing curve between time to death and the frailty index using the 2013 cohort, we identified three frailty index categories with minimally overlapping confidence intervals in the estimated survival curves. In Cox models, the frailty index category was the independent variable, and model covariates included age, sex, and the Liu comorbidity index. We included the Liu comorbidity score to account for comorbidity burden not captured in the frailty index in which each deficit carries an equal weight. We stratified the 2013 cohort into each of the three frailty index categories, and for each of the 53 deficits, calculated the proportion of individuals with that specific deficit.

All analyses were repeated using the 2017 validation cohort. There were no adjustments for multiple comparisons. Model parameter estimates are reported with 95% confidence intervals (95% CIs). Statistical tests were conducted at the α=0.05 level. All statistical analyses were implemented using SAS/STAT 15.1 software (SAS Institute, Cary NC).

Results

Cohort Selection and Characteristics

We assembled frailty index development and validation cohorts of 20,974 and 21,355 patients, respectively. For the development cohort, there were 57,932 adults aged ≥65 years who started dialysis in 2013; however, there were 36,958 exclusions: primarily for death (n=10,733) or absence of continuous Medicare coverage in the first 6 months (n=17,072) (Figure 3). The 2013 cohort had a mean (SD) age of 74.9 (6.6) years, was 43% (n=9099) female, and had median (Q1, Q3) Liu comorbidity score of 7 (³,¹⁰) (Table 1). The 2017 validation cohort had similar cohort selection flow (Supplemental Figure 1), demographics, and comorbidity burden (Table 1).

Figure 3.: Cohort flow diagram for the 2013 incident cohort.

Table 1. - Cohort characteristics

Characteristic	2013 Cohort	2017 Cohort
	n=20,974	n=21,355
Age at index calculation, mean (SD)	74.9 (6.6)	74.6 (6.4)
Female, n (%)	9099 (43)	9037 (42)
Race and ethnicity, n (%)
Hispanic	1995 (10)	1944 (9)
Non-Hispanic Black	4202 (20)	3952 (19)
Non-Hispanic White	13475 (64)	14062 (66)
Other a	1302 (6)	1397 (7)
Liu comorbidity index, b median (Q1, Q3)	7.0 (3.0, 10.0)	7.0 (4.0, 11.0)
ESKD cause, n (%)
Diabetes	9512 (45)	10203 (48)
Hypertension	7545 (36)	7323 (34)
Glomerulonephritis	1163 (6)	1180 (6)
Other c	2239 (11)	2649 (12)
Modality, n (%)
In-center hemodialysis	18840 (87)	18228 (85)
Other	2754 (13)	3127 (15)

SD, standard deviation; Q1, Q3, first and third quartile.

^aOther race/ethnicity includes Asian, American Indian or Alaska Native, Native Hawaiian or Pacific Islander, Other or Multiracial, and Unknown.

^bA claims-based comorbidity index developed in patients on dialysis (¹⁹).

^cOther ESKD causes includes other reasons, missing reasons, and unknown reasons.

Within the 2013 cohort, 3% (n=528) had no deficits, the deficit distribution among those with deficits was right-skewed as the highest number of deficits accumulated by an individual was 38 (of the 53) (Figure 4). Among those with deficits, the mean (SD) number of deficits was 16 (7), corresponding with a mean (SD) frailty index of 0.30 (0.13), and the mean (SD) frailty index for women and men was 0.31 (0.13) and 0.29 (0.13), respectively. There was an increase in frailty index across age groups in both sexes (Table 2). Compared with those receiving other modalities, the frailty index was higher among those receiving in-center hemodialysis (0.31 [0.13] versus 0.24 [0.12]) (P<0.0001). In the 2017 cohort, there was a similar mean (SD) frailty index (0.29 [0.12]) among those with deficits, trends by sex and age group (Supplemental Table 2), and higher frailty index among those receiving in-center hemodialysis compared with other modalities (0.30 [0.12] vs 0.24 [0.11]) (P<0.0001). The correlation between frailty index and comorbidity index was 0.63 for both cohorts.

Figure 4.: Distribution of frailty index deficits in the 2013 cohort. This histogram demonstrates the distribution of deficit count among individuals in the development cohort. The maximum number of deficits by any individual was 38.

Table 2. - Frailty index by age and sex in 2013 cohort

Age Group, yr	Male		Female
	Mean (SD)	n	Mean (SD)	n
65–69	0.25 (0.14)	3104	0.28 (0.13)	2348
70–74	0.28 (0.14)	3002	0.29 (0.13)	2295
75–79	0.29 (0.14)	2589	0.31 (0.13)	2100
80–84	0.31 (0.14)	1977	0.32 (0.13)	1485
85–90	0.32 (0.14)	1203	0.33 (0.13)	871

Frailty Index and Time to Death in the 2013 Cohort

Among 20,974, 18% (n=3842) died over the 12-month observation period, whereas 1% (n=226) were censored for transplantation. In a model adjusted for age and sex, we found each 0.1 point increase in frailty index was associated with a 41% increased hazard of death (hazard ratio [HR], 1.41; 95% CI, 1.37 to 1.44) (Table 3) (effect estimates for all covariates is shown in Supplemental Table 3). C-statistics for the frailty index, the comorbidity index, and the model including both indices were 0.65, 0.65, and 0.66, respectively (Table 3).

Table 3. - Effect estimates and Harrell’s C for frailty index and comorbidity models of time to death and time to first hospitalization in 2013 and 2017 cohorts

Outcomes	Frailty Index a		Comorbidity Index b		Harrell’s C-Statistic
	2013	2017	2013	2017	2013	2017
Time to death, d
Frailty index alone	1.41 (1.37 to 1.44)	1.45 (1.41 to 1.49)	—	—	0.65	0.64
Comorbidity index alone	—	—	1.23 (1.21 to 1.25)	1.29 (1.27 to 1.31)	0.65	0.66
Comorbidity + frailty indices	1.24 (1.20 to 1.28)	1.21 (1.17 to 1.25)	1.14 (1.12 to 1.16)	1.20 (1.18, to1.23)	0.66	0.67
Time to first hospitalization, d
Frailty index alone	1.33 (1.31 to 1.35)	1.35 (1.33 to 1.38)	—	—	0.61	0.61
Comorbidity index alone	—	—	1.17 (1.16 to 1.18)	1.20 (1.19 to 1.21)	0.60	0.60
Comorbidity + frailty indices	1.23 (1.21 to 1.25)	1.22 (1.20 to 1.25)	1.08 (1.07 to 1.09)	1.11 (1.10 to 1.12)	0.61	0.62

All models adjusted for age and sex.

^aHazard ratios are reported for each 0.1 unit (10%) increase in frailty index.

^bHazard ratios are reported for each 2.1 unit (10%) increase in comorbidity index.

Frailty Index and Time to First Hospitalization in the 2013 Cohort

Among 20,974, 55% (n=11,493) experienced at least one hospitalization over the 12-month observation period. In a model adjusted for age and sex, we found each 0.1-point increase in frailty index was associated with a 33% increased hazard of hospitalization (HR, 1.33; 95% CI, 1.31 to 1.35) (Table 3). C-statistics for the frailty index, the comorbidity index, and the model including both indices were 0.61, 0.60, and 0.61, respectively (Table 3).

Frailty Index and Length of Stay in the 2013 Cohort

Among those who experienced at least one hospitalization during follow-up (n=11,493), the median (Q1, Q3) length of stay was 5 (4, 9) days. In a model adjusted for age and sex, we found each 0.1-point increase in frailty index was associated with a 6% increase in average length of stay (Rate Ratio, 1.06; 95% CI, 1.04 to 1.07] (Table 4).

Table 4. - Ratio of average days in the hospital for frailty index and comorbidity models of first hospitalization length of stay in 2013 and 2017 cohorts

Length of Stay	Frailty Index a		Comorbidity Index b
	2013	2017	2013	2017
Frailty index alone	1.06 (1.04 to 1.07)	1.04 (1.03 to 1.06)	—	—
Comorbidity index alone	—	—	1.02 (1.02 to 1.03)	1.020 (1.01 to 1.03)
Comorbidity + frailty indices	1.05 (1.03 to 1.07)	1.04 (1.02 to 1.06)	1.01 (1.00 to 1.02)	1.00 (1.00 to 1.01)

All models adjusted for age and sex.

^aRatios are reported for each 0.1 unit (10%) increase in frailty index.

^bRatios are reported for each 2.1 unit (10%) increase in comorbidity index.

Findings in the 2017 Validation Cohort

In the 12-month observation period for the 2017 cohort, 17% (n=3687) died, 53% (n=11,284) experienced at least one hospitalization, and 1% (n=247) were censored for transplantation. For prediction of death, the C-statistics for the frailty index, the comorbidity index, and the model including both indices were 0.64, 0.66, and 0.67, respectively (Table 3). For prediction of first hospitalization, C-statistic values for the frailty index, the comorbidity index, and the model including both indices were 0.61, 0.60, and 0.62, respectively. Length of stay analyses were similar to the 2013 cohort (Table 4).

Adverse Outcomes and Deficits Stratified by Frailty Index Categories

Using a locally estimated scatterplot smoothing curve for frailty index and time to death in the 2013 cohort (Figure 5, Supplemental Figure 2), we identified three frailty index categories: 0–0.23, 0.24–0.36, and 0.37+. Having a frailty index of 0.24–0.36 or 0.37+ was associated with increased hazards for mortality and time to first hospitalization compared with the reference category (0–0.23) (Figure 6, Supplemental Figure 3). With adjustment for age, sex, and comorbidity, each successively higher frailty category was associated with significantly higher hazards of the outcome (Table 5). Figure 7 shows a subset of deficits present in ≥25% of the cohort and demonstrated the largest difference in proportion of individuals with the deficit (≥0.30) between the lowest (0–0.23) and highest (0.37+) frailty index categories. For example, the proportion with an ambulance claim increased from 6% in the lowest to 76% in the highest frailty index category. Supplemental Figure 4 shows all 53 deficits and their proportions within each frailty index category.

Figure 5.: A locally estimated scatterplot smoothing (LOESS) curve of frailty index and time to death identifies three frailty index categories in the 2013 cohort. The three frailty index categories identified by visual inspection were 0–0.23, 0.24–0.36, and 0.37 or higher.

Figure 6.: Kaplan–Meier curves in the 2013 cohort. (A) Kaplan–Meier curve for time to death by frailty index category. (B) Kaplan–Meier curve for time to first hospitalization by frailty index category. For both plots, the blue line representing the 0–0.23 group, red line representing the 0.24–0.36 group, and green line is the 0.37+ group (those with the highest frailty index values).

Table 5. - Hazard ratios of time to death and time to hospitalization by frailty index categories in 2013 and 2017 cohorts

Frailty Index Categories	Death Event, n (%)	Hazard Ratio (95% Confidence Interval)	Hospitalization Event, n (%)	Hazard Ratio (95% Confidence Interval)
2013 cohort
0–0.23 (n=7446)	795 (11%)	Ref.	2,912 (39%)	Ref.
0.24–0.36 (n=7217)	1252 (17%)	1.28 (1.16 to 1.41)	4,144 (57%)	1.43 (1.36 to 1.51)
0.37+ (n=6311)	1795 (28%)	1.81 (1.64 to 2.00)	4,437 (70%)	1.78 (1.68 to 1.88)
2017 cohort
0–0.23 (n=7592)	771 (10%)	Ref.	2,865 (38%)	Ref.
0.24–0.36 (n=7941)	1269 (16%)	1.16 (1.05 to 1.27)	4,368 (55%)	1.36 (1.29 to 1.43)
0.37+ (n=5822)	1647 (28%)	1.68 (1.51 to 1.86)	4,051 (70%)	1.74 (1.64 to 1.84)

All models were adjusted for age, sex, and Liu comorbidity index. Ref., reference.

Figure 7.: A subset of common frailty index deficits and proportion of patients with each deficit stratified by frailty index category. Deficits shown were present in ≤25% of individuals in the 2013 cohort.

Discussion

This study demonstrates the deficit accumulation index approach can be applied to administrative data from older adults who survived ≥6 months of dialysis for development of a frailty index. This approach to frailty assessment has potential application to any available dataset for research and/or data collected for usual care in clinical practice.

This is the first administrative data–based frailty index derived from a dialysis population. As in other studies, the index increases with age, is higher among women than men, and values peak at 0.7 (¹⁸,²¹), indicating no individual survives to possess all deficits. Similar to a claims-based frailty index, this frailty index is associated with mortality after adjusting for age, sex, and comorbidity (¹⁴). Few studies have explored a frailty index in the dialysis population. In one study using a laboratory-based frailty index (LFI) (index with deficits determined from abnormal values for BP, electrolytes, hemoglobin, and other laboratory values), older adults receiving hemodialysis with a high LFI had lower survival rates than those with low LFI (²²). A separate study developed a transplant waitlist–specific frailty index using medical record and survey responses among transplant candidates (²³). In models with age >70 years and sex adjustments, having a frailty index ≥0.25 was associated with increased risk for death and waitlist withdrawal (²⁴). Our study uniquely adds to the literature because this frailty index is developed from administrative data and in a distinct cohort of older adults who survived 6 months of dialysis.

Although our analyses demonstrate our frailty index had similar prediction performance as a comorbidity index, this frailty index uniquely includes a subset of variables that can change over time. For example, some frailty index deficits reflect recent acute events (e.g., brain imaging, ambulance use). However, the frailty index deficits did not include comorbidities commonly present at dialysis initiation (e.g., hypertension, diabetes). A comorbidity index typically only worsens because once a comorbidity is diagnosed, it is presumed too always be present. In contrast, the frailty index deficits extend beyond comorbidities and can capture dynamics in health status as frailty index improves or worsens (²⁵,²⁶). With this temporality, the frailty index can be assessed longitudinally within individuals. Relatedly, a frailty index is a potentially important metric to understand and mitigate incident frailty in patients of all ages receiving dialysis.

We acknowledge additional research is needed to enhance our frailty index for both research and clinical application. The frailty index’s risk prediction could be improved by adapting it to reflect the multidimensionality assumption inherent to deficit accumulation theory (¹⁸). Such adaptation would require the addition of deficits that represent additional types of data, such as laboratory values and symptoms, signs, and disability, which were not available for this analysis. Laboratory test abnormalities could enhance frailty assessment because they represent subclinical deficits that may precede clinically detectable diagnosis (²⁷,²⁸). With USRDS data, additions may be feasible using Consolidated Renal Operations in a Web-enabled Network data (¹⁶). Self-reported measures of symptoms, signs, and disability can also enhance frailty assessment, as demonstrated in frailty indices derived from comprehensive geriatric assessment data (²⁹). Further, an alternative approach to improve the predictive ability of the frailty index is to create a weighted sum of deficits, rather than a simple sum. However, weighting reduces generalizability. Still, weighting or machine learning algorithms should be considered when the intended use of the frailty index is to accurately predict an individual patient’s survival for clinical decision-making (⁸,¹⁴,³⁰). In addition to efforts to improve predictive ability, further evaluations to validate the index could include analyses to quantify the extent of index agreement with phenotypic frailty in a dialysis cohort, how the frailty index’s strength of association with mortality changes when included in models with other variables that are associated with mortality in the dialysis population (i.e., central venous catheter use, absence of predialysis nephrology care), longitudinal changes in frailty index, and potential modifiers of frailty index. Further optimization and evaluation of this frailty index would provide greater confidence in its use in studies as a potential confounder of treatment-outcome associations and for heterogeneity of treatment effects analyses (^31–33).

This study has clinical implications. Because the frailty index utilizes data that have already been collected (²¹), it has high potential to be used for population health management to screen for the most vulnerable patients. Such screening can be used to optimize resource allocation for both medical and population health services, such as comprehensive geriatric assessment and social services (³⁴,³⁵). Access to those services has the potential to limit functional impairment, which could delay or prevent long-term care. Compared with phenotypic frailty and frailty screening instruments (³⁶), a frailty index is ideal for the dialysis population where mobility impairment is common and resource constraints limit physical performance assessments (³⁷). To be utilized for frailty screening in real-time clinical interactions, this frailty index needs to be adapted for the dialysis electronic medical record. Practically, an electronic health record–based frailty index could allow for real-time frailty assessment for all patients irrespective of insurance status and be integrated in the electronic health record to alert providers of frailty index changes over time or a threshold. Such tools have been developed for patients with CKD (³⁸,³⁹); however, clinicians should use caution in applying a frailty index to predict future outcomes in individual patients.

Our study’s strength lies in our data source, USRDS, which provided the most comprehensive administrative data for the older dialysis population, and in our systematic approach to frailty index development (¹⁸). However, this study does have limitations. First, we acknowledge that use of diagnostic and procedure codes can change with coding trends and depend on clinician behavior. As a result, ascertainment bias is inherent in both the development and calculation of an individual’s frailty index. Efforts to minimize this bias would focus on including claims that are generally stable in use over time. Second, our data-driven approach to identifying deficits led to identification of administrative claims that are not conceptually consistent with being a deficit (e.g., influenza vaccination, electrocardiogram) or nonspecific (e.g., other Medicare fee schedule). Although some tests may be reasonable surrogates for health deficits (e.g., brain imaging for ischemic stroke) (⁴⁰), future work can apply a combination of data-driven and expert opinion to optimize selection of deficits. Third, the frailty index did not include objective physical or cognitive function from clinical assessment among potential deficits. Although administrative claims do not often reflect objective functional data found in medical records, efforts can be made to identify and evaluate surrogate variables and/or incorporate medical record data. Fourth, this frailty index was defined in community-dwelling older adults who survived 6 months of dialysis. There is survival bias inherent to this approach making this index appropriate for clinical and research purposes in the prevalent dialysis population, but not appropriate for older adults initiating dialysis, a group of patients with a high risk of death. As this study serves as a proof of concept, it sets the foundation for a separate frailty index capturing a unique set of deficits present at dialysis initiation. Fifth, this frailty index is not generalizable to younger adults, long-term care residents, or predialysis cohorts, in whom frailty is common. Future work will need to include studies to evaluate this frailty index in these subgroups and/or adapt it for universal applicability. Finally, we did not compare this novel frailty index to other existing frailty indices. It is possible other frailty indices could also be useful in dialysis populations.

In sum, we have developed an administrative data–based frailty index for prevalent community-dwelling older adults receiving dialysis. The frailty index predicts mortality and hospitalization within 12 months of frailty assessment, similar to a dialysis-specific comorbidity index. These findings demonstrate potential application for inclusion of frailty in secondary data analyses and practical frailty screening for dialysis clinics. Through such applications, the frailty index could improve identification and management of the most vulnerable among the dialysis population.

Disclosures

C. Colón-Emeric reports having consultancy agreements and an advisory or leadership role with Amgen and Novartis; reports receiving research funding from UCB Pharma; and reports having patents or royalties as Co-inventor 2 use patents for bisphonate indication in cardiovascular diseases. D.H. Kim reports having consultancy agreements with Alosa Health and VillageMD. J.F. Pendergast reports receiving research funding from Dialysis Centers, Inc.; and reports receiving honoraria from the National Institutes of Health National Institute on Aging/National Institute of Mental Health Advanced Research Institute for training junior scholars to get their first R01 (US$2000 for 3 days of mentoring). J.J. Scialla reports having an advisory or leadership role as Deputy Editor, American Journal of Kidney Diseases. L.E. Boulware reports having an advisory or leadership role with the Association for Clinical and Translational Science, Journal of the American Medical Association Editorial Board, Journal of the American Medical Association Network Online Editorial Board, and the Robert Wood Johnson Clinical Scholars National Advisory Committee. P.L. Ephraim reports having consultancy agreements with Stony Run Consulting. R. Hall reports having consultancy agreements with Bayer, Reata Pharmaceuticals, Otsuka, Travere Pharmaceuticals, and United Health Group; and reports having an advisory or leadership role with the CJASN Editorial Board and the Journal of the American Geriatrics Society Editorial Board. All remaining authors have nothing to disclose.

Funding

This study was supported by the National Center for Advancing Translational Sciences under grant UL1TR002553, National Institute on Aging grants P30AG028716, K76AG059930, R01AG056368, R01AG062713, R01AG071809, and R21AG060227, National Institute of Diabetes and Digestive and Kidney Diseases grant R01DK111952, and the American Society of Nephrology Foundation for Kidney Research.