Domains for a Comprehensive Geriatric Assessment of Older Adults with Chronic Kidney Disease: Results from the CRIC Study

Chronic kidney disease (CKD) is prevalent among older adults, affecting over 38% of those 65 years and older [1]. Such individuals are at high risk of comorbidities and declines in various domains, such as frailty [2, 3], cognitive impairment [3‒5], and poor quality of life [6]. These medical, social, and functional conditions may independently or interactively result in downstream adverse health outcomes, including progression to end-stage kidney disease and mortality [7‒9].

A comprehensive geriatric assessment (CGA) can be used to identify multidimensional needs unique to older patients and help inform multidisciplinary, coordinated care plans to improve outcomes [10‒13]. Implementation of CGAs in older CKD patients helps guide care processes, improve treatment satisfaction, and decrease distress [14‒16]. Previous studies have applied generic CGAs to patients with CKD [17, 18]. However, generic CGAs are not designed for older patients with CKD and do not include geriatric conditions that frequently coexist with CKD. For example, frailty can develop at younger ages (i.e., ≤65 years) [19] and is more common among older adults with CKD than those with normal kidney function [2]. Similarly, kidney disease-specific quality of life is predictive of adverse health outcomes but is not included in generic CGAs [20, 21]. Given the nuances of aging with CKD, finding the right elements to include in the assessment would be the essential first step to building a CKD-specific CGA (CKD-CGA). In combination with the development of an integrated care plan, the CKD-CGA may yield a more practical and timely approach to assessment and care [22]. Using data from 864 community-dwelling older participants (≥55 years) from the Chronic Renal Insufficiency Cohort (CRIC), we 1) identified domains of a CKD-CGA, 2) characterized patterns of these domains in older adults with CKD, and 3) tested the predictive utility of the patterns of domains on adverse health outcomes.

The CRIC is an ongoing, multicenter prospective cohort study for examining risk factors, etiology, diagnosis, and outcomes of adults with CKD [23, 24]. Between 2003 and 2008, 3,939 participants between 21 and 74 years old with mild to moderate CKD were recruited from seven clinical centers in the USA. Annual in-person follow-up data were available through May 2020.

We selected participants enrolled in the Physical Performance Ancillary Study of the CRIC (parent study). The cross-sectional ancillary study was conducted in four centers from April 2008 to February 2010. The ancillary study visit is the baseline visit for the current study unless otherwise noted. Of 1,152 participants enrolled, we excluded 253 participants aged <55 years and 35 individuals who had initiated dialysis at baseline (shown in online suppl. Fig. 1; for all online suppl. material, see www.karger.com/doi/10.1159/000528602). The study protocol was approved by the Institutional Review Boards at each site. All participants provided written informed consent.

Potential domains for the CKD-CGA were selected a priori based on clinical expertise, the literature, and the availability of measures in the CRIC. Seventeen domains were considered: cardiovascular disease, hypertension, diabetes, five components of the physical frailty phenotype [25] (Fried’s physical frailty phenotypes: weakness, exhaustion, slowness, low physical activity, and weight loss), depressive symptoms (Becks Depression Inventory [26]), cognition (Modified Mini-Mental State [27]), five kidney disease quality-of-life components (Kidney Disease Quality of Life-36 [28]: burden, effects, and symptoms of kidney disease and physical and mental component summaries), health literacy (Short Test of Functional Health Literacy in Adults [29]), and medication use (potentially inappropriate medication identified according to 2015 American Geriatric Society Beers Criteria [30]). Each domain was categorized into a binary indicator. Clinically relevant cutoffs for having a geriatric condition were used. If no clinically recognized cutoffs exist, we used the 20th percentile [25]. We used data collected at baseline to assess each geriatric domain. The only exception was health literacy, which was measured at baseline of the CRIC study, and values were assumed to be constant throughout follow-up. Table 1 presents details on considered domains, timing of data collection, description of instruments used, and chosen cutoffs for having a geriatric condition.

The primary outcome of interest for evaluating the predictive utility of the patterns of domains was all-cause mortality. Death was ascertained using linkage with the Social Security Death Master File, retrieval of death certificates or obituaries, review of hospital records, or reports from next-of-kin. Time to mortality was measured in years from baseline.

Our secondary outcomes of interest were dialysis initiation and all-cause hospitalization. Initiation of dialysis was determined by self-report, records from local clinical centers, and linkage with the US Renal Data System at each annual visit. Time to dialysis initiation was measured in years from baseline. The number of hospitalizations in the last 12 months was ascertained at each annual visit. We used data obtained in the year after the baseline visit to predict the risk of hospitalization within 12 months of assessing the domains we identified for the CKD-CGA. Hospitalization was treated as a binary variable; participants either had or did not have any hospitalization.

Sex, race, marital status, education, and household income were measured at baseline of the parent study. Age, smoking status, body mass index (BMI), and estimated glomerular filtration rate (eGFR; using the Chronic Kidney Disease Epidemiology Collaboration equation [31]) were updated annually; values at baseline of the ancillary study were used.

Descriptive statistics, overall and stratified by age groups (55–64 vs. ≥65), were used to describe the study population at baseline. Categorical variables were reported in counts and proportion; comparisons between groups were conducted using χ² tests. Continuous variables were reported in means and standard deviation; comparisons between groups were made using t tests.

Latent class analysis (LCA) is a parametric approach used to empirically identify underlying, distinct subgroups (or classes) based on patterns of binary indicators. LCA was applied to 17 binary indicators of potential geriatric domains. We ran a series of 1- to 5-class models to identify the solution that fit the data best. LCA assumes conditional independence; the indicators are independent of one another because any relationships are fully explained by latent class membership. We assessed this assumption by scrutinizing the standardized bivariate residuals [32].

We estimated the prevalence of class membership, as well as the prevalence of each indicator conditioned on class membership (i.e., conditional probabilities). To evaluate the optimal number of latent classes, we compared models based on four fit indices: Akaike Information Criterion (AIC) [33], Bayesian Information Criterion (BIC) [34], Lo-Mendell-Rubin-adjusted likelihood ratio test (LMR LRT) [35], and bootstrap likelihood ratio test (BLRT) [36]. For AIC and BIC, lower values indicate better fit. LMR LRT and BLRT compare the fit of a model (k-class) to a smaller model (k-1 class); a nonsignificant χ² test (p > 0.05) suggests a comparable fit between the two models. Entropy was estimated to evaluate classification error [37]. The index ranges from 0 to 1, with higher entropy suggesting less error in classifying individuals. Findings of more than one latent class, in which the conditional probabilities of indicators are homogeneous in each class but show separation between classes, would be consistent with evidence for a syndromic nature of CKD-related geriatric conditions [38].

LCA aided the final selection of domains in the CKD-CGA. Comparisons of the difference in conditional probabilities of each geriatric condition across classes were done using Cohen’s d, an effect size measure of the difference between two groups [39]. Effect sizes were categorized as trivial (d < 0.2), small-to-medium (0.2 ≤ d < 0.5), medium-to-large (0.5 ≤ d < 0.8), or large (d > 0.8). Hypertension had the smallest effect size relative to the other geriatric conditions (d = 0.29), suggesting that the domain did not discriminate as well between classes. Therefore, it was removed from the final list of domains to include in the CKD-CGA. LCA was re-estimated with the remaining 16 domains.

For all time-to-event analysis, participants who did not experience an event were censored at loss to follow-up or administratively in May 2020. Cox proportional hazard modeling was conducted to examine time to mortality and dialysis initiation. Violation of the proportional hazard assumption was tested for all Cox proportional hazard models. Logistic regression modeling was conducted to examine hospitalization. For latent class regression (LCR; using latent class membership as the predictor), we manually conducted Vermunt’s three-step approach to account for classification error and to prevent the outcome from influencing class membership [40, 41]. We evaluated model discrimination, the ability of the models to assign a higher probability of outcomes to participants who have the outcome, using concordance statistics (C-statistics) [42, 43]. C-statistics range from 0.5 to 1, with 0.5 indicating model prediction is no better than chance and 1 indicating perfect prediction. C-statistics were calculated from Cox proportional hazard and logistic regression models that used the most-likely posterior class membership – the class individuals had the highest probability to be in given their pattern of geriatric conditions – as the exposure variable [44]. All models were adjusted for age, sex, race, BMI, eGFR, and smoking status.

There was a high proportion of missing values for depressive symptoms (53.1%) and cognition (38.2%) at baseline because these measures were collected biannually. Given the longitudinal nature of the CRIC study, we conducted linear interpolation imputation [45], wherein missing values were imputed by averaging the values from the visit before and after the ancillary study visit, for participants missing values at baseline. Domains other than depressive symptoms and cognition had missingness at baseline ranging from 0% to 7%. This missingness was assumed missing at random conditional on other variables in the model. The full informational maximum likelihood estimator accounted for missing values in LCA and LCR. Multiple imputation was implemented using chained equations (MICE) for multivariable models used to calculate C-statistics [46, 47].

In the year after baseline, 35 participants were missing hospitalization data due to loss to follow-up. LCR in the main analysis accounted for missing values. However, the C-statistics calculation from the multivariable logistic regression models using most-likely posterior class membership as the exposure variable was conducted without these participants. As a sensitivity analysis, we included participants who were previously excluded by using their hospitalization data from the previous year. LCA and LCR were conducted using Mplus 8.6 [48]; other analyses were conducted using Stata 16.1 [49].

The sample included 864 participants followed for up to 12.0 years. The median age was 66.7 years (range 55–80 years). Thirty-nine percent were 55–64 years old, and 61% were ≥65 years; 47.1% were women, 58.7% were white, and 35.4% were black (Table 2). Study participants had a median of 3 geriatric conditions. In comparison to younger participants (age 55–64 years), participants aged ≥65 years had lower household income and eGFR and higher cardiovascular disease and frailty phenotypes: weakness, slowness, and low physical activity.

According to model fit indices (Table 3), AIC was the lowest in a 5-class model, while BIC was the lowest in the 3-class model. However, the largest drop in both AIC and BIC was from the 1-class to the 2-class model, favoring a 2-class solution. The BLRT did not help in determining the optimal class solution. The LMR LRT suggested that the 2-class model fit significantly better than the 1-class model (p < 0.001) but that a 3-class model was no better than a 2-class model (p = 0.51). Together, along with considerations of the interpretability and the practicality of the solutions, a 2-class model was selected.

The first class comprised 34.7% of the sample (shown in Fig. 1). Conditional probabilities of having geriatric conditions in the class were moderate to high (mostly ranging from 26% to 69%; Table 4). We named this class the “high-burden of geriatric conditions” class. The second class labeled the “low-burden of geriatric conditions” class comprised the remaining 65.3% of the sample. Conditional probabilities of having geriatric conditions were low to moderate (mostly ranging from 1% to 27%; Table 4). Both the high- and the low-burden classes, respectively, have a relatively low conditional probability of weight loss (12% vs. 3%), cognitive impairment (7% vs. 1%), and limited health literacy (13% vs. 4%) and relatively high conditional probability of diabetes (60% vs. 36%) and use of potentially inappropriate medication (69% vs. 52%).

The median number of years of follow-up was 10.9 years for mortality and 9.0 years for dialysis initiation. Of 864 participants, 271 died (31.4%) and 137 initiated dialysis (15.9%); 271 individuals (of 829; 32.7%) were hospitalized at least once within 12 months of baseline. The cumulative incidence of death in the high-burden of geriatric conditions class was consistently higher than that of the low-burden of geriatric conditions class throughout follow-up (shown in Fig. 2). At 1, 5, and 10 years of follow-up, the unadjusted cumulative incidence of death was 2.7%, 19.3%, and 42.4% in the high-burden of geriatric conditions class and 0.9%, 8.6%, and 21.3% in the low-burden of geriatric conditions class, respectively (Table 5). The cumulative incidence of dialysis initiation followed a similar trend: 2.8%, 18.0%, and 30.0% in the high-burden class and 1.2%, 7.1%, and 13.7% in the low-burden class at 1, 5, and 10 years, respectively.

After adjusting for age, sex, race, eGFR, BMI, and smoking status, the high-burden of geriatric conditions class had a 2.09-fold increased hazard of death (95% CI: 1.56, 2.78, p < 0.001), a 1.63-fold increased hazard of dialysis initiation (95% CI: 1.06, 2.52, p = 0.028), and a 2.00-fold increased odds of being hospitalized within 12 months of the geriatric assessment (95% CI: 1.38, 2.88, p < 0.001) compared to the low-burden of geriatric conditions class (Table 5). Model discrimination was excellent for time to dialysis initiation (C-statistics = 0.86) and moderate for time to mortality and hospitalization (C-statistics = 0.70 and 0.66, respectively). Predicting outcomes with latent class membership had similar or slightly increased discrimination compared to a model that only included the control variables. Including all 16 geriatric domains in a model, after adjustment, provided the largest increase in discriminatory power (online suppl. Table 1).

Replacing missing values of hospitalization 12 months from baseline using hospitalization 12 months prior to baseline resulted in a stronger association but similar inference as the main analysis (aOR = 2.20, 95% CI: 1.54, 3.14, p < 0.001). Model discrimination remained the same as the main analysis (C-statistics = 0.66).

Using data from the CRIC, we identified 16 domains for a CKD-CGA for adults aged 55 years and older with all stages of CKD. LCA indicated that the clustering of these geriatric conditions was syndromic in nature, tending to manifest together. Thirty-five percent of the participants had a high burden of geriatric conditions and were more likely to experience future adverse health outcomes. Specifically, these participants had a 2.09-fold (95% CI: 1.56, 2.78) increased risk of experiencing death, a 1.63-fold (95% CI: 1.06, 2.52) increased risk of dialysis initiation, and a 2.00-fold (95% CI: 1.38, 2.88) increased risk of hospitalization within the next year compared to participants with a low burden of geriatric conditions. Grouping by burdens of geriatric condition was good at predicting dialysis initiation (C-statistics = 0.86) and fair at predicting mortality and hospitalization (C-statistics = 0.70 and 0.66, respectively).

The number of domains selected for CKD-specific CGAs, instruments used to measure the domains, and cutoffs that define the geriatric conditions differ across studies. Therefore, direct comparisons with other studies are limited. Generally, previous studies found a high prevalence of geriatric conditions among older adults with CKD [14, 50‒52]. This is consistent with our study which found that over a third of older participants experienced a high burden of geriatric conditions. Given the sample composition, our study further reveals that even younger CKD patients (below age 65 years) could experience a high burden of geriatric conditions. Of note, in our study, the prevalence of having depressive symptoms, low quality of life components, and the frailty phenotype exhaustion was especially different between the high-burden of geriatric conditions class compared to the low-burden of geriatric conditions class (42–50% vs. 2–6%, respectively). This suggests that these domains were particularly important in distinguishing individuals between the classes. Like many studies that aimed to identify domains in a CKD-specific CGA, ours included domains on chronic conditions, cognition, depressive symptoms, and frailty. We additionally included less commonly incorporated domains that may more broadly capture the nuances of aging with CKD, such as quality of life, health literacy, and medication use. Declines and problems in these domains have been associated with CKD and subsequent adverse health outcomes [53‒58]. Including these additional domains allowed our CKD-CGA to provide a more holistic assessment of geriatric conditions in the community-living older CKD population.

No other studies of CKD-CGAs of which we are aware have tested the predictive utility of different patterns of domains. One study of a multidimensional prognostic index derived from a generic eight-domain CGA reported that increased index scores were associated with short-term all-cause mortality in hospitalized older patients with CKD [17, 18]. In using the domains we identified for the CKD-CGA, we found that having a high burden of geriatric conditions is predictive of all-cause mortality and dialysis initiation up to 12.0 years and hospitalization within 1 year after assessment. Our study extends upon the current literature, suggesting there may be both short-term and long-term consequences of having multiple geriatric conditions in the CKD population. Additionally, incorporating more geriatric domains would provide a stronger prediction of adverse health outcomes than individual domains alone.

Our study has strengths. To our knowledge, this is the first study to empirically identify the domains of a CKD-CGA using psychometric methods in conjunction with a priori knowledge, to minimize potential bias associated with the subjective selection of domains [59]. The domains we selected for the CKD-CGA are designed to assess older adults at all stages of CKD, improving upon existing CKD-specific CGAs which often only targeted patients in advanced stages of CKD [16, 50, 59]. Further, we leveraged long-term follow-up data of over a decade to assess the predictive utility of different patterns of domains. This study also has limitations. The a priori selected domains for the CKD-CGA were restricted by data availability in the CRIC study. For example, (instrumental) activities of daily living were not collected in the study but are often considered an essential component of a CGA, given the importance of maintaining functional independence [11, 60]. In our data, cognition and depressive symptom measures had a high proportion of missing values due to study design. To address this, we conducted imputation and confirmed that participants with versus without missing data did not differ in sociodemographic characteristics. Another limitation is that the participants in the study may not represent a generalizable sample of the CKD population in the USA. The cross-sectional ancillary study from which we selected our study sample recruited participants with varying follow-up lengths since enrollment into the parent study. Individuals who remained in the parent study and subsequently enrolled in the ancillary study may be healthier. Coupled with the fact that our sample also consists of younger and potentially healthier older adults, our results are likely an underestimation of the effect of these conditions on adverse outcomes.

Our findings have strong clinical implications for nephrology and geriatric practice. The association of the burden of geriatric conditions with adverse health outcomes provides evidence for the value of assessing for geriatric conditions in older adults at all stages of CKD. Our study suggests that those aged ≥55 years have a high burden of geriatric conditions and could benefit from receiving the CKD-CGA early before they become further impaired. Early identification of geriatric conditions has the potential to help initiate discussion on and guide individualized care plans, delay or prevent CKD progression, and ensure maintenance of independence and quality of life [14, 16, 50]. By identifying domains to include in a CKD-CGA, our study has only taken the first step of developing a CGA. The next step in developing the CKD-CGA would be applying the assessment in a clinical and/or research setting along with an integrated care plan. One potential disadvantage in using the CKD-CGA is that implementing CGAs is a multidisciplinary process that can be time-consuming, labor-intensive, and costly. However, a recent qualitative study with 47 health care providers suggested that patients predominantly view the time and content of assessment positively [15]. Providers appreciated the multidimensional approach to care provision [15]. Nevertheless, future research should examine the feasibility, effectiveness, and impact of incorporating the CKD-CGA into nephrology practice. Further validation is crucial in confirming the utility and potential of the CKD-CGA in elevating CKD care in the older population.

We identified 16 domains for a CKD-CGA for older adults of all stages of CKD using psychometric analyses. We identified a class of participants with a relatively high probability of experiencing most geriatric conditions, which has a higher risk of mortality, dialysis initiation, and hospitalization. With further validation in an external cohort, the CKD-CGA has the potential to be used in nephrology practices for assessing geriatric conditions, informing decision-making, and improving health outcomes in older adults with CKD.

This study protocol was reviewed, and the need for approval was waived by the Johns Hopkins Bloomberg School of Public Health Institutional Review Board. All participants provided written informed consent.

The authors have no conflicts of interest to declare.

Funding for this project was obtained through the CRIC study Opportunity Pool Program. Funding for the CRIC study was obtained under a cooperative agreement from National Institute of Diabetes and Digestive and Kidney Diseases (U01DK060990, U01DK060984, U01DK061022, U01DK061021, U01DK061028, U01DK060980, U01DK060963, U01DK060902, and U24DK060990). In addition, this work was supported in part by the Perelman School of Medicine at the University of Pennsylvania Clinical and Translational Science Award NIH/NCATS UL1TR000003, Johns Hopkins University UL1 TR-000424, University of Maryland GCRC M01 RR-16500, Clinical and Translational Science Collaborative of Cleveland, UL1TR000439 from the National Center for Advancing Translational Sciences (NCATS) component of the National Institutes of Health and NIH roadmap for Medical Research, Michigan Institute for Clinical and Health Research (MICHR) UL1TR000433, University of Illinois at Chicago CTSA UL1RR029879, Tulane COBRE for Clinical and Translational Research in Cardiometabolic Diseases P20 GM109036, Kaiser Permanente NIH/NCRR UCSF-CTSI UL1 RR-024131, Department of Internal Medicine, University of New Mexico School of Medicine, Albuquerque, NM R01DK119199. Dr. McAdams-DeMarco was supported by R01AG055781 and R01DK114074. Dr. Rasheeda K. Hall was supported by K76AG059930. Dr. Nadia Chu was supported by K01AG064040.

Research idea and study design: Mara McAdams-DeMarco, Rasheeda K. Hall, and Dorry Segev; data analysis/interpretation: Venus Chiu, Mara McAdams-DeMarco, Alden L. Gross, Nadia M. Chu, Dorry Segev, and Rasheeda K. Hall; statistical analysis: Venus Chiu; supervision or mentorship: Mara McAdams-DeMarco, Alden L. Gross, and Nadia M. Chu. Each author contributed important intellectual content during manuscript drafting and accepts accountability for the overall work by ensuring that questions pertaining to the accuracy or integrity of any portion of the work are appropriately investigated and resolved. All the authors take responsibility that this study has been reported honestly, accurately, and transparently.

The data supporting the findings of this study are available in the NIDDK repository at https://repository.niddk.nih.gov/studies/cric/. Further enquiries can be directed to the corresponding author.

The CRIC study investigators include Lawrence J. Appel, MD, MPH; Jing Chen, MD, MMSc, MSc; Harold I. Feldman, MD, MSCE; Alan S. Go, MD; James P. Lash, MD; Robert G. Nelson, MD, PhD, MS; Mahboob Rahman, MD; Panduranga S. Rao, MD; Vallabh O Shah, PhD, MS; Raymond R. Townsend, MD; and Mark L. Unruh, MD, MS.