We used data from the Massachusetts General Hospital (MGH) COVID-19 Data Registry to identify patients aged 65 years and older who were hospitalized for PCR confirmed SARS-CoV-2 between March 11, 2020 – June 3, 2020 (12). The registry was developed using coded data extraction from the Mass General Brigham Enterprise Data Warehouse and manual chart review of electronic health records. Of the 1,391 patients in the registry, 549 were aged 65 and older. We then identified a subset of older adults who belonged to Mass General Brigham Medicare Accountable Care Organization (ACO) risk contract, which enabled us to use Medicare claims to characterize frailty, utilization, and expenditures. A total of 158 patients were enrolled in the ACO risk contract for at least one month in the year prior to their COVID admission. To ensure validity of frailty classification, we further excluded patients with < 9 months of continuous ACO enrollment in the year prior to hospitalization for a final study cohort of 136 patients (12-months n=122; 10-months n=3; 9-months n=11, Appendix Figure 1). Medicare claims were also used to examine healthcare utilization and costs following hospitalization; patients who died during hospitalization or had less than 1 month of ACO alignment were excluded from the follow-up utilization and cost analytic cohort (n=96). This study received exempt approval from our Institutional Review Board.

We used a CFI algorithm to compute frailty scores for each patient using administrative claims in the 12 months prior to COVID hospitalization (13). The CFI employs a cumulative deficit approach to estimating frailty status using ICD-10 diagnosis codes, current procedural terminology (CPT) codes, and the healthcare common procedure coding system (HCPCS) from Medicare administrative data files over a 12 month look back period (13, 14). Scores range from 0 to 1, with higher scores representing higher levels of frailty. Open-source code for calculating the CFI is available at https://dataverse.harvard.edu. Appendix Table 1 provides a listing of administrative code categories included in the CFI. We classified patients as robust/ pre-frail (CFI <0.25), mildly frail (CFI 0.25 to <0.35) and moderate to severely frail (CFI ≥0.35) (15).

Electronic health records identified date of death and date of last known follow-up in our health system in the cohort for survival analysis. The time frame for survival analysis started at hospital admission and continued for 6 months. Medicare claims characterized utilization per member per month (PMPM) and expenditures among survivors of the COVID-19 hospitalization who remained enrolled in the ACO during follow-up. Utilization characteristics include ED visits defined as treat-and-release (i.e., not resulting in an observation or inpatient stay) and observation visits that did not result in an inpatient admission. Inpatient admissions include all hospital stays (Medical, Surgical, and Psychiatric), excluding skilled nursing stays and inpatient rehabilitation. Total Medicare expenditures (TME) represent the sum of all medical costs (excluding prescription drug costs) billed to Medicare each month over the 6 months following discharge.

We describe the following demographic characteristics at hospital admission: age, sex, race/ethnicity, residence in a nursing home (NH) or assisted living facility (ALF), living alone, and being dual eligible for Medicare and Medicaid. Healthcare utilization and TME in the year prior to COVID hospitalization were characterized using Medicare claims. COVID registry data was used to identify clinical characteristics (e.g., chronic conditions, vital signs, use of supplemental oxygen, and symptoms at hospital admission), admission to the intensive care unit, and hospital length of stay.

We used descriptive statistics to characterize demographic and clinical characteristics for the full sample and by frailty status. Continuous data are presented using means/standard deviations and median/interquartile range. Categorical data are presented using proportions. We constructed a KaplanMeier curve to examine survival in the cohort by frailty status. We then ran Cox proportional-hazards models to evaluate the association between frailty and mortality when adjusted for factors independently associated with survival (residing in a NH/ALF and being dual eligible). To evaluate potential bias in survival analyses, we compared results to survival analyses in the subgroup of patients who had the full 12 months of ACO enrollment prior to admission (n=122).

Given the non-normal distribution of data, Kruskal-Wallis rank sum tests were used to evaluate utilization and a generalized linear model with gamma distribution and log-link function was used to evaluate differences in monthly Medicare expenditures. All analyses were performed using SAS version 9.4 (SAS Institute, Inc., Cary, NC).