Abstract
BACKGROUND: Patient effort to comply with complex medication instructions is known to be related to nonadherence and subsequent medical complications or health care costs. A widely used Medication Regimen Complexity Index (MRCI) has been used with electronic health records (EHRs) to identify patients who could benefit from pharmacist intervention. A similar claims-derived measure may be better suited for clinical decision support, since claims offer a more complete view of patient care and health utilization.
OBJECTIVE: To define and validate a novel insurance claims–based medication complexity score (MCS) patterned after the widely used MRCI, derived from EHRs.
METHODS: Insurance claims and EHR data were provided by HealthPartners (N = 54,988) (Bloomington, Minnesota) and The Johns Hopkins Health System (N = 28,589) (Baltimore, Maryland) for years 2013 and 2017, respectively. Yearly measures of medication complexity were developed for each patient and evaluated with one another using rank correlation within different clinical subgroupings. Indicators for the presence of individually complex prescriptions were also developed and assessed using exact agreement. Complexity measures were then correlated with select covariates to further validate the concordance between MCS and MRCI with respect to clinical metrics. These included demographic, comorbidity, and health care utilization markers. Prescribed medications in each system’s EHR were coded using the previously validated MRCI weighting rules. Insurance claims for retail pharmacy medications were coded using our novel MCS, which closely followed MRCI scoring rules.
RESULTS: EHR-based MRCI and claims-based MCS were significantly correlated with one another for most clinical subgroupings. Likewise, both measures were correlated with several covariates, including count of active medications and chronic conditions. The MCS was, in most cases, more associated with key health covariates than was MRCI, although both were consistently significant. We found that the highest correlation between MCS and MRCI is obtained with patients who have similar counts of pharmacy records between EHRs and claims (HealthPartners: P = 0.796; Johns Hopkins Health System: P = 0.779).
CONCLUSIONS: The findings suggest good correspondence between MCS and MRCI and that claims data represent a useful resource for assessing medication complexity. Claims data also have major practical advantages, such as interoperability across health care systems, although they lack the detailed clinical context of EHRs.
DISCLOSURES: The Johns Hopkins University holds the copyright to the Adjusted Clinical Groups (ACG) system and receives royalties from the global distribution of the ACG system. This revenue supports a portion of the authors’ salary. No additional or external funding supported this work.
The authors have no conflict of interest to disclose.
What is already known about this subject
Medication nonadherence and downstream health outcomes are known to be related to complicated prescribing orders, often termed medication complexity.
A commonly used measure of medication complexity, the Medication Regimen Complexity Index (MRCI), has been successfully adapted to electronic health records (EHRs) but not claims data.
Claims data have several key advantages over what is commonly available through the EHR, including these: (1) interoperability across multiple providers and health systems, (2) more proximal evidence of patient behavior through medication fills, and (3) more standardized data structures and coding that are transferrable across platforms.
What this study adds
A novel claims-derived medication complexity score (MCS) is described and compared with the established EHR-based MRCI to evaluate correspondence with one another and with selected comorbidity markers in 2 large health care systems.
The specifications of the MCS and its components are presented so others may apply this new tool. The introduction of a claims-based MCS is discussed in the context of application within health care delivery systems for the identification of patients at higher risk of poor health outcomes.
Clinicians and patients are often faced with challenges to optimally manage health conditions with pharmaceuticals. When assessing a patient’s drug regimen, “medication complexity” refers to the amount of effort needed by the patient or their caretaker to adhere to all current prescriptions. Prior research has shown that a highly complex regimen is often associated with poor adherence.1-5 This, in turn, is associated with medical complications and increased health care costs.6-11
A recent national survey estimates around 51% of American adults aged 40 years or older take at least 1 medication for a chronic condition and are “somewhat” to “largely” nonadherent by self-report.9 To try to limit how burdensome it is to achieve medication adherence, many organizations employ pharmacists or other qualified providers to periodically review a patient’s standing prescriptions.12 This process is required by Medicare Part-D Prescription Drug Plans as a medication therapy management (MTM) program.13 MTM review has been helpful in managing care for some patients with serious conditions and/or multimorbidity.13-14 The assessment process requires significant organizational resources, however, both in terms of identifying patients who would benefit most and the time commitment of the review.15-17 The application of available electronic information could greatly improve the efficiency and effectiveness of MTM, or comprehensive medication management (CMM), as part of clinical practice and improve quality of care.12,15,18
We have reported on the development of a medication complexity score (MCS) as a tool to assist clinicians with stratifying patients who would benefit most from medication review.15 The justification and basis for this measure came from a panel of 5 experts, each having board certification for pharmacotherapy and/or ambulatory care, who were asked to review a subset of patients from a large administrative claims record. Experts informed us of which types of features they considered important when evaluating the need for pharmacist intervention. The panel identified several features that reflect problems with medication use, including “complex medication instructions” and “low health literacy.” A corresponding MCS was thus developed, applied to the same dataset, and evaluated against other commonly used indices of risk in a retrospective 1-year cohort. The risk indices included the MTM criteria used by Centers for Medicare & Medicaid Services (CMS) and markers from the Johns Hopkins Adjusted Clinical Groups (ACG) System (version 12.0).13,19 The results suggested that having 1 or more complex medication, by some route, form, or frequency of use for the given prescription, corresponds to higher medical, pharmacy, and total costs when compared with either a broadly or narrowly defined set of MTM eligibility criteria.13 It also corresponded to a higher percentage of patients with 1 or more gaps in medication fills.15
Previous studies have developed measures of medication complexity based on both insurance claims and electronic health records (EHRs), but few of these tools have been deployed on an operational basis to inform clinical judgments and stratify subpopulations by need or risk.10,18,20 One exception is the widely applied Medication Regimen Complexity Index (MRCI). This measure, originally developed using paper medical records, has been adapted as an EHR-based composite measure reflecting the concept of medication complexity.18 The MRCI has been well validated and used with various subpopulations.4,21-22 To date, no systematic attempt to develop and assess a version of the MRCI using retail pharmacy insurance claims, rather than EHR data, has been reported.
EHRs and claims offer different perspectives on patient and provider interactions.23 EHRs provide valuable data for contextualizing patient care; however, EHR data from one provider often do not capture the full range of patient contacts. Alternatively, claims can represent a more complete view of utilization across an entire population.24-26 Thus, a claims-based medication complexity measure could be useful in identifying and addressing medication adherence challenges. Moreover, insurance claims indicate that the patient picked up a prescription (a fill), whereas an EHR pharmacy entry usually indicates only that a prescription was written (an order). Prior research in this area has led investigators to conclude that overlap between EHR and claims medication records is dependent on medication class and condition treated, with an estimated 20% of EHR-based medication orders not having a corresponding fill in the claims record within 12 months from the prescription date.27
The MCS is a novel claims-based measure patterned after the MRCI scoring algorithm.15 As part of our development process, this work sought to validate the MCS measure by comparing it to the established EHR-based MRCI across 2 large health care delivery systems. We also assessed continuous and binary versions of the MCS for the identification of patients who may benefit from pharmacist intervention. This study also compared the association between the MCS and MRCI with selected covariates within the study populations, including health care utilization markers. As a related objective, we investigated whether MCS and MRCI were more closely aligned among patients who had similar counts of fills and orders.
We use a retrospective cohort design to evaluate MCS and MRCI concordance in 2 major health care systems, both with EHR and claims data. These data were provided by HealthPartners (HP; Bloomington, Minnesota) for 2013 and The Johns Hopkins Health System (JHHS; Baltimore, Maryland) for 2017. In addition to being large integrated delivery systems with multiple providers and well-established EHRs, each has a large health insurance program from which we obtained insurance claims. This study focused on the subset of patients with continuous insurance coverage also receiving care from the delivery system. Most of the patients within the HP plan relied on HP providers for care delivery; a smaller subset of JHHS-insured enrollees relied on the JHHS providers. This study was approved by the Johns Hopkins School of Public Health Institutional Review Board.
Demographic, diagnostic/case-mix, and health care utilization metrics were calculated using the standardized output of the Johns Hopkins Adjusted Clinical Groups (ACG) System. The ACG System is a widely used and validated case-mix algorithm that assigns patient-level case-mix metrics based on the International Classification of Diseases, Tenth Revision, and other input data in claims or EHR data.19 All ACG metrics were calculated using the full claims records at each study site.
Several variables were selected as descriptive features and covariates. These included patient age; age group; sex; count of chronic conditions; count of concurrent-year active ingredients; condition markers for diabetes, hypertension, ischemia, and congestive heart failure; and annual utilization counts, including emergency department (ED) visits, inpatient hospitalizations, and readmissions within 30 days of discharge. These metrics are commonly used in risk segmentation literature and in assessing the validity of MRCI in the prediction of health outcomes.4,18,28,29-31
Data from each study site included beneficiary eligibility and medical and pharmacy files. The original data were normalized to a local data structure from their original raw format for quality assurance and inspected for completeness and accuracy. All programming was completed in R (version 3.6.0), using the “ggplot2” library for visualizations and base packages for statistical analysis.32
The original MRCI attaches medication-level weights for different complexity levels of route, form, frequency, and additional directions and then sums them together as a composite score.21 These scores are then aggregated across the patient’s active medications to reflect an overall regimen complexity. The MRCI has often been applied manually by a committee of pharmacists or other clinicians, although automated versions have been developed and applied to EHRs.18,21 Even in these automated versions, a significant amount of annotation and cleaning is required to attach component weights to a combination of coded features or free text for route, form, frequency, and additional directions. Once weights are assigned to each level of each feature, MRCI can be applied to the entire EHR without needing to manually review each patient.
Using the results of our previous study, we adapted and modified the subcomponents of the EHR-based MRCI to develop the claims-based MCS.15 Because much of the component information in MRCI can be identified in claims, it was possible to use administrative records of pharmacy reimbursement to calculate an analogous, although not identical, version of the MRCI, forming the basis of the MCS we evaluated in this study.
When using EHR as its source, component A of the MRCI encodes the complexity of the medication’s route and form. For example, an oral tablet is considered less complex than an oral film or subcutaneous injection, which then corresponds to a lower weight. Component B of the MRCI refers to prescribed medication frequency, with more frequent use generally being higher in assigned weight. Component C of the MRCI covers additional directions for use (eg, “with meals” and “scheduled”) and is more likely to be entered as secondary free text or as a part of the prescribed frequency (eg, “QD at bedtime”).
Components A through C of the MRCI were fully available in the EHR, but only component A was available within claims. Therefore, it was possible to calculate the standard MRCI using our EHR data, but because of the content of the insurance data, the claims-based MCS differed from the MRCI in several ways. First, both of our claims records made use of the US Food and Drug Administration’s National Drug Code (NDC). The NDC number system is specific to each individual type of medication, and therefore, each NDC has a known route and form. This meant it was possible to develop an analog for the MRCI’s component A weights tied to each NDC that could be used across different health systems. We also used the NDC-based approach to calculate component A of the MRCI for the HP EHR, but for JHHS, it had to be annotated. This required a pharmacist member of our study team to assign component A weights to unique combinations of medication route and form, which could then be appended to the JHHS EHR for MRCI calculation. NDC listings were provided by the Multum Lexicon Plus, from Cerner Multum, Inc.33 A sample reference of component A weights assigned by route and form is made available in the Supplementary Materials of this article (Supplementary Table 1, available in online article).
Second, claims data often lack information on use frequency. Thus, component B weights required modification. The medication quantity and days supply were instead used to approximate MRCI weights by first calculating the ratio of quantity to days’ supply and then classifying the ratio into categories corresponding to weights, like those outlined by the original authors of MRCI.21 One important nuance to this approach is that a ratio of quantity to days’ supply assumes a consistent interval in days per unit administered. For example, a ratio of 60:30 assumes 2 units per day are consumed (a half day per unit), although the prescribing frequency may reflect a different schedule altogether (eg, as needed, hourly). Less-than-daily administrations were given a weight of 2 by the original MRCI rubric and were similarly inferred by a quantity-to-supply ratio less than 1 for claims, then weighted accordingly. Acknowledging the possibility that a daily medication may slightly differ in count and days’ supply (eg, 30 tablets for a 28-day supply), component B weights for claims followed an interval scheme of rounding down to the nearest integer for weight. For example, a patient with 7 units to 14 days resembles a frequency of every other day and would have a component B weight of 2, since it is less than daily. A patient with 15 units to 14 days would most resemble a daily frequency and, rounding down, would be assigned a 1 for component B. A full account of component B mapping is provided in Supplementary Table 2.
Third, both HP and JHHS EHR contained component C information as part of the text for medication frequency. This was evaluated and annotated using the original MRCI’s rubric for component C weights. Because no frequency information was contained in claims for either HP or JHHS health care systems, it was not possible to approximate component C values for claims. Therefore, our claims-based MCS consisted of components A and B, whereas the EHR-based MRCI was scored using components A through C. Both MCS and MRCI were calculated as a sum of their constituent weights at the medication level and then summed for the full 1-year sample to obtain a single MCS and MRCI measure for each patient. An indicator for individual complex medications was also created at the medication level but aggregated as a binary value for patients having had any such medications prescribed within. These indicators signify that at least 1 medication was prescribed that had both component A and component B weights equal to or greater than 3.
To ensure as complete EHR and claims data as possible, we selected study patients who relied on the provider systems for outpatient care and were insured by each organization’s health plan for both medical and pharmacy services for the entire year. Patients without pharmacy records or outpatient services were excluded. Additionally, patients who were not found in both the claims and EHR of either program were excluded, as were individuals for whom a valid MCS or MRCI score could not be calculated because of data quality issues. Patients were not required to have matching pharmacy encounter records between claims and the EHR.
We identified 121,993 unique patients in the HP health care system data of 2013, of which 111,975 had outpatient pharmacy records (Figure 1). A total of 54,988 HP patients who had at least 1 valid pharmacy record in both the EHR and claims, who were fully enrolled for all of 2013, and who were aged less than 65 years were selected for inclusion. The JHHS study site’s health plan had 273,375 unique patients in 2017, of whom 28,589 were selected for inclusion using the same criteria. We noted the most common exclusion reason for JHHS was that the patient was not found in the EHR (222,355).
FIGURE 1 Cohort Selection Logic for Study Populations
Separate MCS-to-MRCI comparisons were made for HP and JHHS. Patient characteristics were summarized by clinical variables and compared across data sources. Continuous MCS and MRCI were compared using measures of central tendency (ie, mean and SD) and Spearman correlation coefficients. Corresponding P values were considered significant at a low α level of 0.001 because of the relatively large volume of patients in either health care system. Binary complexity was compared using percentage of exact agreement. Subsets of patients were formed for different age ranges, sex, counts of chronic conditions, and active ingredients to allow inspection of MCS-MRCI correlation strength for different subpopulations. Additional Spearman coefficients were produced for either MCS or MRCI and selected demographic, diagnostic, and utilization metrics.
A separate set of correlation statistics for continuous MCS and MRCI were also produced for groupings of low and high counts of medication records (ie, fills in claims or orders in EHRs) to inspect correspondence of MCS and MRCI for patients with similar or dissimilar counts of records between claims and the EHR. Fills reflect any recorded medication being dispensed per the claims record. EHR pharmacy records contain medication orders only. Neither claims nor EHR pharmacy records were adjusted by count of refills for this analysis. The number of fills or orders were split around the 50th percentile, such that a patient could be identified as having a low count of orders (ie, EHR records less than EHR median) and a high count of fills (ie, claims records more than claims median) or vice versa. Patients labeled low for EHR and low for claims or high for EHR and high for claims were regarded as having similar amounts of information reflected in both the EHR and claims. These patients were seen to both seldom interact with providers and have little medication use (low-low pair) or frequently interact with providers and have much medication use (high-high pair). Conversely, dissimilar information composed of high orders and low fills or low orders and high fills, and by extension, inform us of a discrepant or incomplete pattern of provider encounters and medication use.
The HP and JHHS study cohorts differed modestly by patient makeup (Table 1). Patients in the HP sample were somewhat older (40.0, SD = 17.5 vs 31.2, SD = 18.2), with the largest proportional difference being observed for the age group less than 20 years (16.7% vs 31.0%). Both samples were disproportionately more representative of female patients (60.4% vs 66.1%). Counts of chronic conditions, active ingredients, and disease indicators reveal similar case-mix between HP and JHHS, although the latter appears to represent a more chronically sick population by comparison. This is consistent with health care utilization metrics with HP patients having fewer average annual ED visits (0.17, SD = 0.5 vs 0.82, SD = 1.81), hospital admissions (0.08, SD = 0.4 vs 0.20, SD = 0.75) and readmissions to the hospital within 30 days (0.01, SD = 0.2 vs 0.03, SD = 0.41).
|
| Variable | HealthPartners | Johns Hopkins Health System |
|---|---|---|
| Age, years | 40 ± 17.5 | 31.2 ± 18.2 |
| <20 | 9,208 (16.7) | 8,866 (31.0) |
| 20-34 | 10,404 (18.9) | 7,362 (25.8) |
| 34-44 | 8,761 (15.9) | 4,292 (15.0) |
| 45-54 | 12,114 (22.0) | 4,162 (14.6) |
| 55-64 | 14,501 (26.4) | 3,907 (13.7) |
| Sex | ||
| Female | 33,229 (60.4) | 18,897 (66.1) |
| Male | 21,759 (39.6) | 9,692 (33.9) |
| Chronic conditions | 2.8 ± 2.8 | 3.2 ± 3.7 |
| Active ingredients | 6.9 ± 6.7 | 7.4 ± 8.1 |
| Diabetes | 4,723 (8.6) | 3,376 (11.8) |
| Hypertension | 14,407 (26.2) | 8,739 (30.6) |
| Ischemic heart disease | 1,278 (2.3) | 1,568 (5.5) |
| Congestive heart failure | 965 (1.8) | 584 (2.0) |
| ED visits | 0.17 ± 0.5 | 0.82 ± 1.81 |
| Inpatient hospital stays | 0.08 ± 0.4 | 0.20 ± 0.75 |
| 30-day readmissions | 0.01 ± 0.2 | 0.03 ± 0.41 |
Data are presented as mean ± SD or n (%). Study samples: N = 54,988 at HP and 28,589 at JHHS. Counts reflect the number of patients with a given demographic or diagnostic condition, whereas means and SDs are provided for age in years, counts of chronic conditions, active ingredients, ED visits, inpatient hospital stays, and readmissions to a hospital within 30 days of discharge.
ED = emergency department.
We observed claims-based MCS to be consistently 2 to 4 times greater than the corresponding EHR-based MRCI (Table 2). SDs, however, suggest a high degree of variability within each subgroup. Average MCS increased with successive levels of age (27.8-58.6 for HP; 37.7-102.9 for JHHS), count of chronic conditions (25.6-66.9; 34.7-95.1), and count of active ingredients (7.9-52.7; 13.1-75.8). Similarly, average MRCI increased with each of the same subgroups in most instances. Coefficients were obtained for the correlation between MCS and MRCI in the full sample for HP and JHHS (P = 0.556; 0.627). Exact agreement for the binary indicators were moderate to high. Spearman coefficients increased from the lowest to the highest levels of each subgrouping for the HP sample (age: P = 0.526-0.589; chronic conditions: P = 0.449-0.529; and active ingredients: P = 0.336-0.475) and JHHS sample (age: P = 0.609-0.642; chronic conditions: P = 0.544-0.561; and active ingredients: P = 0.412-0.557). This increase did not appear to be perfectly linear in all cases, however. Agreement for binary indicators of MCS and MRCI did not appear to consistently increase with groupings of age, chronic conditions, or active ingredients.
|
| Patient group | MRCI Mean (SD) | MCS Mean (SD) | Health Partners | Johns Hopkins Health System | ||||
|---|---|---|---|---|---|---|---|---|
| Spearman rhoa | % Exact | MRCI Mean (SD) | MCS Mean (SD) | Spearman rhoa | % Exact | |||
| Full sample | 11.8 (11.6) | 43.3 (52.5) | 0.556 | 79.0 | 24.3 (33.5) | 59.3 (73.3) | 0.627 | 76.5 |
| Age, y | ||||||||
| < 20 | 10.8 (12.2) | 27.8 (31.3) | 0.526 | 50.7 | 17.0 (19.4) | 37.7 (46.7) | 0.609 | 61.0 |
| 20-34 | 11.6 (12.9) | 32.7 (40.3) | 0.529 | 87.2 | 20.1 (26.6) | 45.6 (51.6) | 0.552 | 86.3 |
| 35-44 | 13.1 (15.9) | 39.2 (47.2) | 0.550 | 85.8 | 25.2 (31.9) | 60.8 (67.3) | 0.611 | 85.1 |
| 45-54 | 14.6 (16.4) | 48.6 (57.9) | 0.574 | 83.7 | 33.3 (44.9) | 87.4 (95.4) | 0.654 | 81.1 |
| 55-64 | 12.8 (14.5) | 58.6 (63.2) | 0.589 | 83.1 | 38.2 (48.1) | 102.9 (102.6) | 0.642 | 78.9 |
| Sex | ||||||||
| Female | 12.2 (14.1) | 45.7 (54.6) | 0.565 | 80.1 | 24.8 (34.5) | 61.8 (75.6) | 0.624 | 78.3 |
| Male | 9.3 (8.8) | 39.5 (48.7) | 0.545 | 77.4 | 23.4 (31.4) | 54.6 (68.5) | 0.635 | 73.0 |
| Count of chronic conditions treated during year | ||||||||
| 1 | 10.8 (10.5) | 25.6 (23.8) | 0.449 | 79.6 | 15.3 (15.0) | 34.7 (35.1) | 0.544 | 75.3 |
| 2 | 17.1 (18.0) | 33.5 (32.3) | 0.466 | 80.9 | 18.2 (18.3) | 43.3 (43.7) | 0.509 | 78.9 |
| 3+ | 4.9 (5.1) | 66.9 (67.2) | 0.529 | 79.5 | 37.9 (44.8) | 95.1 (93.2) | 0.561 | 76.8 |
| Count of active ingredients in filled prescriptions | ||||||||
| 1 | 6.3 (5.8) | 7.9 (6.7) | 0.336 | 80.4 | 8.5 (9.3) | 13.1 (14.8) | 0.412 | 71.9 |
| 2 | 14.6 (15.5) | 13.6 (10.5) | 0.226 | 80.7 | 11.5 (11.3) | 21.7 (19.6) | 0.445 | 72.7 |
| 3+ | 11.8 (11.6) | 52.7 (56.1) | 0.475 | 78.5 | 29.9 (37.4) | 75.8 (79.6) | 0.557 | 77.3 |
aIndicates test significance at P < 0.001.
EHR = electronic health record; MCS = medication complexity score (novel measure developed by this study, calculated from pharmacy claims data); MRCI = Medication Regimen Complexity Index (established measure calculated from electronic health records); y = years.
MCS was also significantly correlated with selected covariates for each data source. Coefficient strength varied across variables but was generally aligned with or exceeded the estimate for MRCI (Figure 2). Age was moderately correlated with MCS for both HP and JHHS (P = 0.245; 0.329). A positive correlation for sex (P = 0.078; 0.072) was indicative of higher MCS for the female sex. A moderate to strong relationship was obtained for count of chronic conditions (P = 0.557; 0.568) and active ingredients (P = 0.769; 0.747). Several comorbidities showed a smaller relationship with MCS (diabetes: P = 0.285; 0.320, hypertension: P = 0.327; 0.396, ischemia: P = 0.148; 0.230, and congestive heart failure: P = 0.132; 0.177). Correlations for health care utilization measures were consistently small across HP and JHHS (ED visits: P = 0.171; 0.209, inpatient stays: P = 0.150; 0.225, and 30-day readmission: P = 0.071; 0.143). Each Spearman coefficient observed was significant at the P < 0.001 level (Supplementary Table 3).
FIGURE 2 Patient-Level Correlation Between Continuous Measures Medication Complexity and Selected Covariates at 2 Study Sites
Spearman coefficients for the MCS-MRCI comparison were markedly higher when looking at just the subset of patients with similar counts of records between EHR and claims (Table 3). The correlation was lowest for patients who had comparatively few medication orders in the EHR but substantial fills in claims, as when a patient goes outside the health care system to receive health services (HP: P = 0.156; JHHS: P = 0.066). This subgroup made up 19.5% of the sample for HP and 11.3% for JHHS. Correlations were much higher among patients with similar counts of records between the EHR and claims (HP: P = 0.796; JHHS: P = 0.779), which also accounted for a much larger share of either dataset’s patient cohort (HP: 72.6%; JHHS: 79.9%). Interestingly, the percentage of exact agreement between binary MCS and MRCI was not consistently affected by subset of record counts; instead, the agreement fluctuated between 80% and 90% for HP and 75% and 80% for JHHS.
|
| Record count subset | HP | JHHS | ||||
|---|---|---|---|---|---|---|
| Spearman rhoa | % Exact | n (%) | Spearman rhoa | % Exact | n (%) | |
| Full sample | 0.572 | 86.2 | 45,780 (100.0) | 0.627 | 76.5 | 28,589 (100.0) |
| eHigh-cHigh | 0.500 | 79.3 | 13,866 (30.3) | 0.500 | 75.2 | 10,812 (37.8) |
| eLow-cLow | 0.284 | 90.7 | 19,351 (42.3) | 0.391 | 76.0 | 12,042 (42.1) |
| eHigh-cLow | 0.409 | 82.0 | 3,627 (7.9) | 0.209 | 77.7 | 2,507 (8.8) |
| eLow-cHigh | 0.156 | 88.8 | 8,936 (19.5) | 0.066 | 81.8 | 3,228 (11.3) |
| All similar | 0.796 | 85.9 | 33,217 (72.6) | 0.779 | 75.6 | 22,854 (79.9) |
| All dissimilar | −0.382 | 86.8 | 12,563 (27.4) | −0.396 | 80.0 | 5,735 (20.1) |
| Exclude eLow-cHigh | 0.764 | 85.6 | 36,844 (80.5) | 0.734 | 75.8 | 25,361 (88.7) |
Subsets of patients were identified as having high or low (partitioned at the 50th percentile) medication counts using either EHR (e) or claims (c) records. Similar and dissimilar subsets were defined as instances in which EHR and claims were both or differentially low and high, hence the subset convention (eg, eHigh-cHigh, eHigh-cLow). According to Hartman (1977) and Steimler (2004), values between 75% and 90% demonstrate an acceptable level of agreement and not a strong agreement.
aIndicates test significance at P < 0.001.
EHR = electronic health record; HP = HealthPartners; JHHS = Johns Hopkins Health System; MCS = medication complexity score (novel measure developed by this study, calculated from pharmacy claims data); MRCI = Medication Regimen Complexity Index (established measure calculated from electronic health records).
Medication nonadherence and downstream health outcomes are known to be related to complicated prescribing orders. The MRCI has been successfully adapted to the EHR but not claims data. We have developed and evaluated a claims-derived MCS to address challenges associated with accessing and analyzing EHR data in many instances, specifically, the lack of information from out-of-system providers and reliance on medication orders rather than fills. We then compared the EHR-based MRCI with MCS to evaluate concordance to one another and to selected comorbidity markers in 2 large health care systems.
Our findings document moderate to high intercorrelations between the novel claims-based MCS and the established EHR-based MRCI. This remained true across diverse clinical subgroupings for age and comorbidity using secondary electronic data from 2 separate health care systems. By partitioning the sample by similar and dissimilar medication record counts across the EHR and claims, we isolated cases in which the correlation between our metrics was poor. This tended to occur when a patient had an abundance of fills without corresponding orders. As expected, a better correspondence across these 2 measures were identified when a greater consistency was found between counts of medication orders and fills (eg, high-high and low-low).
Correlations between the MCS and key covariates also met or exceeded that of the MRCI, potentially because of the expanded information available in claims. Prior research has demonstrated significant, high mean MRCI scores for patients known to have specific chronic conditions, such as diabetes (23.0-29.4) and heart disease (21.0-26.7).22,29-31 Our findings reinforce the observation that medication complexity varies with the number of patient conditions and active ingredients. The coefficient observed for the MCS and active ingredients was particularly high, as it was with the MRCI in its introductory publication (P = 0.9).21 The original publication also noted moderate size correlations for age and sex (P = 0.34; 0.49) that failed to reach significance likely because of limited sample size. We obtained low to moderate correlations for MCS and MRCI with age and sex, and both reached significance. Correlations for MCS and utilization measures were relatively small but also like those obtained in MRCI research for prior year hospitalization (P = 0.109).29
The MCS fully relies on transparent medication-level coding (eg, NDC, quantity, days’ supply) and thus is more readily automated than the MRCI. The MCS might also improve decision support tools and allow pharmacists to rapidly identify and enroll patients in CMM programs by prioritizing patients who are at elevated risk of nonadherence, hospitalization, and ED visits.15 It remains necessary to test the additive predictive value of MCS over traditional indices of risk and comorbidity, however. Researchers of medication complexity and adherence issues can more readily use MCS to leverage the benefits of claims data: access to medication fills and more complete information across multiple providers or systems.
This study has several important limitations. First, information captured by the EHR and claims often do not overlap fully.26,34 EHRs are specific to a health care system and may not have medications across all providers. Conversely, patients do not always pay for medication through insurance (eg, over-the-counter medications). This limits the scope of medications in claims to a degree. Second, claims tend to lack a comparable clinical context to understanding course of care. Such granularity might also aid care decisions. However, claims arguably offer a more complete and valid record for patient adherence than does the EHR (ie, fills vs orders). Third, we did not evaluate medication complexity with respect to adherence as part of this validation because of the difficulty interpreting differential adherence for fills vs orders. Our panel of pharmacy experts had also identified medication cost as contributing to patient nonadherence; thus, modeling the risk of certain health outcomes, including adherence, was outside the scope of this work.
Finally, there was some threat of bias being introduced in the differential component weighting for MRCI and MCS, especially for component B weights. Our patient selection criteria were another potential source of bias. For example, patients with intermittent care or significantly greater comorbidity may have been lost by requiring 12 months of continual enrollment and an upper age limit of 64 years. This was a significant limitation, as we have reason to believe Medicare recipients and chronically underserved populations could benefit most from targeted CMM.35,36
The present work represents a comprehensive step in the validation of a claims-based medication complexity measure (ie, MCS) that is generally comparable to a previously validated EHR-based measure (ie, MRCI). Future efforts in expanding the development and application of the MCS and other similar metrics should focus on further understanding its measurement properties, practical advantages, and disadvantages relative to complexity indices derived from the EHR.
ACKNOWLEDGMENTS
This work was made possible by previous database management, support, and conceptualization from the team at the Center for Population Health IT. We also thank HealthPartners (Bloomington, MN) for sharing data and technical support.
| 1. | Bazargan M, Smith J, Yazdanshenas H, Movassaghi M, Martins D, Orum G. Non-adherence to medication regimens among older African-American adults. BMC Geriatrics. 2017;17(1):63. doi:10.1186/s12877-017-0558-5 Crossref, Medline, Google Scholar |
| 2. | Ayele AA, Tegegn HG, Ayele TA, Ayalew MD. Medication regimen complexity and its impact on medication adherence and glycemic control among patients with type 2 diabetes mellitus in an Ethiopian general hospital. BMJ Open Diab Res Care. 2019;7(1):e000685. doi:10.1136/bmjdrc-2019-000685 Crossref, Medline, Google Scholar |
| 3. | Stange D, Kriston L, von-Wolff A, Baehr M, Dartsch DC. Reducing cardiovascular medication complexity in a German university hospital: Effects of a structured pharmaceutical management intervention on adherence. J Manag Care Spec Pharm. 2013;19(5):396-407. doi:10.18553/jmcp.2013.19.5.396 Link, Google Scholar |
| 4. | Alves-Conceicão V, Rocha KSS, Silva FVN, Silva ROS, da Silva DT, de Lyra-Jr DP. Medication regimen complexity measured by MRCI: A systematic review to identify health outcomes. Ann Pharmacother. 2018; 52(11):1117-34. doi:10.1177/1060028018773691 Crossref, Medline, Google Scholar |
| 5. | Lee JS, Yang J, Stockl KM, Lew H, Solow BK. Evaluation of eligibility criteria used to identify patients for medication therapy management services: a retrospective cohort study in a Medicare Advantage Part D population. J Manag Care Spec Pharm. 2016;22(1):22-30. doi:10.18553/jmcp.2016.22.1.22 Link, Google Scholar |
| 6. | Pantuzza LL, Ceccato MDGB, Silveira MR, Junqueira LMR, Reis AMM. Association between medication regimen complexity and pharmacotherapy adherence: a systematic review. Eur J Clin Pharmacol. 2017;73(11):1475-89. doi:10.1007/s00228-017-2315-2 Crossref, Medline, Google Scholar |
| 7. | Chang H-Y, Richards TM, Shermock KM, et al. Evaluating the impact of prescription fill rates on risk stratification model performance. Med Care. 2017;55(12): 1052-60. doi:10.1097/MLR.000000 0000000825 Crossref, Medline, Google Scholar |
| 8. | Ma X, Jung C, Chang H-Y, Richards TM, Kharrazi H. Assessing the population-level correlation of medication regimen complexity and adherence indices using electronic health records and insurance claims. J Manag Care Spec Pharm. 2020;26(7):860-71. doi:10.18553/jmcp.2020.26.7.860 Link, Google Scholar |
| 9. | Langer Research Associations. Medication Adherence in America: a national report card. National Community Pharmacists Association. 2013. Accessed July 2, 2021. http://www.ncpa.co/adherence/AdherenceReportCard_Full.pdf Google Scholar |
| 10. | Iuga AO, McGuire MJ. Adherence and health care costs. Risk Manag Healthc Policy. 2014;7:35-44. doi:10.2147/RMHP.S19801 Medline, Google Scholar |
| 11. | Weir DL, Motulsky A, Abrahamowicz M, et al. Failure to follow medication changes made at hospital discharge is associated with adverse events in 30 days. Health Serv Res. 2020;55(4):512-23. doi:10.1111/1475-6773.13292 Crossref, Medline, Google Scholar |
| 12. | Sheehan OC, Kharrazi H, Carl KJ, et al. Helping Older Adults Improve Their Medication Experience (HOME) by addressing medication regimen complexity in home healthcare. Home Healthc Now. 2018;36(1):10-19. doi:10.1097/NHH.0000000000000632 Crossref, Medline, Google Scholar |
| 13. | Chavez-Valdez AL. CY 2020 medication therapy management program guidance and submission instructions. Centers for Medicare & Medicaid Services. Updated 2019. Accessed April 17, 2020. https://www.cms.gov/Medicare/Prescription-Drug-Coverage/PrescriptionDrugCovContra/Downloads/Memo-Contract-Year-2020-Medication-Therapy-Management-MTM-Program-Submission-v-041019-.pdf Google Scholar |
| 14. | Wittayanukorn S, Westrick SC, Hansen RA, et al. Evaluation of medication therapy management services for patients with cardiovascular disease in a self-insured employer health plan. J Manag Care Pharm. 2013;19(5):385-95. doi:10.18553/jmcp.2013.19.5.385 Link, Google Scholar |
| 15. | Bishop, MA, Chang, HY, Kitchen, CA, Weiner, J, Shermock, KM. Development of measurable criteria to identify and prioritize patients for inclusion in comprehensive medication management programs within primary care programs. J Manag Care Spec Pharm. 2021;27(8): 1009-18. doi:10.18553/jmcp.2021.27.8.1009 Link, Google Scholar |
| 16. | Dakwa DS, Marshall VD, Chaffee BW. The impact of drug order complexity on prospective medication order review and verification time. J Am Med Inform Assoc. 2020;27(2):284-93. doi:10.1093/jamia/ocz188 Crossref, Medline, Google Scholar |
| 17. | Zullig LL, Jazowski SA, Wang TY, et al. Novel application of approaches to predicting medication adherence using medical claims data. Health Serv Res. 2019;54(6):1255-62. doi:10.1111/1475-6773.13200 Crossref, Medline, Google Scholar |
| 18. | McDonald MV, Peng TR, Sridharan S, et al. Automating the medication regimen complexity index. J Am Med Inform Assoc. 2013;20(3):499-505. doi:10.1136/amiajnl-2012-001272 Crossref, Medline, Google Scholar |
| 19. | Health Services Research & Development Center at the Johns Hopkins University Bloomberg School of Public Health. The Johns Hopkins ACG System Version 11.0 Technical Reference Guide. The Johns Hopkins University Bloomberg School of Public Health; 2014. Google Scholar |
| 20. | Kharrazi H, Ma X, Chang H-Y, Richards TM, Jung C. Comparing the predictive effects of patient medication adherence indices in EHR and claims-based risk stratification models. Popul Health Manag. 2021;24(5):601-09. doi:10.1089/pop.2020.0306 Crossref, Medline, Google Scholar |
| 21. | George J, Phun Y-T, Bailey MJ, Kong DCM, Stewart K. Development and validation of the medication regimen complexity index. Ann Pharmacother. 2004;38(9):1369-76. doi:10.1345/aph.1D479 Crossref, Medline, Google Scholar |
| 22. | Schoonover H, Corbett CF, Weeks DL, Willson MN, Setter SM. Predicting potential postdischarge adverse drug events and 30-day unplanned hospital readmissions from medication regimen complexity. J Patient Saf. 2014;10(4):186-91. doi:10.1097/PTS.0000000000000067 Crossref, Medline, Google Scholar |
| 23. | Pandya C, Chang H-Y, Kharrazi H. Electronic health record-based risk stratification: a potential key ingredient to achieving value-based care. Popul Health Manag. 2021;24(6):654-56. doi:10.1089/pop.2021.0131 Crossref, Medline, Google Scholar |
| 24. | Kharrazi H, Wang C, Scharfstein D. Prospective EHR-based clinical trials: the challenge of missing data. J Gen Intern Med. 2014;29(7):976-78. doi:10.1007/s11606-014-2883-0 Crossref, Medline, Google Scholar |
| 25. | Wilson J, Bock A. The benefit of using both claims data and electronic medical record data in health care analysis. Optum, Inc. 2011. Accessed July 2, 2021. https://www.optum.com/content/dam/optum/resources/whitePapers/Benefits-of-using-both-claims-and-EMR-data-in-HC-analysis-WhitePaper-ACS.pdf Google Scholar |
| 26. | Kharrazi H, Chi W, Chang H-Y, et al. Comparing population-based risk-stratification model performance using demographic, diagnosis and medication data extracted from outpatient electronic health records versus administrative claims. Med Care. 2017;55(8):789-96. doi:10.1097/MLR.0000000000000754 Crossref, Medline, Google Scholar |
| 27. | Rowan CG, Flory J, Gerhard T, et al. Agreement and validity of electronic health record prescribing data relative to pharmacy claims data: a validation study from a US electronic health record database. Pharmacoepidemiol Drug Saf. 2017;26:963-72. doi:10.1002/pds.4234 Crossref, Medline, Google Scholar |
| 28. | Jeffery AD, Hewner S, Pruinelli L, et al. Risk prediction and segmentation models used in the United States for assessing risk in whole populations: a critical literature review with implications for nurses’ role in population health management. JAMIA Open. 2019;2(1):205-14. doi:10.1093/jamiaopen/ooy053 Crossref, Medline, Google Scholar |
| 29. | Brysch EG, Cauthon KAB, Kalich BA, Sarbacker GB. Medication Regimen Complexity Index in the elderly in an outpatient setting: a literature review. Consult Pharm. 2018;33(9):484-96. doi:10.4140/TCP.n.2018.484 Crossref, Medline, Google Scholar |
| 30. | Negewo NA, Gibson PG, Wark PA, Simpson JL, McDonald VM. Treatment burden, clinical outcomes, and comorbidities in COPD: an examination of the utility of medication regimen complexity index in COPD. Int J Chron Obstruct Pulmon Dis. 2017;12:2929-2942. doi:10.2147/COPD.S136256 Crossref, Medline, Google Scholar |
| 31. | Libby AM, Fish DN, Hosokawa PW, et al. Patient-level medication regimen complexity across populations with chronic disease. Clin Ther. 2013;35(4):385-98.e1. doi:10.1016/j.clinthera.2013.02.019 Crossref, Medline, Google Scholar |
| 32. | R. Version 3.6.0. R Foundation; 2017. Accessed July 2, 2021. https://www.R-project.org/ Google Scholar |
| 33. | Cerner Multum, Inc. Lexicon Plus. 2020. https://www.cerner.com/solutions/drug-database Google Scholar |
| 34. | Kharrazi H, Weiner JP. A practical comparison between the predictive power of population-based risk stratification models using data from electronic health records versus administrative claims: setting a baseline for future EHR-derived risk stratification models. Med Care. 2018;56(2):202-03. doi:10.1097/MLR.0000000000000849 Crossref, Medline, Google Scholar |
| 35. | Miller GE, Sarpong EM, Davidoff AJ, Yang EY, Brandt NJ, Fick DM. Determinants of potentially inappropriate medication use among community dwelling older adults. Health Serv Res. 2016;52(4):1534-49. doi:10.1111/1475-6773.12562 Crossref, Medline, Google Scholar |
| 36. | Durfey, SNM, Kind AJH, Buckingham WR, DuGoff EH, Trivedi AN. Neighborhood disadvantage and chronic disease management. Health Serv Res. 2019;54 suppl 1(suppl 1):206-16. doi:10.1111/1475-6773.13092 Crossref, Medline, Google Scholar |
| 37. | HealthPartners. Quick facts about HealthPartners. 2020. Accessed May 13, 2020. https://www.healthpartners.com/hp/about/quick-facts/index.html Google Scholar |
| 38. | Johns Hopkins HealthCare LLC. Johns Hopkins Medicine. 2020. Accessed July 2, 2021. https://www.hopkinsmedicine.org/johns_hopkins_healthcare/ Google Scholar |
| 39. | Community Physicians. Johns Hopkins Medicine. 2020. Accessed July 2, 2021. https://www.hopkinsmedicine.org/community_physicians/patient_information/index.html Google Scholar |