Abstract
Background
Medically complex, disabled adults have high 30-day readmission rates. However, physical functioning is not routinely included in risk-adjustment models. We examined the association between multimorbidity with readmissions and mortality using a physical functioning weighted International Classification of Diseases (ICD)-coded multimorbidity-weighted index (MWI-ICD) representing 84 conditions.
Methods
We included Medicare beneficiaries with ≥1 hospitalization 2000–2015 who participated in a Health and Retirement Study interview before admission. We computed MWI-ICD by summing physical functioning weighted conditions from Medicare claims. We examined 30-, 90-, and 365-day postdischarge mortality using multivariable logistic regression and length of stay through zero-inflated negative binomials. Models adjusted for age, sex, race/ethnicity, body mass index, smoking status, physical activity, education, net worth, and marital status/living arrangement.
Results
The final sample of 10 737 participants had mean ± standard deviation (SD) age 75.9 ± 8.7 years, MWI-ICD 14.9 ± 9.0, and 20% had a 30-day readmission. Adults in the highest versus lowest quartile MWI-ICD had 92% increased odds of 30-day readmission (odds ratio [OR] = 1.92, 95% confidence interval [CI]: 1.65–2.22). A 1-point increase in MWI-ICD was associated with 24% increased odds of 30-day readmission (OR = 1.24, 95% CI: 1.18–1.31). A 1-point increase in MWI-ICD was associated with 32% increased odds of death within 365-day postdischarge (OR = 1.32, 95% CI: 1.25–1.40). Readmitted participants with the highest versus lowest quartile MWI-ICD had 37% increased number of expected hospitalized days (incidence rate ratio = 1.37, 95% CI: 1.17–1.59).
Conclusion
Among Medicare beneficiaries, multimorbidity using MWI-ICD is associated with an increased risk of readmissions, mortality, and longer length of stay. MWI-ICD appears to be a valid measure of multimorbidity that embeds physical functioning and presents an opportunity to incorporate functional status into claims-based risk-adjustment models.
Keywords: Comorbidity, Hospitalization, Multiple chronic conditions
Older adults with multimorbidity (multiple coexisting chronic conditions), disability, and social complexity have among the highest 30-day readmission rates (1–3). Individual-level characteristics such as these vary across diverse hospital populations but are not routinely available in claims-based risk-adjustment models. As a result, safety-net hospitals that serve disproportionately more high-need, high-cost patients incur greater readmission penalties, independent of provider performance (3–5). On the other hand, it is conceivable that including social determinants of health in risk-adjustment models could result in accepting lower standards of care for hospitals serving more socially disadvantaged populations. It is thus critical for risk-adjustment models to account for patient-level characteristics and variation in hospital- and provider-level quality of care in evaluating differences in hospital performance.
Traditional Centers for Medicare and Medicaid Services (CMS) risk-adjustment models for readmission use administrative claims data with age, sex, and select comorbidities based on diagnostic codes. A known limitation of these models is the lack of key individual-level covariates. First, individual characteristics such as social determinants of health, lifestyle and behavioral habits, and disability are known risk factors for readmission (1–3) but are neither routinely collected nor readily available in administrative data. Second, clinical characteristics are based on a limited inventory of chronic conditions, most often from hospital-based comorbidity metrics such as the Charlson Comorbidity Index and Elixhauser Comorbidity Score (6–8). These measures weight conditions based on future mortality risk and health care cost and utilization but exclude several less prevalent conditions that may also increase readmission risk (9). Furthermore, traditional comorbidity measures tend to underestimate disease burden. For example, conditions assessed from a defined period of hospital encounters underestimate outpatient diagnoses that may be incompletely documented at hospital admission. Thus, only a fraction of multimorbidity is typically included in claims-based risk-adjustment models.
Multimorbidity is nearly ubiquitous among hospitalized older adults and thus important to comprehensively assess in readmission risk-adjustment models. Furthermore, hospitalized older adults with multimorbidity often have functional limitations. The association between multimorbidity and hospital readmission risk has not been studied using a person-centered measure that weights chronic conditions by their impact on physical functioning. A new multimorbidity-weighted index (MWI) weights chronic conditions by the Short Form-36 physical functioning scale (10) and thus captures both cumulative disease burden and physical functioning. The MWI was developed and validated in community-dwelling adults with reliable self-report, and subsequently mapped to and validated using International Classification of Diseases (ICD-9) and ICD-10 codes in nationally-representative U.S. adults and Medicare beneficiaries (11) for use in claims data. MWI has been validated and shown to predict several important health outcomes, including long-term physical (12), cognitive (13), and social (14) functioning, health-related quality of life (15), hospitalized days (16), office visit utilization (17), and all-cause and suicide mortality (12), thus providing multiple potential uses.
Thus, the purpose of this study was to examine the predictive ability of the ICD-coded MWI (MWI-ICD) on 30-day hospital readmission risk among Medicare beneficiaries. Secondary outcomes included hospital length of stay and short- and long-term postdischarge mortality. We also sought to compare MWI-ICD with the most frequently used comorbidity measures in risk-adjustment, including the Charlson index, Elixhauser score, and simple disease count (6–8).
Method
Sample
Our sample takes advantage of a unique data linkage between 2 complementary data sets. The Health and Retirement Study (HRS) is a nationally-representative longitudinal cohort of >38 000 U.S. adults aged ≥51 years followed since 1992. HRS participants are interviewed biennially, starting with a face-to-face interview at baseline and subsequent telephone or face-to-face interviews. Participants provide information on physician-diagnosed medical conditions, functional status, living situation, employment, income, and health behaviors (18). HRS data are linked at the individual level to Medicare claims from the Centers for Medicare and Medicaid Services (CMS) for consenting HRS participants.
We included Medicare beneficiaries with 100% Fee-for-Service and continuous enrollment in Part A 1-year prior to hospital admission. We used the Medicare Provider Analysis and Review (MedPAR) file to obtain admission and discharge dates. For inclusion, participants must have had at least 1 hospitalization between January 1, 2000 and September 30, 2015 (prior to the conversion from ICD-9 to ICD-10 diagnostic codes) and participated in at least 1 HRS interview before admission. Participants completed the HRS survey biennially and contributed multiple observations of individual-level characteristics. If participants had readmissions during the study period, we used covariates preceding their initial admission to adjust for individual-level covariates.
Our sample did not include participants who died during the initial hospitalization or who were admitted to a skilled nursing facility upon discharge. We excluded participants without an HRS interview preceding the admission date (N = 550) and those missing HRS covariates needed for adjustment (N = 464). This study was approved by the University of Michigan and UCLA Institutional Review Boards (HUM00128383 and IRB#20-002145).
Multimorbidity Measurement and Assessment
The exposure of interest is the ICD-9 Clinical Modification (ICD-9-CM) coded MWI. MWI is a person-centered measure of multimorbidity that includes 84 conditions weighted by their impact on the 10-item Short Form-36 physical functioning scale (10). Thus, MWI-ICD represents both chronic disease burden and physical functioning because chronic conditions are weighted by physical functioning (9,11).
We used the CMS Medicare Parts A and B (inpatient and outpatient) and carrier files (PB data files) between 1991 and 2015 to obtain ICD-9-CM diagnostic codes representing 84 chronic conditions used to calculate MWI-ICD. For positive disease case ascertainment, we used the CMS Chronic Conditions Warehouse method (19) of 1 inpatient or 2 outpatient ICD-codes within a 2-year period. MWI conditions were considered present from the date of diagnosis used in the positive case ascertainment and carried forward in subsequent years. For diagnosed conditions, we also performed a lookback between 1991 and 2015, and if conditions were diagnosed even earlier, we used the date of first diagnosis. Four conditions in MWI-ICD that could potentially be definitively treated (cataract, peptic ulcer, cystitis, and thyroid nodule) were considered present if diagnosed within a year prior to admission date and were not carried forward following the diagnosis date. MWI-ICD was evaluated continuously as a standardized variable and in quartiles.
We also assessed the Charlson Comorbidity Index (6), Elixhauser Comorbidity Score (8), and disease count for comparison with MWI-ICD. Disease count included the same summed conditions as the MWI-ICD but were unweighted. The Charlson index was the sum of up to 19 chronic conditions weighted 1, 2, 3, or 6 based on their associations with 1-year mortality risk in hospitalized patients (6). The Elixhauser score was the sum of up to 30 conditions associated with inpatient mortality, cost, and length of stay that were each weighted 1 point (8). To provide more direct comparisons between MWI-ICD and existing measures, we applied the same methods used in MWI-ICD, which includes a comprehensive lookback period and carry-forward of conditions from the first date of positive case ascertainment. Thus, multimorbidity values were higher in Charlson, Elixhauser, and disease count compared with the typical use of these measures that draw from a single encounter or time period for case ascertainment.
30-Day Readmission Assessment
The primary outcome, 30-day readmissions, was obtained from the CMS MedPAR file. Readmissions were a binary variable that indicated a readmission within 30 days of first admission. If any subsequent admission occurred after the first admission and within the 30-day time period from first admission date, the patient was considered to have a 30-day readmission.
Secondary Outcomes
We assessed 2 secondary outcomes, including hospital length of stay and mortality postdischarge.
Hospital length of stay, a measure of hospital efficiency, was computed by subtracting the discharge date from the admission date in days for each separate admission for participants with a readmission.
Mortality following hospital discharge from the initial admission was measured at 30-, 90-, and 365-day postdischarge to capture short, medium, and long-term mortality, respectively. Immediate or 30-day mortality is an important outcome but has been known to underestimate mortality following hospitalization compared with when longer follow-up (eg, 90 and 365 days) is captured (20–22).
Mortality data were obtained from the CMS Master Beneficiary Summary File variable “date of death” between January 1, 2000 and September 30, 2015. Nearly all (99%) dates of death were validated by CMS based on a compilation of data from Medicare claims from the Medicare Common Working File, the Social Security Administration, online date of death submitted by kin, and the Railroad Retirement Board (23). Participants with incomplete date of death information (eg, only month and year of death) due to heterogeneous data sources had the day of death set as the last day of the month.
Covariates
Covariate data were obtained from the most recent biennial HRS interview preceding the initial hospital admission date. We included potential confounders and additional predictors in the associations with readmission and mortality risk, such as body mass index and physical activity (proxy for physical functioning), which would more greatly attenuate MWI-ICD than other measures, to better demonstrate the independent prediction of MWI-ICD conditions.
Covariates adjusted for in the models included age (continuous), sex, race, and ethnicity (non-Hispanic White, non-Hispanic Black, Hispanic, and other race), education (<12, 12, 13–15, and ≥16 years), marital status/living arrangement (married and/or living with a domestic partner, unmarried and living with another individual, unmarried and living alone), household net worth (quartiles), body mass index (<18.5, 18.5–24.9, 25–29.9, and ≥30 kg/m2), vigorous physical activity (< 3 or ≥3 times/week), and smoking status (current, former, and never smoker).
Statistical Analysis
We examined demographic and clinical participant characteristics through descriptive statistics, including means, standard deviations, and frequencies. To assess the bivariate association between 30-day readmissions with demographic factors, we used the Student’s t test and Pearson’s Chi-squared test for continuous and categorical variables, respectively.
To assess the association between MWI-ICD and the primary outcome, 30-day readmission risk, we used multivariable logistic regression adjusted for potential confounders. We also examined mortality at 30-, 90-, and 365-day postdischarge from the initial hospital admission. For postdischarge mortality, we used multivariable logistic regression and Cox proportional hazards modeling fully adjusted for potential confounders. We presented Kaplan-Meier survival curves for 365-day survival postdischarge by MWI-ICD quartiles.
To examine MWI-ICD and length of stay, we used zero-inflated negative binomial models. We selected this model due to the strong right-skewed distribution of hospitalized days, which is even more extreme in the general community-dwelling population because most individuals are never hospitalized. To account for the fact that most individuals are never hospitalized, we assigned a length of stay of zero days to participants without a readmission (which more than sufficed to represent the general population with no hospital admissions), and we examined the number of hospitalized days during readmission among readmitted participants. We used bootstrapping with 500 resamplings with replacement to obtain confidence intervals for the mean length of stay for each quartile of MWI-ICD.
To assess model fit and discrimination, we used prevalence distributions, the Akaike information criterion (AIC) (24), and concordance (C)-statistics (25). All models were adjusted for age, sex, race, ethnicity, education, living arrangement/marital status, household net worth, body mass index, vigorous physical activity, and smoking status from the HRS interview. The effect of MWI-ICD on 30-day readmissions and postdischarge mortality was presented as odds ratios with 95% confidence intervals and 2-sided p values. All analyses were conducted using SAS version 9.4 (SAS Institute, Inc., Cary, NC) and Stata version 15 (StataCorp LLC, College Station, TX).
Results
Participant Characteristics
Between January 1, 2000 and September 30, 2015, there were 15 180 Medicare beneficiaries admitted to the hospital. After excluding participants without 100% fee-for-service and continuous Part A enrollment prior to admission, an HRS interview prior to admission, and a complete set of covariates, the final sample included 10 737 adults. Participants with a 30-day readmission comprised 19.8% (N = 2 129) of the final sample. Within 365-day postdischarge from the initial hospital admission, 1 545 (14.4%) participants died.
The mean (standard deviation [SD]) age was 75.9 (8.7) years, and mean MWI-ICD was 14.9 (9.0). Compared with participants without a readmission, participants who were readmitted were more likely to be older (76.5 vs 75.7 years), male (45.5% vs 42.7%), less physically active (76.6% vs 74.4% with vigorous physical activity < 3 times/week), current smokers (14.7% vs 12.8%), and have a high MWI-ICD (Table 1). There were no significant differences by race and ethnicity, education, and marital status/living arrangement.
Table 1.
Participant Characteristics by Readmission Status Among Medicare Beneficiaries in the Health and Retirement Study
| Covariate | Sample, N=10 737 | Readmission Status | p Value* | |
|---|---|---|---|---|
| Yes (N=2129) | No (N=8 608) | |||
| Age (years), mean (SD) | 75.9 (8.7) | 76.5 (8.7) | 75.7 (8.7) | .0004 |
| ICD-coded multimorbidity-weighted index (MWI-ICD) | <.0001 | |||
| Mean (SD) | 14.9 (9.0) | 16.5 (9.1) | 14.5 (8.9) | |
| Median (IQR) | 13.5 (8.1, 20.3) | 15.3 (9.4, 22.2) | 13.0 (7.7, 19.7) | |
| Sex, N (%) | .0223 | |||
| Male | 4 646 (43.3) | 968 (45.5) | 3 678 (42.7) | |
| Female | 6 091 (56.7) | 1 161 (54.5) | 4 930 (57.3) | |
| Race and ethnicity, N (%) | .3495 | |||
| Non-Hispanic White | 8 422 (78.4) | 1 663 (78.1) | 6 759 (78.5) | |
| Non-Hispanic Black | 1 466 (13.7) | 311 (14.6) | 1 155 (13.4) | |
| Hispanic | 652 (6.1) | 121 (5.7) | 531 (6.2) | |
| Other | 197 (1.8) | 34 (1.6) | 163 (1.9) | |
| Education (years), N (%) | .7583 | |||
| <12 | 3 516 (32.7) | 689 (32.4) | 2 827 (32.8) | |
| 12 | 3 689 (33.9) | 740 (34.8) | 2 899 (33.7) | |
| 13–15 | 1 830 (17.0) | 364 (17.1) | 1 466 (17.0) | |
| ≥16 | 1 752 (16.3) | 336 (15.8) | 1 416 (16.4) | |
| Household net worth (quartiles), N (%) | .0020 | |||
| Q1 (<$30 500) | 2 684 (25.0) | 531 (25.0) | 2 153 (25.0) | |
| Q2 ($30 501–121 400) | 2 683 (25.0) | 594 (28.0) | 2 089 (24.3) | |
| Q3 ($121 401–340 000) | 2 688(25.0) | 520 (24.4) | 2 168 (25.1) | |
| Q4 (>$340 000) | 2 682 (25.0) | 484 (22.7) | 2 198 (25.5) | |
| Body mass index (kg/m2),N (%) | .3167 | |||
| <18.5 | 297 (2.8) | 57 (2.7) | 240 (2.8) | |
| 18.5–24.9 | 3 758 (35.0) | 782 (36.7) | 2 976 (34.6) | |
| 25–29.9 | 4 040 (37.6) | 777 (36.5) | 3 263 (37.9) | |
| ≥30 | 2 642 (24.6) | 513 (24.1) | 2 129 (24.7) | |
| Marital status/living arrangement,N (%) | .1236 | |||
| Married and/or lives with partner | 5 922 (55.2) | 1 133 (53.2) | 4 789 (55.6) | |
| Unmarried, lives with another individual | 1 581 (14.7) | 322 (15.1) | 1 259 (14.6) | |
| Unmarried, lives alone | 3 234 (30.1) | 674 (31.7) | 2 560 (29.7) | |
| Vigorous physical activity (# times/week), N (%) | .0381 | |||
| <3 | 8 038 (74.9) | 1 631 (76.6) | 6 407 (74.4) | |
| ≥3 | 2 699 (25.1) | 498 (23.4) | 2 201 (25.6) | |
| Smoking status, N (%) | .0242 | |||
| Never | 4 448 (41.4) | 890 (41.8) | 3 558 (41.3) | |
| Former | 4 876 (45.4) | 925 (43.4) | 3 951 (45.9) | |
| Current | 1 413 (13.2) | 314 (14.7) | 1 099 (12.8) |
Notes: ICD = International Classification of Diseases; IQR = interquartile range; Q = quartile; SD = standard deviation.
*Student’s t test for continuous variables and Pearson’s Chi-squared test for categorical variables.
Multimorbidity Characteristics
Participants with a readmission had a significantly higher median MWI-ICD than participants without a readmission (15.3 vs 13.0, p < .0001), although both groups had high multimorbidity on average compared with the general Medicare population (11) and community-dwelling adults (12). Compared with prior measures of multimorbidity, MWI-ICD had the widest distribution at both the low and high ends of multimorbidity, ranging from 0 to 66.1 (Figure 1). In contrast, the ranges for simple disease count (0–31), the Charlson index (0–20), and the Elixhauser score (0–19) were less than half that of MWI-ICD.
Figure 1.
Distribution of multimorbidity metrics among Medicare beneficiaries in the Health and Retirement Study. MWI = multimorbidity-weighted index.
MWI-ICD also had the least left-censoring compared with other metrics. A strong right-skewed distribution, particularly at the very low end, was most pronounced in the Charlson index. With Charlson, 31% had a value of 0–1 (13.5% with value = 0, 17.5% with value = 1), and the interquartile range was limited to 0–3. With Elixhauser, the right-skewed distribution persisted but was less prominent than Charlson at the very low end (2.4% with score = 0, 6.7% with score = 1), and the interquartile range was 3–7. Even simple disease count, defined using the same conditions as MWI-ICD but unweighted, had a right-skew but had less left-censoring than Charlson and Elixhauser (0.7% with 0 conditions, 2.1% with 1 condition) and a broader interquartile range of 5–10 conditions. In contrast, with MWI-ICD, 0.7% had a value of 0, and the interquartile range spanned 8.1–20.3 (Figure 1).
MWI-ICD had the most unique values and was closest to a truly continuous measure. Due to left-censoring and a limited number of unique values available for integer-based metrics, including Charlson, Elixhauser, and disease count, several participants’ multimorbidity values had ties resulting in unbalanced quartiles. Due to differences in the distribution and strong skew, comparisons for even continuous measures standardized or unbalanced quartiles precluded our ability to directly compare measures meaningfully beyond examining distributions.
30-Day Readmissions and Length of Stay
With increasing MWI quartiles, the odds of readmission increased monotonically, even after adjustment for all covariates. Participants in the highest quartile MWI-ICD had a 92% higher odds of readmission compared with those in the first quartile (odds ratio [OR] = 1.92, 95% confidence interval [CI]: 1.65–2.22; Table 2). As a continuous measure, a 1-point increase in MWI-ICD was associated with a 24% statistically significant higher odds of readmission (OR = 1.24, 95% CI: 1.18–1.31). The C-statistic for 30-day readmission for MWI-ICD was 0.58, which was essentially identical to that of other measures (simple disease count 0.57, Charlson index 0.58, and Elixhauser score 0.58).
Table 2.
30-Day Readmission Risk by Continuous Standardized Multimorbidity Measures and MWI-ICD Quartiles Among Medicare Beneficiaries in the Health and Retirement Study, N = 10 737
| Model* | 30-Day Readmission OR† (95% CI) | p Value | p for Trend | AIC | C-Statistic |
|---|---|---|---|---|---|
| Charlson, continuous | 1.23 (1.17, 1.29) | <.0001 | 10604.95 | 0.579 | |
| Elixhauser, continuous | 1.27 (1.21, 1.33) | <.0001 | 10587.88 | 0.584 | |
| Disease count, continuous | 1.18 (1.12, 1.24) | <.0001 | 10637.29 | 0.569 | |
| MWI-ICD, continuous | 1.24 (1.18, 1.31) | <.0001 | 10606.70 | 0.580 | |
| MWI-ICD, quartiles | <.0001 | 10599.42 | 0.582 | ||
| Q4 | 1.92 (1.65, 2.22) | <.0001 | |||
| Q3 | 1.60 (1.38, 1.85) | <.0001 | |||
| Q2 | 1.29 (1.12, 1.50) | .0006 | |||
| Q1 (Reference) | 1.00 | Reference |
Notes: AIC = Akaike information criterion; C = concordance; CI = confidence interval; ICD = International Classification of Diseases; MWI = multimorbidity-weighted index; N/A = not applicable; OR = odds ratio; Q = quartile.
*The multimorbidity measures have different units so cannot be directly compared head-to-head.
†Adjusted for age, sex, race, ethnicity, education, household net worth, body mass index, smoking status, vigorous physical activity, and marital status/living arrangement.
Among participants with a readmission, MWI-ICD was associated with a monotonic increase in the mean length of stay. MWI-ICD was associated with decreased odds of zero readmission days (OR = 0.46, 95% CI: 0.37–0.54) for those in the fourth versus first quartile MWI-ICD. Among patients with readmissions, MWI-ICD was associated with a 37% (95% CI: 1.16–1.57) statistically significant higher mean number of hospitalized days for those in the fourth versus first quartile MWI-ICD. As MWI-ICD increased, the incidence rate of a longer mean length of stay increased.
As a continuous variable, a 1-point increase in MWI-ICD was associated with decreased odds of zero readmission days (OR = 0.98, 95% CI: 0.97–0.98). Among patients with readmissions, MWI-ICD was associated with a 1% (95% CI: 1.00–1.01) statistically significant higher mean number of hospitalized days (Supplementary Table 1), which persisted after full adjustment for covariates.
Mortality
Among the sample of participants with at least 1 hospitalization (N = 10 737), mortality postdischarge from the initial hospital admission was as follows: 414 (3.9%) died within 30 days, 780 (7.3%) participants died within 90 days, and 1 545 (14.4%) participants died within 365-day postdischarge.
Participants with the highest quartile MWI-ICD had an increasingly higher odds of postdischarge mortality within 30 days (OR = 1.68, 95% CI: 1.21–2.33), 90 days (OR = 1.86, 95% CI: 1.46– 2.37), and 365 days (OR = 2.16, 95% CI: 1.81–2.59) postdischarge from the initial admission compared with those in the first quartile (Table 3). The C-statistics for MWI-ICD as a continuous or categorical variable were all 0.72 for 30-, 90-, and 365-day postdischarge mortality (Table 3). The Kaplan-Meier survival curve by MWI-ICD quartiles is shown in Figure 2.
Table 3.
Mortality Risk at 30-, 90-, and 365-Day Postdischarge Following Initial Hospital Admission, by Continuous Standardized Multimorbidity Measures and MWI-ICD Quartiles, N = 10 737
| Model* | Mortality OR† (95% CI) | p Value | p for trend | AIC | C-Statistic |
|---|---|---|---|---|---|
| 30-Day postdischarge mortality, N = 414 | |||||
| Charlson, continuous, | 1.60 (1.48, 1.73) | <.0001 | 3194.22 | 0.766 | |
| Elixhauser, continuous | 1.43 (1.30, 1.57) | <.0001 | 3257.967 | 0.738 | |
| Disease count, continuous | 0.96 (0.87, 1.07) | .4476 | 3310.04 | 0.712 | |
| MWI-ICD, continuous | 1.15 (1.04, 1.27) | .0054 | 3303.07 | 0.717 | |
| MWI-ICD, quartiles | .0041 | 3304.03 | 0.718 | ||
| Q4 | 1.68 (1.21, 2.33) | .0020 | |||
| Q3 | 1.55 (1.11, 2.16) | .0100 | |||
| Q2 | 1.46 (1.04, 2.05) | .0280 | |||
| Q1 (Reference) | 1.00 | Reference | |||
| 90-Day postdischarge mortality, N = 780 | |||||
| Charlson, continuous | 1.70 (1.58, 1.81) | <.0001 | 4974.585 | 0.766 | |
| Elixhauser, continuous | 1.53 (1.42, 1.64) | <.0001 | 5098.892 | 0.739 | |
| Disease count, continuous | 1.03 (0.95, 1.11) | .5103 | 5226.692 | 0.708 | |
| MWI-ICD, continuous | 1.22 (1.13, 1.31) | <.0001 | 5200.41 | 0.715 | |
| MWI-ICD, quartiles | <.0001 | 5203.36 | 0.715 | ||
| Q4 | 1.86 (1.46, 2.37) | <.0001 | |||
| Q3 | 1.63 (1.28, 2.09) | <.0001 | |||
| Q2 | 1.42 (1.10, 1.83) | .0064 | |||
| Q1 (Reference) | 1.00 | Reference | |||
| 365-Day postdischarge mortality, N = 1 545 | |||||
| Charlson, continuous, standardized | 1.86 (1.77, 1.96) | <.0001 | 7633.413 | 0.771 | |
| Elixhauser, continuous | 1.63 (1.54, 1.72) | <.0001 | 7893.869 | 0.742 | |
| Disease count, continuous | 1.12 (1.06, 1.19) | <.0001 | 8167.367 | 0.705 | |
| MWI-ICD, continuous | 1.32 (1.25, 1.40) | <.0001 | 8089.14 | 0.716 | |
| MWI-ICD, quartiles | <.0001 | 8106.22 | 0.715 | ||
| Q4 | 2.16 (1.81, 2.59) | <.0001 | |||
| Q3 | 1.71 (1.42, 2.05) | <.0001 | |||
| Q2 | 1.37 (1.14, 1.65) | .0009 | |||
| Q1 (Reference) | 1.00 | Reference |
Notes: AIC = Akaike information criterion; C = concordance; CI = confidence interval; ICD = International Classification of Diseases; MWI = multimorbidity-weighted index; OR = odds ratio; Q = quartile.
*The multimorbidity measures have different units so cannot be directly compared head-to-head.
†Adjusted for age, sex, race, ethnicity, education, household net worth, body mass index, smoking status, vigorous physical activity, and marital status/living arrangement.
Figure 2.
Kaplan–Meier plot for 365-day survival postdischarge from the initial admission by ICD-coded multimorbidity-weighted index quartiles. ICD = International Classification of Diseases; MWI = multimorbidity-weighted index; Q = quartile.
Among participants with at least one 30-day readmission (N = 2 129), mortality was higher than those with an admission but not necessarily a readmission. There were 252 (11.8%) deaths at 30-day postdischarge, 371 deaths (17.4%) at 90-day postdischarge, and 572 deaths (26.9%) at 365-day postdischarge from the initial hospitalization among those who experienced a readmission.
Discussion
In this study of nationally-representative Medicare beneficiaries, a new ICD-coded MWI was associated with 30-day readmission risk, longer length of stay, and 30-, 90-, and 365-day postdischarge mortality. This study used unique data linkages between CMS Medicare hospital admissions, outpatient encounters, and mortality data and the nationally-representative HRS to adjust for granular individual-level covariates. After robust adjustment for sociodemographic and behavioral and lifestyle predictors, MWI-ICD was attenuated but remained a significant predictor of increased readmissions, length of stay, and mortality risk.
This study expands the use of MWI-ICD to predict health care utilization outcomes in older hospitalized adults, despite MWI-ICD not being specifically designed for these outcomes nor acutely ill populations. MWI-ICD predicted 30-day readmissions with the same discrimination as did prior measures (Charlson, Elixhauser, and disease count), which was poor for all measures and not unexpected given none of these measures was designed for readmission risk. Concordance statistics for mortality were similar for MWI-ICD, disease count, and modestly higher for claims-based Charlson and Elixhauser measures that were designed to predict mortality.
More importantly, one of the greatest strengths and advantages of MWI-ICD over prior measures is its broad distribution at both the low and high ends of multimorbidity, including the least left-censoring and 0 values, and minimal ties to more precisely quantify each person’s multimorbidity. Even in this sample of older hospitalized Medicare beneficiaries with some of the highest multimorbidity in the nation, Charlson and Elixhauser were highly right-skewed. Their skewed distributions persisted despite applying the same methods used in MWI-ICD, which includes a comprehensive lookback period for disease ascertainment and carrying forward chronic conditions, to mitigate the tendency of claims-based measures to underestimate multimorbidity. With Charlson, 31% had a value of 0–1 (13.5% 0, 17.5% 1, IQR 0–3, range 0–20), while with MWI-ICD, 0.7% had a value of 0 (IQR 8.1–20.3, range 0–66.1). This study demonstrates the underestimation of multimorbidity when using Charlson and Elixhauser, even among hospitalized adults for which these measures were developed. This underestimation is likely due to the limited inventory of conditions (19 in Charlson, 30 in Elixhauser), and right-skewed distributions, despite correcting for disease underestimation.
In prior studies, MWI-ICD improved the precision and provided a better model fit for predicting mortality risk and future physical functioning compared with Charlson, Elixhauser, and disease count in national community-dwelling adults and Medicare beneficiaries at large. In the first study using Medicare beneficiaries, MWI-ICD had the broadest distribution, most unique values, least left-censoring, and most precision and model fit to predict future physical functioning, and equivocal or higher C-statistics for mortality risk compared with Charlson, Elixhauser, and disease count despite not being weighted to mortality (11). In mutually adjusted models with MWI, disease count, and Charlson with 10-year mortality risk, MWI and simple disease count were more strongly associated with mortality than was the Charlson index (11). The association of MWI with mortality was the strongest of all measures and more than double that of simple disease count. In the second study, an expanded ICD-coded MWI (MICD) also had the broadest distribution, most unique values, least left-censoring, and provided the best model fit (lowest AIC, highest C-statistic or coefficient of determination) to predict 10-year morality and 8-year future physical functioning in HRS-Medicare (26). Finally, in 3 large samples of community-dwelling adults, MWI persisted as a significant independent predictor of 10-year mortality, even after removing Charlson conditions from MWI (12).
In contrast with prior studies, Charlson and Elixhauser provided modestly better discrimination for mortality risk than MWI-ICD in this study, which may be due to 3 key differences. First, the present sample included acutely ill older adults with at least 1 hospitalization, which more closely reflects the sample of adults from which Charlson was developed, while MWI was developed in community-dwelling adults as a more broadly applicable measure of multimorbidity. Second, to provide more direct comparisons between measures, we improved the typical performance of Charlson and Elixhauser by applying a comprehensive lookback period and carrying forward conditions. This helped to reduce the tendency of claims-based measures to underestimate diseases due to their typical reliance on ICD-codes from single encounters or time periods. Third, case ascertainment and thus multimorbidity values will vary based on the macro and ICD codes used. To compute Charlson (and Elixhauser), this study used the Quan (27) macro that includes many broadly grouped parent ICD codes (rather than individual ICD-codes), and thus more conditions may have been captured than in the original Charlson and Deyo adaptation to ICD-9. MWI-ICD includes specific ICD codes and removes lower hierarchy codes not specified as chronic. Nonetheless, concordance statistics were within the range of prior studies examining Charlson and mortality risk in samples of hospitalized adults, and closely matched those of patients admitted with specific conditions, including acute coronary syndrome and hip fracture (28,29).
Multimorbidity has increasingly emerged as an important construct to capture in aging and older adults. Multimorbidity accumulates over the life span. Thus, measures that span the spectrum of conditions present in younger, middle-aged, and older adults are needed. Claims-based measures such as Charlson and Elixhauser used data from hospitalized patients and are skewed toward older, acutely ill hospitalized patients, and disease inclusion is based on non-patient-centered outcomes (mortality, cost, and utilization). In contrast, MWI includes 84 conditions and is weighted to a universal outcome, physical functioning, relevant to community-dwelling individuals of all ages. Physical functioning is upstream to more irreversible outcomes such as disability, frailty, and ultimately mortality, and thus can be targeted earlier for intervention. In prior studies, MWI was also associated with outcomes relevant to or downstream to physical functioning, including long-term cognitive and social functioning, disability, poor health-related quality of life, and mortality. The generalization of these associations has been assessed using several cohorts of community-dwelling adults and Medicare beneficiaries (1–5,30). Despite the importance of physical functioning, it is neither routinely assessed nor available in claims data. To fill this gap, MWI was developed to incorporate physical functioning into multimorbidity measurement so it can be incorporated in large studies where in-person assessments of physical functioning are infeasible. To facilitate its use, MWI is readily and freely available for self-reported physician-diagnosed conditions and as ICD-9 and ICD-10 coded conditions (MWI-ICD), with statistical code published online (11).
This study has limitations. First, in this sample, the use of the Charlson index and Elixhauser score to measure multimorbidity resulted in a highly right-skewed distribution with multiple ties, limiting direct head-to-head comparisons for outcome prediction. Ties were problematic even in this large data set due to strong left-censoring, especially in integer measures (Charlson, Elixhauser, and disease count) versus evenly distributed quartiles in continuous MWI-ICD. This is an important consideration for other studies attempting to compare such measures, as quartiles are highly unbalanced, and even standardized continuous measures do not account for different scales in each measure so must be interpreted with caution. Second, in claims data, chronic conditions to compute multimorbidity come from encounter diagnostic codes subject to coding error and bias such as upcoding (31). However, they remain essential for health services research and include more chronic conditions compared with limited inventories of self-reported chronic conditions typically assessed in population-based surveys. In studies with limited time periods, such as readmissions studies, multimorbidity tends to be underestimated because data typically come from a single baseline or brief period preceding hospitalization. To mitigate this, this study applied an extensive lookback starting in 1991 and carried forward 80 ICD-coded MWI chronic conditions until the date of admission. This lookback and carryforward were applied to conditions in all multimorbidity measures to reduce disease underestimation and compare the measures more directly, and thus helped the performance of Charlson, Elixhauser, and disease count compared with their typical use.
Third, the HRS survey from which individual characteristics were obtained for covariate adjustment is asynchronous with the MWI-ICD exposure obtained from Medicare data. However, most claims studies are not linked to granular individual-level data. We used the most recent HRS survey preceding admission to robustly adjust for individual-level characteristics, including socioeconomic status. Finally, this sample is limited to Medicare beneficiaries with a hospitalization, so our findings on readmission risk and postdischarge mortality cannot be generalized to younger, healthier populations.
In conclusion, a new ICD-coded MWI is a valid measure of multimorbidity that embeds physical functioning in its measurement and predicts readmission, length of stay, and postdischarge mortality. Compared with other metrics, MWI-ICD has the broadest distribution to measure multimorbidity. MWI-ICD is weighted to physical functioning, a universal value for individuals of all ages. MWI-ICD is high performing for several individual and health system level outcomes in community-dwelling adults of all ages and older hospitalized adults. Thus, MWI-ICD presents an opportunity to incorporate functional data into risk-adjustment models without the need to measure physical functioning, thus providing an invaluable tool to predict hospitalization outcomes among aging and older adults exhibiting functional decline through their chronic conditions.
Supplementary Material
Supplementary data are available at The Journals of Gerontology, Series A: Biological Sciences and Medical Sciences online.
Acknowledgments
The author thanks Suja Kumar and Mohammed Kabeto for assistance with the analysis.
Funding
This work was supported by the National Institute on Aging at the National Institutes of Health (grant number K23AG056638 to M.Y.W.).
Conflict of Interest
None declared.
Prior Presentations
This work was presented at the Society of General Internal Medicine Annual Meeting, Distinguished Professor of Hospital Medicine session, April 22, 2021 virtual meeting, and published in abstract form: J Gen Intern Med. 2021;36(Suppl 1):94–95.
Author Contributions
Study concept and study design, data acquisition and analysis, interpretation of data, manuscript writing, and approval of the final version for publication (M.Y.W.).
References
- 1. Greysen SR, Stijacic Cenzer I, Auerbach AD, Covinsky KE. Functional impairment and hospital readmission in Medicare seniors. JAMA Intern Med. 2015;175(4):559–565. doi: 10.1001/jamainternmed.2014.7756 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2. Kansagara D, Englander H, Salanitro A, et al. Risk prediction models for hospital readmission: a systematic review. JAMA. 2011;306(15):1688–1698. doi: 10.1001/jama.2011.1515 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3. Murray F, Allen M, Clark CM, Daly CJ, Jacobs DM. Socio-demographic and -economic factors associated with 30-day readmission for conditions targeted by the hospital readmissions reduction program: a population-based study. BMC Public Health. 2021;21(1):1922. doi: 10.1186/s12889-021-11987-z [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4. Gu Q, Koenig L, Faerberg J, Steinberg CR, Vaz C, Wheatley MP. The Medicare Hospital Readmissions Reduction Program: potential unintended consequences for hospitals serving vulnerable populations. Health Serv Res. 2014;49(3):818–837. doi: 10.1111/1475-6773.12150 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5. Singh S, Lin YL, Kuo YF, Nattinger AB, Goodwin JS. Variation in the risk of readmission among hospitals: the relative contribution of patient, hospital and inpatient provider characteristics. J Gen Intern Med. 2014;29(4):572–578. doi: 10.1007/s11606-013-2723-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6. Charlson ME, Pompei P, Ales KL, MacKenzie CR. A new method of classifying prognostic comorbidity in longitudinal studies: development and validation. J Chronic Dis. 1987;40(5):373–383. doi: 10.1016/0021-9681(87)90171-8 [DOI] [PubMed] [Google Scholar]
- 7. Deyo RA, Cherkin DC, Ciol MA. Adapting a clinical comorbidity index for use with ICD-9-CM administrative databases. J Clin Epidemiol. 1992;45(6):613–619. doi: 10.1016/0895-4356(92)90133-8 [DOI] [PubMed] [Google Scholar]
- 8. Elixhauser A, Steiner C, Harris DR, Coffey RM. Comorbidity measures for use with administrative data. Med Care. 1998;36(1):8–27. doi: 10.1097/00005650-199801000-00004 [DOI] [PubMed] [Google Scholar]
- 9. Wei MY, Kawachi I, Okereke OI, Mukamal KJ. Diverse cumulative impact of chronic diseases on physical health-related quality of life: implications for a measure of multimorbidity. Am J Epidemiol. 2016;184(5):357–365. doi: 10.1093/aje/kwv456 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10. Ware JE, Snow KK, Kosinski M, Gandek B.. SF-36 Health Survey: Manual and Interpretation Guide. Health Institute, New England Medical Center; 1997. [Google Scholar]
- 11. Wei MY, Ratz D, Mukamal KJ. Multimorbidity in Medicare beneficiaries: performance of an ICD-coded multimorbidity-weighted index. J Am Geriatr Soc. 2020;68(5):999–1006. doi: 10.1111/jgs.16310 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12. Wei MY, Mukamal KJ. Multimorbidity, mortality, and long-term physical functioning in 3 prospective cohorts of community-dwelling adults. Am J Epidemiol. 2018;187(1):103–112. doi: 10.1093/aje/kwx198 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13. Wei MY, Levine DA, Zahodne LB, Kabeto MU, Langa KM. Multimorbidity and cognitive decline over 14 years in older Americans. J Gerontol A Biol Sci Med Sci. 2020;75(6):1206–1213. doi: 10.1093/gerona/glz147 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14. Luster JE, Ratz D, Wei MY. Multimorbidity and social participation is moderated by purpose in life and life satisfaction. J Appl Gerontol. 2022;41(2):560–570. doi: 10.1177/07334648211027691 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15. Wei MY, Mukamal KJ. Multimorbidity and mental health-related quality of life and risk of completed suicide. J Am Geriatr Soc. 2019;67(3):511–519. doi: 10.1111/jgs.15678 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16. Han BH, Tuazon E, Wei MY, Paone D. Multimorbidity and inpatient utilization among older adults with opioid use disorder in New York city. J Gen Intern Med. 2022;37(7):1634–1640. doi: 10.1007/s11606-021-07130-8 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17. Wei MY, Poe J. Multimorbidity and primary care office visit frequency. J Gen Intern Med. 2020;35(Suppl 1):S203–204. doi: 10.1007/s11606-020-05890-3 [DOI] [Google Scholar]
- 18. Documentation. Health and Retirement Study. 2022. https://hrs.isr.umich.edu/documentation [Google Scholar]
- 19. Chronic Conditions Data Warehouse. Centers for Medicare & Medicaid Services. Accessed October, 2022. https://www2.ccwdata.org/web/guest/condition-categories-chronic [Google Scholar]
- 20. Rigberg DA, McGory ML, Zingmond DS, et al. Thirty-day mortality statistics underestimate the risk of repair of thoracoabdominal aortic aneurysms: a statewide experience. J Vasc Surg. 2006;43(2):217–222; discussion 223. doi: 10.1016/j.jvs.2005.10.070 [DOI] [PubMed] [Google Scholar]
- 21. Hirji S, McGurk S, Kiehm S, et al. Utility of 90-day mortality vs 30-day mortality as a quality metric for transcatheter and surgical aortic valve replacement outcomes. JAMA Cardiol. 2020;5(2):156–165. doi: 10.1001/jamacardio.2019.4657 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 22. Talsma AK, Lingsma HF, Steyerberg EW, Wijnhoven BP, Van Lanschot JJ. The 30-day versus in-hospital and 90-day mortality after esophagectomy as indicators for quality of care. Ann Surg. 2014;260(2):267–273. doi: 10.1097/SLA.0000000000000482 [DOI] [PubMed] [Google Scholar]
- 23. Jarosek S. Death Information in the Research Identifiable Medicare Data.Research Data Assistance Center. Accessed February, 2022. https://resdac.org/articles/death-information-research-identifiable-medicare-data [Google Scholar]
- 24. Akaike H. Information theory as an extension of the maximum likelihood principle. presented at: Second International Symposium on Information Theory; 1973; Akademiai Kiado, Budapest. [Google Scholar]
- 25. Harrell FEH. Regression Modeling Strategies: With Applications to Linear Models, Logistic and Ordinal Regression, and Survival Analysis. 2nd ed. Springer; 2015. [Google Scholar]
- 26. Wei MY, Luster JE, Ratz D, Mukamal KJ, Langa KM. Development, validation, and performance of a new physical functioning-weighted multimorbidity index for use in administrative data. J Gen Intern Med. 2021;36(8):2427–2433. doi: 10.1007/s11606-020-06486-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 27. Quan H, Sundararajan V, Halfon P, et al. Coding algorithms for defining comorbidities in ICD-9-CM and ICD-10 administrative data. Med Care. 2005;43(11):1130–1139. doi: 10.1097/01.mlr.0000182534.19832.83 [DOI] [PubMed] [Google Scholar]
- 28. Hautamäki M, Lyytikäinen LP, Mahdiani S, et al. The association between Charlson comorbidity index and mortality in acute coronary syndrome—the MADDEC study. Scand Cardiovasc J. 2020;54(3):146–152. doi: 10.1080/14017431.2019.1693615 [DOI] [PubMed] [Google Scholar]
- 29. Varady NH, Gillinov SM, Yeung CM, Rudisill SS, Chen AF. The Charlson and Elixhauser scores outperform the American society of anesthesiologists score in assessing 1-year mortality risk after hip fracture surgery. Clin Orthop Relat Res. 2021;479(9):1970–1979. doi: 10.1097/CORR.0000000000001772 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 30. Joynt KE, Orav EJ, Jha AK. Thirty-day readmission rates for Medicare beneficiaries by race and site of care. JAMA. 2011;305(7):675–681. doi: 10.1001/jama.2011.123 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 31. Schneeweiss S, Avorn J. A review of uses of health care utilization databases for epidemiologic research on therapeutics. J Clin Epidemiol. 2005;58(4):323–337. doi: 10.1016/j.jclinepi.2004.10.012 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.