Effect of frailty on resource use and cost for Medicare patients

We used 2013 data for Medicare patients with supplementary commercial health insurance from the MarketScan^® (Truven Analytics) database. All patients aged 65 and above with insurance coverage for at least the first 3 months of 2013 were extracted from the insurance eligibility data table. Inpatient, outpatient and pharmacy records for this cohort were also extracted. Records for services received through 31 March 2013 were separated and used for the construction of baseline measures. Records from 1st April and later were used for constructing the follow-up (FU) cost and event measures. All risk measures, resource use and cost variables were aggregated at the patient level. We excluded all patients under age 65 and patients with diagnoses at baseline of cancer, paraplegia or late effects of stroke since they would more than likely carry a different risk profile. All patient records for a 3-month baseline period were used to construct measures of frailty and morbidity to be used for the baseline description. All outpatient records were used to construct the Elixhauser comorbidity score [15]; all inpatient records were used to construct the Quan version of the Charlson score, which is calculated as the weighted sum of 12 chronic disease variable identified by examining all ICD-9 codes on patients hospital discharge records, with a possible range from 0 to 28 [1,8]. The Faurot frailty index [13] was used to identify frailty at baseline. The score uses 24 variables defined based on ICD-9 diagnosis codes and procedure, and durable medical equipment billing codes present in inpatient, outpatient and durable medical equipment billing files. The codes measure use of preventive services, which are indicators of robustness and are weighted with a minus weight because they were protective against frailty in the prediction model. Other codes capture symptoms, chronic disease diagnoses, functional limitations and use of medical equipment, which are all weighted positively as indicators of increasing levels of frailty. The scores are summed and used to classify patients at baseline into three groups described as frail (score >5), prefrail (score 1–4.9) and robust (score <1) [14]. Polypharmacy was defined by a record of five or more prescriptions filled for different medications (defined by generic name) for each month during the baseline period [16,17], and a record of one or more dispensed prescriptions that met the 2012 Beers Criteria list [18,19] of potentially inappropriate medications (PIMs) and classes to avoid in older adults, was used as a measure of PIM use. A binary indicator of any hospital admission during baseline, and patient data on age, sex, region and residence in a nonmetropolitan area (rural) were used to define their demographic characteristics. Patient race was not used because this variable is not available in the data source. Only patients with at least 90 days of insurance coverage were retained in the analytical dataset.

A total of 3,742,540 Medicare patients with supplementary private insurance coverage, for a mean of 352 days (standard deviation [SD] 44.9) in 2013, were included in the final dataset (Table 1). Their mean cost for the FU period was US$10,139 (SD 29,398), mean age was 74.1 years (SD 7.0) with 54.5% females. Overall, 14.4% of patients in the cohort had one or more emergency department (ED) visits, 11.4% had at least one hospital admission and 3.8% had an admission to a SNF during the FU period. Of the patients in the cohort, 208,528 (5.6%) met the criteria for being frail, 738,462 (19.7%) were prefrail and 2,795,550 (74.7%) were classified as robust based on baseline observations. There was a slightly higher representation of females in the frail and prefrail groups, and also higher mean age. Overall, 3.5% of patients had one or more hospital admissions during the baseline period, with large differences for frail patients (22.6%), compared with prefrail (7.5%) and robust patients (1.0%). Overall, 12.4% of patients had polypharmacy recorded during the baseline period, with greater rates observed for frail and prefrail patients, and 38.7% patients received one or more prescriptions for drugs on Beers list during the baseline period. The mean Charlson comorbidity score was 0.087 (SD 0.558) and ranged from 0 to15 at baseline, with higher values in frail and prefrail patients. The ten most common diagnoses measured by the Elixhauser index during the baseline period in order of prevalence include hypertension, diabetes, cardiac arrhythmia, diabetic complications, chronic obstructive pulmonary disease, senility, congestive heart failure, chronic renal failure, cardiac valve problems and asthma, with consistently higher rates observed by frailty status. The majority of patients resided in a metropolitan area (84.5%), with 24% from the north-east, 28% from a north-central state, 26% from the south, 21% from the west and 2% lacking state information (Table 1).

In unadjusted analyses the risk of hospital admissions, ED visits and admission to SNF during the 9-month FU period, as well as the amount of resources used (hospital days, ED visits and SNF days) was examined for the patient cohorts classified by baseline frailty. Robust patients had consistently lower risk of medical care use and cost, followed byprefrail patients, with frail patients having the highest use of medical care and highest cost (Table 2).

We estimated the marginal cost of being frail or prefrail compared with being robust using five multivariable models (Table 3), which used increasingly comprehensive measures to control for the effect of baseline variables on cost. In Model 1 we only controlled for the effects demographics (age and sex) and geographic characteristics (region and rural residence). In Model 2 we added any baseline hospital use (yes/no) and the Charlson comorbidity score, which are common control variables used in adjusted cost analyses. In Model 3 we added controls for baseline polypharmacy and use of drugs included on Beers list to the control variables included in Model 2. Model 4 contains the same variables as Model 2, except we added 26 individual binary indicators of the comorbid conditions in the Elixhauser index. Finally, in Model 5 we included all the baseline control variables. The model cost estimates for robust patients varied little as a result of adding baseline control variables. However, the marginal cost difference between robust, prefrail and frail patients moderated substantially as we added controls for baseline differences in comorbid conditions and medication patterns. The estimated cost difference between robust and frail patients decreased from US$23,117 in Model 1 to US$10,690 in Model 5. The cost difference between the estimates for pre-frail and frail was US$12,739 in Model 1 and US$5,171 in Model 5. The trends in cost estimates for the models and the unadjusted costs are displayed in Figure 1.

Similar findings were observed for other outcomes of interest. We examined the differences between robust, prefrail and frail patients on risk of hospital admissions, SNF admissions and ED visits, and found statistically significant and clinically meaningful differences between robust, pre-frail and frail patients (Table 4). The same pattern was observed for medical resource use for robust, prefrail and frail patients on mean use of hospital days, number of emergency department visits and days in an SNF for patients with use of these services (Table 4). We also observed increased risk of death during a hospital stay by frailty level (total death rates could not be measured because no information on out-of-hospital deaths is included in the dataset).

The estimates for the models that controlled for all available baseline variables (Model 5) demonstrated a consistently smaller difference between frail, prefrail and robust cohorts than those estimated by Model 1 where we only controlled for age, sex, region, rural location and study days. The estimated adjusted odds ratios of hospital admission and SNF admission had the largest differences between Model 1 and Model 5, as would be expected because Model 1 did not control for the presence of comorbid conditions. The differences between Model 1 and Model 5 were much smaller when resource use differences were estimated for only those patients who used the specific type of care.

We examined differences in acute medical events and cost of care for Medicare patients in 2013 using the Faurot [13] frailty index derived from data over a 3-month baseline period. We also used the baseline data to assign two other comorbidity measures: the Charlson score [22] and the Elizhouser index [15], which are commonly used to control for comorbid conditions in studies using archival billing data. We controlled for hospital use, polypharmacy and the use of any drugs on Beers list of PIMs [19]. The proxy FS classified patients at baseline into: robust, prefrail or frail groups. The mean cost of care for the 9-month FU period for Frail patients was US$25,320, compared with US$16,305 for the prefrail, and US$8,099 for robust patients. Our cost estimate for the frail group is slightly lower than the 2012 Medicare cost derived from the report of cost for frail patients provided by Figuerosa et al. [23] in their Appendix Table 3. These authors report that the US$3836 mean Medicare preventable spending for frail beneficiaries constitutes 9.3% of total cost. Thus, their frail cohort would have an estimated 9-months cost of US$30,935 in 2012 compared with our estimate of US$25,230 for 2013. However, Figuerosa et al. [23] classified frailty using a hierarchical ‘waterfall approach’, so their estimates include the marginal cost of comorbid conditions, and would therefore be higher than ours. Their frailty measure classified 8.6% of their cohort as frail. We observed a slightly lower rate of 5.56% in our cohort, which is similar to the 6.4% of patients with five to six difficulties in activities of daily living reported by Kautter [24] in the study of Medicare risk adjustment for the frail elderly. We observed a difference of US$17,220 in FU cost between frail and robust patients, and a difference of US$8205 between prefrail and robust individuals. This is a 213% higher FU cost for frail patients and 101% for prefrail patients compared with those who are robust. This difference between frail and robust patient in FU cost narrowed to US$10,690 (103%) in the fully controlled model (Model 5). A reduction of similar magnitude (US$5,519; 53%) was observed for the FU costs for the prefrail and robust comparison.

The importance of using a comprehensive set of baseline variables, including risk adjustors, constructed from the information imbedded in the International Classification of Disease (ICD), Current Procedural Terminology (CPT), Healthcare Common Procedure Coding System (HCPCS)  and the National Drug Code (NDC) codes in baseline data can be fully appreciated by reviewing the very large differences in the odds of events (admission, ED visit and SNF risk) that we observed for Model 1 (unadjusted) and Model 5 (fully adjusted, Table 4). Two-stage modeling approaches are often needed for comparing costs of important resources, such as days in hospital when many patients have no use of this type of resource (many 0 values). We observed a very large difference in the adjusted odds ratio estimates across all types of resource use for Models 1 and 5 for patients classified by frailty. This finding demonstrates the importance of the baseline control variables for the comparison of utilization of specific resources and their cost using two-part regression models.

We increasingly use archival billing data to evaluate the effects of intervention to improve outcomes and decrease costs for Medicare patients. These studies depend on careful matching of treatment and control populations at baseline to avoid selection bias. It is clear from the findings in this study that we need to include frailty, prefrailty and robust status, as well as several comprehensive measures of comorbidities and medication use in our baseline control measures because they noticeably have the ability to contribute independently to bias in our estimates if they are unbalanced in a population. The use of archival data for assessment of quality improvement and policy evaluation holds great promise for reaching the goal of a ‘learning health system’. However, the sparse amount of clinical measures that are available in billing data require that we are diligent in using the information imbedded in the ICD, CPT, HCPCS and NDC codes in baseline data to construct and validate proxy measures for medical conditions that may inject selection bias in our estimates. We found that patient frailty at baseline is an important predictor of medical resource use and cost for Medicare patients, independent of patient demographics, comorbidity burden and medication use patterns. Thus, frailty should be included as a baseline control and/or matching variable in comparative effectiveness and program evaluation studies that compare resource use and cost for Medicare beneficiaries.

The study has several limitations. All studies that use archival billing data to compare cost and outcomes for groups of patients are at risk of potential misclassification bias due to errors in ICD-9 diagnosis codes and CPT/HCPCS billing codes. This is especially the case with a score, such as frailty, which depends on the identification of billing codes for durable medical equipment during a short baseline period, because some frail patients may have received the equipment before the baseline period and may still be using it. This issue will tend to underidentify frailty and bias costs toward the null for frail patients. The fact that we limited the baseline period to 3 months could induce ascertainment bias for both frailty and comorbidity measurements. When designing archival studies, trade-offs must be made between using the available data for the baseline period, which may limit the data available for the FU time period, and potentially make the baseline measures less relevant due to differential effects of aging across population strata. Further, billing data contain errors and may have entries that are adjustment records for disallowed payments. Most of these records are removed during the data cleaning and dataset construction. However, our study used >285 million records to construct the analytical dataset, thus it is likely that we missed some adjustment records. Our mean cost per patient in the study was US$10,139 over 9-months FU, or approximately US$13,519 when annualized. This amount includes 2013 payments by Medicare, plus payments by a commercial insurer who provided supplementary Medicare insurance. The Kaiser Family Foundation reports that mean Medicare expenditure per beneficiary in 2013 was US$10,827, or US$2,692 less that our estimate. This relatively small cost difference increases our confidence in these cost estimates.

Frailty, even when identified in an imperfect manner through the use of diagnosis and procedure codes, increases the estimated mean cost of medical care by as much as US$17,220 over a 9-month time period. However, because frailty increases with age and often co-occurs with comorbid conditions, polypharmacy and use of PIMs medications, it is important to differentiate the effects of frailty (and prefrailty) on cost from effects imposed by comorbidities and medication burdens. We found that once we controlled for patient demographics, comorbidities and medication use patterns, the cost contribution of frailty remained, but was much smaller. Thus, it is important that researchers differentiate between effects of frailty, and effect of disease and medication burden in their studies. Each factor contributes to cost, but while often correlated, are not synonymous. Further, our findings highlight the importance of using a comprehensive approach to identify control variable for cost comparisons. The conceptual framework proposed by the National Academy of Medicine [3] provides a good start. It captures demographic, geographic, behavioral, physiologic and disease-related variables, which should all be considered for inclusion as control variables in cost comparisons. The change from ICD-9 to ICD-10 codes with the latter’s greater ability to capture social determinants of health, combined with the work to identify demographic factors [4] and develop a common taxonomy that segments individuals in a health system’s population based on the care they need as well as how often they need it [3] should be used to guide the selection of control variables used in future cost comparison studies.

The authors have no relevant affiliations or financial involvement with any organization or entity with a financial interest in or financial conflict with the subject matter or materials discussed in the manuscript. This includes employment, consultancies, honoraria, stock ownership or options, expert testimony, grants or patents received or pending, or royalties.

No writing assistance was utilized in the production of this manuscript.

The authors state that they have obtained appropriate institutional review board approval or have followed the principles outlined in the Declaration of Helsinki for all human or animal experimental investigations. In addition, for investigations involving human subjects, informed consent has been obtained from the participants involved.

This work is licensed under the Attribution-NonCommercial-NoDerivatives 4.0 Unported License. To view a copy of this license, visit http://creativecommons.org/licenses/by-nc-nd/4.0/