Comparing Survey-Based Frailty Assessment to Medicare Claims in Predicting Health Outcomes and Utilization in Medicare Beneficiaries

Abstract

Objectives:

To assess two models for the prediction of health utilization and functions using standardized in-person assessments of frailty and administrative claims-based geriatric risk measures among Medicare fee-for-service beneficiaries aged 65 years and above.

Methods:

Outcomes of hospitalizations, death, and functional help were investigated for participants in the 2011 National Health and Aging Trends Study. For each outcome, multivariable logistic regression model was used to investigate claims-based geriatric risk and survey-based frailty.

Results:

Both claims-based and survey-based models showed moderate discrimination. The c-statistic of the standardized frailty models ranged from 0.67 (for any hospitalization) to 0.84 (for any IADL [instrumental activities of daily living] help). Models using administrative data ranged from 0.71 (for any hospitalization) to 0.81 (for any IADL help).

Discussion:

Models based on existing administrative data appear to be as discriminate as survey-based models. Health care providers and insurance plans can effectively apply existing data resources to help identify high-risk individuals for potential care management interventions.

Introduction

Identifying older individuals at high risk for poor health outcomes and high expenditures is an important goal from a clinical and health system perspective. Potential benefits include the establishment of organizational strategies and services that can help optimize care for older adults. Several approaches for characterizing those at highest risk have emerged in recent years. These include the frailty phenotype, the accumulation of deficits model, and the geriatric syndrome (Fried et al., 2001; Rockwood & Mitnitski, 2007; Tinetti, Inouye, Gill, & Doucette, 1995).

Frailty is a complex clinical manifestation that often co-occurs with comorbidities and disability but can exist independently (Fried, Ferrucci, Darer, Williamson, & Anderson, 2004; Lau, Kwan, & Cheung, 2016). About 15% of adults aged 65 and older are considered frail, and frailty is associated with high risk of adverse events such as hospitalizations, mortality, and institutionalization (Bandeen-Roche et al., 2015; Drubbel et al., 2013; Kelaiditi et al., 2016; Moore, White, Washington, Coenen, & Elixhauser, 2017; Pugh et al., 2014; Rosenberg, 2012; Sirois et al., 2017; Woo, Leung, & Morley, 2012; Zaslavsky et al., 2016; Zaslavsky et al., 2017). Alternative approaches to characterizing frailty have been developed. Fried et al. (2001) view frailty as a phenotype measured by five indicators focused on physical frailty. Rockwood describes frailty as an accumulation of deficits focused on the assumption that an individual will be more frail when more things are wrong (Rockwood & Mitnitski, 2007).

The concept of “geriatric syndrome” or “geriatric conditions” also has emerged to capture the multifactorial and heterogeneous features that are prevalent in older people and do not always fit within discrete clinical diagnoses (Ahmed, Mandel, & Fain, 2007; Cigolle, Langa, Kabeto, Tian, & Blaum, 2007; Lee, Cigolle, & Blaum, 2009; Tinetti et al., 1995). These constructs share most risk factors with frailty and are hypothesized to result in a cascade of additional risk factors, leading to final outcomes of dependence and death. These factors are prevalent in frail older adults, they affect quality of life and disability, and they have multiple underlying risk factors involving multiple organ systems (Inouye, Studenski, Tinetti, & Kuchel, 2007; Kane, Shamliyan, Talley, & Pacala, 2012; Koroukian et al., 2016; Schiltz et al., 2017; Wang, Shamliyan, Talley, Ramakrishnan, & Kane, 2013). The literature, both from a biological and clinical perspective, continues to explore factors causing and/or associated with frailty (Chamberlain et al., 2016; Clegg, Young, Iliffe, Rikkert, & Rockwood, 2013; Inouye et al., 2007; Kane et al., 2012; Sternberg, Wershof Schwartz, Karunananthan, Bergman, & Mark Clarfield, 2011; Walston et al., 2006; Xue, 2011).

Widely used approaches for measuring frailty require individual assessment of key indicators (Fried et al., 2001; Rockwood & Mitnitski, 2007). Accordingly, the majority of previous studies examining the association between frailty and its related risk factors on health care utilization and health-related outcomes have derived frailty measures from patient/respondent surveys or cohort studies (Basic & Shanley, 2015; Carey et al., 2008; Cigolle et al., 2007; Di Bari et al., 2014; Lee et al., 2009; Pollack et al., 2017; Schiltz et al., 2017; Theou, Brothers, Mitnitski, & Rockwood, 2013; Zaslavsky et al., 2016; Zaslavsky et al., 2017). Although some specialized long-term care support service initiatives, such as the Program of All-Inclusive Care for the Elderly (PACE), conduct frailty assessments on a routine basis, such assessments are not common in other health care settings for older populations, despite the fact that more than 50% of Americans age 65 and older meet criteria as frail or pre-frail (Bandeen-Roche et al., 2015; Hirth, Baskins, & Dever-Bumba, 2009; Saucier, Burwell, & Gerst, 2005). As a result, these specialized evaluations may reach only a small subset of those at risk. Use of administrative claims data from outpatient and inpatient services could provide a useful and inexpensive alternative of identifying and managing frail individuals, if a crosswalk between claims and frailty assessment were developed.

Previous studies have attempted to measure frailty through claims data. Both Kim and Segal developed a claims-based frailty index using a range of clinical diagnoses (Kim et al., 2017; Segal et al., 2017). Using a deficit accumulation model, Kim demonstrated that durable medical equipment claims, diagnoses for degenerative disease of the central nervous system, cardiovascular diseases, and cerebrovascular diseases were most strongly associated with the deficit accumulation index (Kim et al., 2017). In contrast, Segal used the frailty phenotype model as the standard reference to find associated conditions (Segal et al., 2017). We expand on these studies by examining 10 claims-based geriatric risk measures that may be more reflective of conditions in need of care management interventions, and which are, in most cases, less dependent on specific individual disease diagnoses and have been shown to be highly predictive of poor outcomes among older adults (Kan et al., 2018). To the extent feasible, most of these risk measures reflect conditions that are aligned with those included under the geriatric syndrome construct.

The first objective of this article was to examine the relationship between this newly developed claims-based measures of geriatric risk, which is more focused on care management intervention concepts than previous measures, and a more standard measure of frailty within a nationally representative sample of Medicare beneficiaries aged 65 and older. The standard measure of frailty is a modified version of the Fried frailty phenotype. The secondary objective was to compare these two alternative risk measures (independently and combined) on the association with health utilization and outcomes pertinent to a geriatric population with special needs.

Method

Data Source

Data are from the 2011 survey and the 2010–2012 linked Medicare claims (inpatient, outpatient, carrier, skilled nursing, and home health files) for participants in the National Health and Aging Trends Study (NHATS), a nationally representative sample of Medicare beneficiaries aged 65 and older in the United States drawn from Medicare enrollment files (Kasper & Freedman, 2017). Annual, face-to-face interviews are conducted by trained surveyors who collect detailed information on physical and cognitive capacity, functional status, demographic, and other characteristics. The response rate for 2011 was 71% (Kasper & Freedman, 2017).

Sample

To construct claims-based risk measures and be included in our study, NHATS participants had to be continuously enrolled in Medicare fee-for-service Parts A and B from January 1, 2010, through December 31, 2012, or until their death, whichever was earlier. They also needed to complete an inperson interview in 2011 (spanning from May to October) to obtain the information used to assess frailty status. We excluded persons living in nursing homes in 2011 who were not eligible for an in-person interview. This yielded an analytic sample of 4,457 Medicare beneficiaries (see Appendix B for further sample description).

Frailty Indicator

The Fried frailty phenotype was constructed using five adapted criteria to describe frailty: exhaustion, low physical activity, shrinking, low walking speed, and weakness (Bandeen-Roche et al., 2015). Exhaustion and low physical activity were self-reported measures; participants met the criteria if they reported having low energy or never walked for exercise or engaged in vigorous activities, respectively. Shrinking was derived from either self-report of unintentional weight loss of 10 or more pounds or body mass index (BMI) less than 18.5 kg/m². Low walking speed and weakness were measured activities of usual-pace walking trials and maximum dominant hand grip strength. Participants met the criteria if walking trials were at or below the 20th percentile of the weighted population distribution within four sex-by-height categories or grip strength was at or below the 20th percentile within eight sex-by-BMI categories, respectively. An individual was classified as frail if ≥3 criteria were met, pre-frail if 1 or 2 criteria were met, and robust if none of the criteria were met based on the distribution (Appendix A). Additional details of these measures can be found in the original paper (Bandeen-Roche et al., 2015).

Geriatric Risk Measures Derived From Medicare Administrative Data

We proposed and operationalized 10 factors related to frailty and commonly used in definitions and measurements of geriatric risk (Inouye et al., 2007; Kane et al., 2012; Sternberg et al., 2011; Wang et al., 2013). We limited our measures to those that are potentially documented by clinicians in electronic health records (EHRs) or administrative records using International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) codes. The measures used were an extension of the “frailty” marker in the Johns Hopkins University ACG® system (Version 11), which have been demonstrated to predict adverse outcomes in geriatric populations (McIsaac, Bryson, & van Walraven, 2016; Sternberg et al., 2012). This earlier tool was originally developed to identify conditions related to factors negatively impacting activities of daily living (ADL) in Medicare patients (Weiner et al., 1996). To update this tool, we examined the literature and measures were furthered refined by clinicians and geriatricians on our team who manage frail older adults with multiple chronic conditions and/or disability.

The 10 current geriatric risk measures in our model include walking difficulty, history of falls, dementia/cognitive impairment, vision impairment, urinary incontinence, fecal incontinence, presence of decubitus/pressure ulcer, malnutrition, weight loss, and lack of social support. These 10 individual risk factors were identified as present or absent (i.e., one occurrence) based on ICD-9-CM codes in claims for the period starting from 2010 until the survey-interview month in 2011. Examples of diagnoses in each category were presented in a previous article (a sample list is included in Appendix C) (Kan et al., 2018).

Covariates

Age, sex, and race were self-reported. For some analyses, we also used ICD-9-CM codes to derive two comorbidity risk adjusters from claims data. The widely used Johns Hopkins ACG System (Version 11.0) was used to derive the number of major aggregated diagnostic groups (ADGs) and number of hospital-dominant morbidity types (The Johns Hopkins ACG® System). The ACG System categorizes all claims diagnosed conditions into 32 ADGs, where eight major ADGs are a subset of morbidity types that are associated with very high expected resource use. Hospital-dominant morbidity types represent diagnoses that, when present, are associated with higher probability of future hospitalizations (Johns Hopkins Bloomberg School of Public Health, 2014). We include a count of the number of these morbidity types in our analyses.

Outcomes

Six health utilization and functional outcome measures were used to assess concurrent and predictive validity. Three future outcomes were obtained from Medicare records: any occurrence of an inpatient hospitalization or a preventable hospitalization within 12 months of interview month and death within 12 months of interview month, where interview month spans from May to October. Three concurrent functional status measures were created from NHATS 2011 interview data, indicating the health status at the time of interview: receiving help with at least one household activity for health and functioning reasons, hereafter referred to as IADL (instrumental activity of daily living), receiving help with at least one mobility activity, and receiving help with at least one self-care activity.

Preventable hospitalizations were defined by identified conditions and associated Diagnostic Related Groups (DRGs) which can often be prevented or safely and effectively managed without hospitalization pertinent to a frail older population (Walsh et al., 2012). Functional status used measures previously validated in the NHATS data to measure disability and functioning (Freedman et al., 2011). IADL assistance was measured by report of help with laundry, shopping, meal preparation, banking and finances, or medications/injections in the last month for health or functioning reasons. Mobility help was measured by report of help with transferring (getting out of bed), getting around inside, or going outside in the last month. Self-care help was measured by help with toileting, eating, dressing, or bathing in the last month.

Statistical Analyses

Prevalence of claims-based geriatric risk factors was examined by frailty status. To determine frailty associations with health utilization and outcomes, logistic regressions were used. Individuals were categorized by standardized frailty status (robust, pre-frail, frail) and presence of claims-based geriatric risk factors (0 factors present, 1 factor present, ≥2 factors present) based on the distribution of the number of factors (Appendix A) (Kan et al., 2018).

First, base models were run with only demographic measures (age, race, and gender). We then ran models with and without the two disease-based morbidity measures derived from claims to evaluate the degree to which geriatric risk and frailty measures had independent effects in predicting outcomes. We also run models with both claims and survey components to play out the scenario with both types of information available (i.e., a health system) and the utility of such models. These models all controlled for age, race, and gender. C-statistics were used to compare model performance. Because of the nature of non-nested models and complex survey design, we also present Akaike information criterion (AIC) as another measure of model selection to compare the relative quality among models for parsimony (Yao, Li, & Graubard, 2015). We validated out-of-sample performance of our models by running leave-one-out cross-validation (LOOCV) and reported the error rates. We run LOOCV given the nature of complex survey design, and given bootstrapping methods rely on equal-probability sampling within stratum (Opsomer & Miller, 2005). Because there may be concerns of different look-back and follow-up periods in our sample, we conducted sensitivity analyses standardizing both time periods and find our results hold (Appendix D). We standardize the look-back period to the minimum look-back period present in the sample, which was 16 months (the longest look-back period was 21 months). We then separately standardize the follow-up period by excluding decedents and examine outcomes within the 12-month interview window. Analyses were done using SAS 9.3 and with NHATS weights and survey design variables (SAS Institute, 2014).

Results

Our sample included 4,457 individuals, who when weighted represent 20 million fee-for-service Medicare beneficiaries. Fifty-seven percent were female, 85% were White, and the cohort had a mean age of 76. Approximately 16% of the cohort met the frail criteria and 10% had ≥2 geriatric risk factors at time of interview in 2011 (Table 1). For our claims-based risk measures, nearly 15% of the cohort were identified as having difficulty walking, 8% had falls, 6% had weight loss, and 5% had dementia (Table 2).

Table 1.

Characteristics of Beneficiaries With Continuous Enrollment in Fee-for-Service Medicare Preceding NHATS Interview Month in 2011.

	Data source	Weighted % (n)	SE
Age, M (SD)	S	76.21	(0.16)
Sex	S
Female		56.6 (2587)	(0.90%)
Male		43.4 (1870)	(0.90%)
Race	S
White		85.3 (3282)	(0.86%)
Black		7.0 (872)	(0.43%)
Hispanic		4.7 (194)	(0.51%)
Other		2.9 (109)	(0.48%)
Frailty	S
Robust		38.6 (1513)	(0.95%)
Pre-frail		45.6 (2067)	(0.95%)
Frail		15.8 (877)	(0.71%)
Geriatric risk	C
0 factors		71.9 (3045)	(0.91%)
1 factor		18.6 (904)	(0.72%)
2+ factors		9.5 (508)	(0.51%)
Comorbidity measures, M (SD)	C
Hospital-dominant conditions count		0.39	(0.01)
Major ADG count		2.05	(0.04)
Outcomes
Inpatient hospitalization	C	21.0 (1046)	(0.71%)
Preventable hospitalization	C	4.5 (245)	(0.33%)
Mortality	C	4.2 (236)	(0.30%)
Any help with IADL	S	22.9 (1274)	(0.64%)
Any help with mobility	S	19.3 (1024)	(0.62%)
Any help with self-care	S	17.1 (901)	(0.62%)

Note. n = 4,457. ADGs and hospital-dominant conditions are morbidity clusters that are a component of the ACG case-mix system (Johns Hopkins Bloomberg School of Public Health, 2014). NHATS = National Health and Aging Trends Study; S = survey assessment; C = claims; ADG = aggregated diagnostic groups; IADL = instrumental activity of daily living. For age, hospital-dominant conditions count, and major ADGs, means are reported and SEs are in parentheses.

Table 2.

Prevalence of Geriatric Risk Factors (Claims-Based) and Frailty Status (Survey-Based) by Geriatric Risk Factor.

	Robust	Pre-frail	Frail	Total
Risk factors from claims	n = 1,513	n = 2,067	n = 877	n = 4,457
Fecal incontinence	a	a	a	0.32% (0.09%)
Presence of pressure ulcer/decubitus ulcer	a	a	8.49% (1.21%)	1.99% (0.27%)
Dementia	2.15% (0.36%)	4.49% (0.43%)	12.44% (1.27%)	4.84% (0.30%)
History of falls	4.49% (0.54%)	7.82% (0.57%)	19.63% (1.72%)	8.40% (0.49%)
Malnutrition	a	a	2.29% (0.57%)	0.70% (0.15%)
Social support challenges	a	a	a	a
Urinary incontinence	3.69% (0.60%)	3.76% (0.46%)	7.79% (1.05%)	4.40% (0.36%)
Vision impairment	1.03% (0.25%)	1.08% (0.23%)	2.00% (0.40%)	1.20% (0.15%)
Weight loss	2.29% (0.40%)	5.37% (0.54%)	14.39% (1.16%)	5.60% (0.35%)
Walking difficulty	7.59% (0.70%)	14.50% (1.04%)	32.97% (2.09%)	14.75% (0.78%)

Note. Frailty status is from 2011 NHATS. Percentages are weighted prevalence of any occurrence of geriatric risk factor in Part A or Part B claims from 2010 through 2011. Standard errors are reported in parentheses. Cannot report due to small sample size. NHATS = National Health and Aging Trends Study.

^aIndicates unweighted cell size <11 persons.

For each claims-based risk factor, the proportion identified with that risk increased by frailty status, with the lowest prevalence among those who are robust and highest prevalence among those who are frail (Table 2). For example, proportion of beneficiaries categorized as having fallen increased from 4% in the robust cohort, to 8% in the pre-frail cohort, and to nearly 20% in the frail cohort; proportion with walking difficulty increased from 8% in the robust cohort, to 15% in the pre-frail cohort, and to nearly 33% in the frail cohort.

Both claims-based and survey-based models show moderate discrimination in predicting the six outcomes we evaluated (Table 3). The c-statistic of survey-based models range from 0.67 (any hospitalization) to 0.84 (any IADL help); the claims-based models range from c-statistic 0.71 (any hospitalization) to 0.81 (any IADL help). The models applying risk information derived from claims data showed slightly better performance than the models using risk factors derived from the in-person survey assessments when predicting the three outcomes defined using Medicare data: any hospitalization (c-statistic 0.71 vs. 0.67 respectively), preventable hospitalization (c-statistic 0.76 vs. 0.72), and death (c-statistic 0.80 vs. 0.79). Claims-based models showed similar or decreased performance compared to survey-based models in predictions of survey-derived outcomes: help with IADL (c-statistic 0.81 vs. 0.84 respectively), mobility (c-statistic 0.79 vs. 0.82), and self-care (c-statistic 0.80 for both models). When information from both claims and survey assessments were combined, model performance increased over either of the other models alone, for all outcomes. Within each outcome, error rates are lower when using frailty or geriatric risk models compared to demographics alone (Table 3). The LOOCV measure helps assess model predictive performance, with lower error rates suggesting lower misclassification.

Table 3.

C-statistic for Risk Models Predicting Utilization and Functional Outcomes by Data Source.

	Data source
Outcome	Demographic only	Survey assessment onlya	Claims onlyb	Claims and survey assessmentc
Inpatient hospitalization	0.62	0.67	0.71	0.72
LOOCV error rate	17.41%	16.97%	16.22%	16.06%
Preventable hospitalization	0.65	0.72	0.76	0.78
LOOCV error rate	5.12%	5.04%	5.00%	4.95%
Any help with IADL	0.74	0.84	0.81	0.86
LOOCV error rate	17.18%	13.89%	15.03%	13.04%
Any help with mobility	0.69	0.82	0.79	0.85
LOOCV error rate	16.00%	12.96%	13.96%	12.18%
Any help with self-care	0.67	0.80	0.80	0.84
LOOCV error rate	14.98%	12.69%	12.68%	11.57%
Mortality	0.71	0.79	0.80	0.82
LOOCV error rate	4.84%	4.62%	4.76%	4.58%

Note. All figures reflect c-statistic of predictive model. All models adjust for age, race, and gender. Outcome of inpatient hospitalization, preventable hospitalization, and death indicates occurrence within 12 months of 2011 NHATS interview month; outcomes of any help with IADL, mobility, and self-care indicate help received as reported on the 2011 NHATS interview. LOOVV = leave-one-out cross-validation; IADL = instrumental activity of daily living; NHATS = National Health and Aging Trends Study.

^aSurvey assessment-only models utilize frailty status (robust, pre-frail, frail) as predictors.

^bClaims-only models utilize geriatric risk categories (0 factors met, 1 factor met, and 2+ factors met) and comorbidity measures as predictors.

^cClaims and survey assessment models utilize frailty status, geriatric risk categories, and comorbidities as predictors.

Table 4 presents odds ratios, C-statistics, and the AIC for models for two key outcomes: any hospitalization and any IADL help. The odds ratios for both the standardized frailty and the claims-based geriatric risk measures decreased when morbidity measures were added to the model but model performance increased. For example, in survey-based models, pre-frail status increased the odds of any hospitalization in the next 12 months by 2.04 times compared to robust status (p<.05). This odds ratio decreased to 1.69 times when morbidity measures are included in the model. The c-statistic for the survey-based model increased from 0.67 to 0.72 when morbidity measures are added. This pattern held for claims-based models and across models for any IADL help as the outcome. The AIC also suggest that models with morbidity measures performed better than models without and that claims-based models perform as well as survey-based models.

Table 4.

Comparison of Odds Ratios and Model Performance for Hospitalization and Help With IADLs by Data Source With and Without Comorbidity Measures.

	Survey assessment only		Claims only		Claims and survey assessment
	Unadjusteda	Adjusted	Unadjusted	Adjusteda	Adjusteda
Inpatient hospitalization (OR)
Frailty construct
Robust (ref)	—	—			—
Pre-frail	2.04 [1.72, 2.42]	1.69 [1.42, 2.02]			1.69 [1.41, 2.02]
Frail	4.07 [3.15, 5.26]	2.51 [1.92, 3.27]			2.41 [1.85, 3.13]
Geriatric risk factor
0 factors met (ref)			—	—	—
1 factor met			1.87 [1.57, 2.23]	1.28 [1.04, 1.57]	1.21 [0.98, 1.48]
2+ factors met			3.19 [2.60, 3.93]	1.47 [1.18, 1.84]	1.31 [1.05, 1.63]
Comorbidity measures
Count of major ADGs		1.31 [1.23, 1.40]		1.32 [1.23, 1.41]	1.28 [1.19, 1.37]
Count of serious hospital-dominant conditions		1.30 [1.17, 1.43]		1.33 [1.20, 1.47]	1.30 [1.17, 1.44]
C-statistic	0.67	0.72	0.66	0.71	0.72
AIC	19,659,940	18,584,672	19,853,862	18,817,432	18,559,056
Any help with IADL (OR)
Frailty construct
Robust (ref)	—	—			—
Pre-frail	3.47 [2.67, 4.49]	3.01 [2.30, 3.92]			2.99 [2.29, 3.91]
Frail	21.37 [16.17, 28.23]	15.42 [11.61, 20.47]			14.00 [10.58, 18.53]
Geriatric risk factor
0 factors met (ref)			—	—	—
1 factor met			2.71 [2.27, 3.24]	2.00 [1.66, 2.42]	1.78 [1.45, 2.19]
2+ factors met			7.01 [5.73, 8.58]	3.76 [3.01, 4.70]	2.87 [2.20, 3.73]
Comorbidity measures
Count of major ADGs		1.35 [1.25, 1.46]		1.32 [1.22, 1.41]	1.23 [1.13, 1.34]
Count of serious hospital-dominant conditions		1.04 [0.91, 1.18]		1.13 [0.99, 1.29]	1.04 [0.91, 1.18]
C-statistic	0.84	0.85	0.79	0.81	0.86
AIC	17,912,682	17,303,555	16,248,986	15,618,486	15,325,730

Note. n = 4,457. Bold font indicates that the OR is significant at .05 level. 95% confidence intervals are in square brackets. All models also include age, race, and gender. Outcome of inpatient hospitalization indicates any inpatient hospitalization within 12 months of NHATS interview month; outcome of any help with IADL indicates help received as reported on the 2011 NHATS interview. See text for further description of variables, models, and data sources. IADL = instrumental activity of daily living; OR = odds ratio; ADG = aggregated diagnostic groups; AIC = Akaike information criterion; NHATS = National Health and Aging Trends Study.

^aIndicates the model included on Table 2.

Discussion

We have developed a geriatric risk measure using information found in health claims data that can assist in identifying individuals with characteristics associated with the frailty phenotype and poor outcomes. In a large nationally representative community-dwelling 65+ population, our claims-based risk models performed, in most cases, as well or better than the in-person frailty survey assessments for a range of important outcomes. This study contributes to the expanding literature for deriving risk measures from available claims data or other secondary sources like EHRs for the measurement of health status for aging populations. The approach reported here, even though it relied on ICD codes assigned by clinicians and reported on Medicare claims, moves beyond the traditional disease-specific comorbidity measures generally used as a high-risk finding tool within older populations.

Our findings showed that prevalence of claims-derived geriatric risk factors was substantially higher among Medicare enrollees assessed in-person as frail compared to those assessed as robust. Our claims-based models were better at predicting outcomes of any hospitalization, preventable hospitalization, and death, when compared to the survey assessment-based frailty models. Moreover, our claims-based geriatric risk factors were independently predictive of increased risk of health care utilization, such as preventable inpatient stays, and adverse outcomes, such as death and functional limitations, beyond traditional morbidity measures. Furthermore, when both survey- and claims-based metrics of frailty and geriatric risk were combined, model performance increased for all outcomes, suggesting they are tapping into dimensions that are, at least in part, unique.

Unlike much of the previous literature, our article attempts to provide a framework showing how claims-based risk factors could offer practical opportunities to identify patients for targeted care management services for a specific target population; for example, those enrolled in a health plan or cared for by an integrated delivery system. Also, we provide a type of agreement validation by comparing and contrasting how our measure is associated with widely accepted functional status measures and with reference standard frailty assessments.

Our model improves on existing work in claims-based indices in that it does not solely rely on data-driven clinical diagnosis domains (i.e., congestive heart failure, stroke, Charlson Comorbidity Index) and it excludes non-modifiable or prior use factors (e.g., race, gender, and inpatient admission in the past 6 months) that are present in similar proposed indices (Kim et al., 2017; Segal et al., 2017). While these non-modifiable factors may be strong predictors of higher utilization and poor outcomes, our model offers prognostic utility beyond these factors by including them separately and still demonstrating independent and strong association of our geriatric risk factors on utilization and outcomes. These two previous studies, utilizing the Fried phenotype and the Rockwood deficit accumulation concepts, found moderate discrimination (c-statistic ranged from 0.60 to 0.75) in predicting health utilization and functional outcomes, and our models find comparable and, at times, higher model performance. Although these previous studies attempted to approximate a validated frailty index, our study does not attempt to do so. We expand beyond frailty to attempt to capture other geriatric characteristics and anchor our predictions against a validated frailty index.

This study provides some support that existing administrative datasets can be used to identify risk factors that appear to be statistically and conceptually associated with more widely accepted constructs of frailty, functional status, and geriatric syndrome. In this manner, our geriatric risk indicator has the potential to provide capabilities that extend beyond the traditional morbidity-based risk adjusters and also provide tangible targets for care management strategies. For example, by concentrating on the most prevalent risk factors or individuals with overlapping risk factors, health plans or other entities could identify specific cohorts of individuals for whom targeted population health interventions might mitigate adverse outcomes, such as sub-groups of older populations who are high-cost and high-need and not easily identified using traditional comorbidity measures alone (Long et al., 2017).

Furthermore, our model could be used to help identify those at highest risk of adverse outcomes within large general populations of older adults who could be referred for further targeted in-person assessments. This could broaden the use of frailty screening to health care settings other than those that serve specialized populations (e.g., within PACE plans) where it is typically done (Carey et al., 2008; Chrischilles et al., 2014; Hirth et al., 2009; Kelaiditi et al., 2016; Pugh et al., 2014; Rosen et al., 2001; Rosenberg, 2012; Sirois et al., 2017).

Although it has a number of unique strengths, this study also has several limitations. When we considered our 10 risk factors, severity or seriousness of each factor could not be reliably assessed with existing Medicare claims data. While we do account for the presence of one or more risk factors, the factors themselves were not leveled in terms of severity. It is unclear if accounting for severity would result in better risk prediction. Second, the relationship that was found between geriatric risk and outcomes is an association and causal inferences cannot be made. In addition, comorbidity covariates such as major ADG count may have reflected some of the individual geriatric risk measures to a limited extent, but both measures continue to be strong predictors across our models. We also have different look-back and follow-up periods for individuals based on which month they were interviewed for the NHATS assessment. When we conduct sensitivity analyses standardizing the look-back period and excluding decedents, we find our results hold, suggesting the differences do not alter our findings significantly. Finally, we acknowledge that our concepts are dependent on documentation in claims, which may be subject to biases for reimbursement purposes. For example, previous studies have demonstrated the infrequent use of ICD-9 V codes to capture social determinants of health (Torres et al., 2017). There may be particular concerns, in addition to our social support codes, on underestimating the true incidence of conditions that are often not reported (i.e., falls) or apply only in particular settings (i.e., pressure ulcers). However, we have demonstrated previously that these claims model have performed as well as some more time-sensitive data sources, such as electronic health records (Kan et al., 2018).

At the present time, claims remain the source for developing risk measures like the geriatric risk scores used here. As EHRs mature, the clinical granularity of information within clinician’s free-text notes will increasingly be available to those wishing to undertake geriatric risk assessment among older populations using secondary data (Kan et al., 2018). Our findings suggest that whatever the level of availability of health care data now or in the future, geriatric risk measurement tools such as the one reported here can and should be used to identify high-need sub-groups among older populations. It is only through the use of routine population level predictive risk case identification that effective and efficient care can be delivered to the frail sub-group of older persons; a group that for the foreseeable future will continue to increase with the aging of the population.

Appendix A

Distribution of Geriatric Risk Prevalence.

	Frequency	%	SE (%)
Number of risk factors
0	3,045	72.1	(0.91)
1	904	18.7	(0.72)
2	337	6.3	(0.38)
3	118	2.0	(0.20)
4	41	0.9	(0.17)

Note. Percentages are weighted prevalence of any occurrence of geriatric risk factor in Medicare claims from 2010 through 2011.

Appendix B

Sample Selection and Characteristics of Excluded Sample.

	N	Weighted %
Age
65–69	766	32.9
70–74	720	24.0
75–79	669	17.4
80–84	642	12.7
85–89	434	8.2
90+	389	4.7
Sex
Male	1,415	42.2
Female	2,205	57.8
Race/ethnicity
White, non-Hispanic	2,269	75.7
Black, non-Hispanic	916	9.8
Other, non-Hispanic	121	4.2
Hispanic	277	9.2
Don’t know/refused	36	1.1

Note. n = 3,620. Demographics of excluded NHATs participants (as described above). Information taken from 2011 NHATs survey. NHATS = National Health and Aging Trends Study.

Appendix C

Sample of Geriatric Risk Concept Measures Look-Up Table: ICD-9-CM Codes. Absence of fecal control.

ICD code	ICD code description (billable codes only)	Frailty concept
78760	FULL INCONTINENCE-FECES (Begin 2010)	AFC
Decubitus ulcer.
ICD code	ICD code description (billable codes only)	Frailty concept
7070	DECUBITUS ULCER (End 2004)	DEC
70700	DECUBITUS ULCER SITE NOS (Begin 2004)	DEC
70701	DECUBITUS ULCER ELBOW (Begin 2004)	DEC
70702	DECUBITUS ULCER UP BACK (Begin 2004)	DEC
70703	DECUBITUS ULCER LOW BACK (Begin 2004)	DEC
70704	DECUBITUS ULCER HIP (Begin 2004)	DEC
70705	DECUBITUS ULCER BUTTOCK (Begin 2004)	DEC
70706	DECUBITUS ULCER ANKLE (Begin 2004)	DEC
70707	DECUBITUS ULCER HEEL (Begin 2004)	DEC
70709	DECUBITUS ULCER SITE NEC (Begin 2004)	DEC
Dementia.
ICD code	ICD code description (billable codes only)	Frailty concept
2901	PRESENILE DEMENTIA	DEM
29010	PRESENILE DEMENTIA	DEM
29011	PRESENILE DELIRIUM	DEM
29012	PRESENILE DELUSION	DEM
29013	PRESENILE DEPRESSION	DEM
2902	SENILE DEMENTIA WITH DELUSIONAL OR DEPRESSIVE FEATURES	DEM
29020	SENILE DEMENTIA WITH DELUSIONAL FEATURES	DEM
29021	SENILE DEMENTIA WITH DEPRESSIVE FEATURES	DEM
2903	SENILE DEMENTIA WITH DELIRIUM	DEM
2904	VASCULAR DEMENTIA	DEM
Falls.
ICD code	ICD code description (billable codes only)	Frailty concept
E8842	FALL FROM CHAIR	FAL
E8843	FALL FROM WHEELCHAIR (Begin 1995)	FAL
E8844	FALL FROM BED	FAL
E8845	FALL FROM FURNITURE NEC (Begin 1995)	FAL
E8846	FALL FROM COMMODE (Begin 1995)	FAL
V1588	PERSONAL HISTORY OF FALL (Begin 2005)	FAL
E884.9	OTHER FALL FROM ONE LEVEL TO ANOTHER	FAL
E880.0	FALL, ESCALATOR	FAL
E880.1	FALL ON OR FROM SIDEWALK CURB	FAL
E880.9	OTHER STAIRS OR STEPS	FAL
Malnutrition and/or catabolic illness.
ICD code	ICD code description (billable codes only)	Frailty concept
261	NUTRITIONAL MARASMUS	MAL
262	OTH SEVERE MALNUTRITION	MAL
263	MALNUTRITION OF MODERATE DEGREE	MAL
263.1	MALNUTRITION OF MILD DEGREE	MAL
263.2	ARRESTED DEVELOPMENT FOLLOWING PROTEIN-CALORIE MALNUTRITION	MAL
264.2	VITAMIN A DEFICIENCY WITH CORNEAL XEROSIS	MAL
264.3	VITAMIN A DEFICIENCY WITH CORNEAL ULCERATION AND XEROSIS	MAL
264.4	VITAMIN A DEFICIENCY WITH KERATOMALACIA	MAL
264.5	VITAMIN A DEFICIENCY WITH NIGHT BLINDNESS	MAL
264.6	VITAMIN A DEFICIENCY WITH XEROPHTHALMIC SCARS OF CORNEA	MAL
Social support.
ICD code	ICD code description (billable codes only)	Frailty concept
V600	LACK OF HOUSING	SSN
V601	INADEQUATE HOUSING	SSN
V602	ECONOMIC PROBLEM	SSN
V604	NO FAMILY ABLE TO CARE	SSN
V605	HOLIDAY RELIEF CARE	SSN
V609	HOUSING/ECONO CIRCUM NOS	SSN
Major problems of urine retention or control.
ICD code	ICD code description (billable codes only)	Frailty concept
5964	ATONY OF BLADDER	URC
59653	BLADDER PARALYSIS	URC
59655	DETRUSOR SPHINC DYSSYNERG	URC
78834	INCONTINENCE-W/O SENSE	URC
78837	CONTINUOUS LEAKAGE	URC
788.33	Mixed incontinence	URC
Severe vision impairment.
ICD code	ICD code description (billable codes only)	Frailty concept
3690	PROFOUND IMPAIRMENT, BOTH EYES	VIS
36900	BOTH EYES BLIND-WHO DEF	VIS
36901	TOT IMPAIRMENT-BOTH EYES	VIS
36902	ONE EYE-NEAR TOT/OTH-NOS	VIS
36903	ONE EYE-NEAR TOT/OTH-TOT	VIS
36904	NEAR-TOT IMPAIR-BOTH EYE	VIS
36905	ONE EYE-PROFOUND/OTH-NOS	VIS
36906	ONE EYE-PROFOUND/OTH-TOT	VIS
36907	ONE EYE-PRFND/OTH-NR TOT	VIS
36908	PROFOUND IMPAIR BOTH EYE	VIS

Note. Note this is a partial list, not for application. If further information is desired, please contact authors. ICD = International Classification of Diseases.

Appendix D

Comparison of Odds Ratios and Model Performance for Hospitalization by Data Source With and Without Comorbidity Measures Using Standardized Look-back Period and Excluding Decedents.

	Inpatient hospitalization (OR)
	Standardized look-back period
	Survey assessment only		Claims only		Claims and survey assessment
n = 4,457	Unadjusted	Adjusted	Unadjusted	Adjusted	Adjusted
Frailty construct
Robust (ref)	—	—			—
Pre-frail	2.04	1.69			1.68
Frail	4.07	2.51			2.40
Geriatric risk factor
0 factors met (ref)			—	—	—
1 factor met			1.90	1.29	1.22
2+ factors met			3.31	1.52	1.33
Comorbidity measures
Count of major ADGs		1.31		1.32	1.28
Count of serious hospital-dominant conditions		1.30		1.33	1.30
C-statistic	0.67	0.72	0.66	0.71	0.72
AIC	19,659,940	18,584,672	19,848,705	18,812,179	18,556,521
	Inpatient hospitalization (OR)
n = 4,221	Excluding decedents
Frailty construct
Robust (ref)	—	—			—
Pre-frail	2.05	1.71			1.70
Frail	3.63	2.35			2.24
Geriatric risk factor
0 factors met (ref)			—	—	—
1 factor met			1.86	1.30	1.23
2+ factors met			3.19	1.54	1.39
Comorbidity measures
Count of major ADGs		1.32		1.32	1.28
Count of serious hospital-dominant conditions		1.25		1.28	1.25
C-statistic	0.65	0.70	0.65	0.70	0.70
AIC	17,802,627	16,941,566	17,913,715	17,114,246	16,911,336

Note. Bold font indicates that the OR is significant at .05 level. All models also include age, race, and gender. Outcome of inpatient hospitalization indicates any inpatient hospitalization within 12 months of NHATS interview month; outcome of any help with IADL indicates help received as reported on the 2011 NHATS interview. Restricted look-back period to the minimum look-back period present in the sample (16 months). Excluded decedents in follow-up period (shortest follow-up period is 4 months in the decedents cohort). OR = odds ratio; ADG = aggregated diagnostic groups; AIC = Akaike information criterion; IADL = instrumental activity of daily living; NHATS = National Health and Aging Trends Study.