Factors associated with physicians’ prescriptions for rheumatoid arthritis drugs not filled by patients

This study was designed as a retrospective cohort study of RA patients identified from Optum’s de-identified Integrated Claims-Clinical dataset, which combines adjudicated administrative claims with Humedica’s EHRs. The longitudinal clinical repository from Humedica is derived from 50+ healthcare provider organizations in the USA that include over 600 hospitals and 6500 clinics and treat more than 63 million patients. The integrated dataset includes historical administrative claim data from pharmacy claims, physician claims, and facility claims, linked with EHRs including medications prescribed and administered, laboratory results, vital signs, body measurements, diagnoses, and procedures. Clinical information from EHRs was derived from both structured EHR data fields (e.g., biometric and clinical observations) and unstructured free text using natural language processing (NLP) (e.g., signs, diseases and symptoms). The data extracted from free text using NLP rather than free text itself were made available for this study. The integrated dataset was statistically de-identified under the Expert Determination method consistent with the Health Insurance Portability and Accountability Act.

The integrated dataset contains linked claims and EHRs for about 120,000 RA patients from 2007 to 2015, from which a MTX cohort and a B/T cohort were independently extracted. The MTX cohort consists of patients newly prescribed oral or injectable methotrexate. The B/T cohort consists of patients newly prescribed a biologic DMARD or tofacitinib. The biologics and tofacitinib in the B/T cohort were treated as a single group of drugs in this study to focus on whether a patient initiated treatment after receiving a biologic or tofacitinib prescription or infusion order.

For inclusion in this analysis, patients were first required to have ≥ 2 RA diagnoses (International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM): 714.0, 714.2, or 714.81) made by a physician (excluding laboratory, X-ray, and other provisional diagnoses) at least 7 days apart and within 12 months from inpatient and outpatient claims data. Subsequently, from EHRs, the first written prescription of MTX or the first prescription or infusion order for B/T after the first RA diagnosis was identified and defined as the index prescription. We conducted the search for the first RA diagnosis from claims and the first drug prescription from EHRs after diagnosis starting from 2007 onward in order to find patients at the earliest stage of disease possible allowed by the data. In addition, the following inclusion and exclusion criteria were applied. Patients must have been ≥ 18 years old at the time of the first RA diagnosis. Patients must have ≥ 12 months of continuous pharmacy and medical insurance coverage before and ≥ 2 (for MTX) or ≥ 3 (for B/T) months of continuous coverage after index prescription. Patients in the MTX cohort who ever had previous evidence of use of MTX, biologics, or tofacitinib in claims data using all available data before the index date were excluded. Patients in the B/T cohort who ever had evidence of use of biologics or tofacitinib in claims data using all available data before the index date were excluded. Also excluded were patients with diagnoses of other autoimmune diseases in claims data before the index, including psoriasis or psoriatic arthritis, inflammatory bowel disease, and ankylosing spondylitis or cancer (excluding nonmelanoma skin cancer). Patients with RA-related inpatient claims within 2 (for MTX) or 3 (for B/T) months after the index were also excluded as they may have received treatment in the hospital setting. This cohort definition is supported by previous research demonstrating that the positive predictive value of identifying RA patients exceeds 85% [14]. The 12-month period before the index prescription was defined as the baseline period during which potential predictors associated with primary nonadherence were identified from claims or EHRs.

The primary nonadherence concept was predicated on linking a new written prescription or infusion order in the EHR data to its filling/administration status in pharmacy claims or medical claims (for infused RA therapies). This study defined a new prescription as the very first prescription or infusion order of MTX or B/T after RA diagnosis. To find out how successfully these sample selection criteria identified new prescriptions, we calculated the percentage of patients with the first RA diagnosis in claims occurring at or after the first ever recorded EHR activity. This criterion guaranteed overlap of EHRs with claims data after RA diagnosis. Thus, the calculated percentage also indicates the percentage of index prescriptions being the first prescription ever recorded in the EHR data after RA diagnosis according to our definition of new prescriptions. This is a conservative measure (i.e., actual percentage of index prescriptions being the first prescription could be higher) as a patient could have been registered in an EHR system before the first recorded EHR activity occurred.

Primary nonadherence was defined as a new prescription of MTX or a new prescription or infusion order of B/T written by a physician as recorded in EHRs but not filled or administered within 2 months for MTX or 3 months for B/T based on claims. Since a drug filled/administered in claims could not be directly linked to a specific prescription in EHRs as administrative claims and EHR prescriptions/orders belong to two unlinked data systems, we searched for any MTX and any B/T filled/administered in claims after the index prescription and attributed the first such claim to the index prescription. Thus, we gave “credit” for any filled or administered B/T therapy, despite the fact that formulary restrictions might preclude use of certain targeted therapies and might require a different therapy to be used first. We compared the individual index B/T prescription with the first filled/administered B/T after the index, expecting that they would agree in most cases. More time was allowed for B/T initiation than for MTX as patients often need more time to clear clinical screening and insurance hurdles (e.g., screening for latent tuberculosis, prior authorization) and may face practical barriers such as scheduling and travel arrangement for infusions [15]. The time window of 2 and 3 months in the definition of primary nonadherence to MTX and B/T was also confirmed empirically by examining the distribution of time elapsed between a new prescription and the first filled or administered MTX and B/T. Infused and some self-injected biologics were identified in medical claims through the Healthcare Common Procedure Coding System, and all oral drugs and the majority of self-injected biologics were identified in pharmacy claims through National Drug Codes (NDCs).

The following baseline potential predictors for primary nonadherence were identified from claims or EHRs either at the index date or during 12 months before the index date. Using clinical judgment and subject matter expertise of the rheumatologist authors, 398 and 416 potential predictors at baseline were identified for the MTX and B/T cohorts, respectively. Baseline patient characteristics (30 variables for both the MTX and B/T samples) included age, sex, average household income (imputed values at the level of zip codes), percentage with college education (imputed values at the level of zip codes), geographic divisions, integrated delivery system (vs multispecialty practice), commercial insurance (vs Medicare), insurance products (e.g., health maintenance organization, point of service, preferred provider organization, exclusive provider organization), administrative service only, and a consumer-driven health plan which is typically associated with high out-of-pocket expenses [16]. Clinical and drug-related characteristics at baseline (17 variables for MTX and 19 variables for B/T) included the count of all active drug ingredients (not just RA related), count of chronic conditions, count of inpatient hospitalizations, count of emergency visits (all of the counts were derived from claims), number of days from RA diagnosis (derived from claims) to index prescription (derived from EHRs), number of RA diagnoses recorded in claims, calendar year of index prescription (from EHRs), prescription of both MTX and B/T within 2 months prior to index (from EHRs), and route of administration of index B/T prescriptions (namely, infusion, subcutaneous injection, or oral).

From structured EHR data fields, clinical observations (16 variables for both MTX and B/T) such as body mass index (BMI), smoking status (namely, never smoked, current smoker, not smoking now, previously smoking), pain score on a 0–10 scale, pulse, and respiration rate were included. When multiple clinical observations were available, the one closest to the index was chosen. A separate indicator for missing data was added for each clinical observation. From the NLP data extracted from free text in EHRs, we generated binary indicators (nine variables for both MTX and B/T) for the presence or absence of signs, diseases, and symptoms (pain, RA, swelling, tenderness, anxiety, depression, fatigue, weakness, and arthritis) as their mentions in free clinical text may serve as a dichotomous indicator for severity. No other measures of RA disease activity or severity were found from the current version of NLP data. Note that the NLP data extracted from free text were generated by Optum for general research purposes rather than specifically for this study or for RA.

In addition, the Johns Hopkins Adjusted Clinical Groups® (ACG®) risk adjustment/case mix system version 11.0 [17] was applied to claims data to generate general comorbidity indicators based on ICD-9-CM diagnosis codes from inpatient and outpatient claims and NDCs from pharmacy claims. These included diagnosis-based conditions called Expanded Diagnosis Clusters (EDCs; 253 variables for MTX and 234 for B/T), pharmacy-based conditions called Rx-defined Morbidity Groups (RxMGs; 59 variables for MTX and 58 for B/T), and Aggregated Diagnosis Groups (ADGs; 32 variables for both MTX and B/T). EDCs and RxMGs cover a large aggregate set of comorbidities including depression, anxiety, and sleep disorder. RxMGs represent treated conditions and do not completely overlap with EDCs which are based solely on diagnosis codes. ADGs are an even higher level aggregation of diagnosed conditions based on clinical criteria such as time limited or not, requiring primary or specialty care, or addressing physical health or psychosocial needs, as well as expected need for healthcare resources.

Patient characteristics at baseline were summarized using mean, median, standard deviation (SD), and percentage. A chi-square test was applied to test association between two categorical variables when appropriate.

Entering all of the 400 or so potential predictors selected based on clinical judgment into a predictive model for primary nonadherence may result in overfitting given the relatively small sample size as well as inclusion of potentially irrelevant variables. To address the high-dimensional input data, the potential predictors were tested for their importance in predicting primary nonadherence using L1-penalized LASSO (least absolute shrinkage and selection operator) logistic regression. LASSO logistic regression conducts variable selection and model estimation at the same time [18]. Specifically, each cohort was first randomly divided into training (75%) and test (25%) samples. In the training sample, a 10-fold crossvalidated misclassification error was evaluated across a wide range of the L1-penalty parameter to determine the largest such penalty that was within one standard error of the parameter that achieved the minimum error. The one standard error criterion tends to select a more parsimonious model without compromising much predictive power. This is especially important for this study as we aimed to develop an interpretable, transparent model as a simple tool for clinical application to identify patients at risk of primary nonadherence. To obtain unbiased estimates of the variables selected by the LASSO logistic regression, a regular multivariate logistic regression model was estimated using the entire training sample. Age, sex, and race were forced into the final regression models regardless of variable selection by LASSO regression.

To assess generalizability of model estimates, the logistic regression estimated with the training sample was applied to the test sample. The area under the curve (AUC) of the receiver operating characteristic and the Hosmer–Lemeshow goodness-of-fit test were estimated in the test sample. To visually show the calibration of the model in the test sample, mean observed probability was plotted against mean predicted probability in each of the 10 deciles of predicted probabilities. All programming was performed in R version 3.2.2 [19] with the two key packages dplyr version 0.4.3 [20] and glmnet version 2.0–5 [21].

Our definition of primary nonadherence was based on identifying filling/administration of MTX or B/T in claims data. Although the current data system does not allow direct linkage of a drug filled/administered to a specific prescription, attribution can be inferred based on the temporal relationship since, in most cases, the first B/T filled/administered was the same as the index B/T prescription. We speculate that the few discrepancies observed between the index B/T prescription and the first filled/administered drug, the majority of which involved etanercept and adalimumab, could be due to insurance and other reasons.

Although we aimed to find patients in the earliest stage of RA allowable by the data, it is possible that some of them may have had a RA diagnosis before the first recorded RA diagnosis found in claims due to left-censoring in incomplete claims data. This was suggested in the B/T cohort by the small number of patients prescribed non-anti-TNF biologics or tofacitinib as the seemingly first-line therapy. These drugs typically are not used as first-line therapy after MTX failure, suggesting left-censoring. However, it is noteworthy that the mean age at first RA diagnosis in our sample is not drastically different from that of incident cohorts from other claims-based studies [27]. Despite the fact that all of the RA patients may not have been incident cases, index prescriptions were “incident” in the sense of virtually all of them being new prescriptions after RA diagnosis and no filling or administration of corresponding drugs recorded in claims for at least 12 months prior to index prescription. This study focused on filling and administration of new prescriptions rather than new prescriptions among strictly defined incident RA cases.

Another limitatoin is that filling a drug does not necessarily mean taking the drug by the patient. Note that this, however, is not an issue for infused biologics since actual administration was recorded in medical claims. Nevertheless, it is interesting to study whether a prescription was even filled and picked up by the patient since this is a necessary action to take the medication. Since this study focused on patient adherence behavior, we tried to identify as many patient-reported variables as possible. However, formal instruments for patient-reported outcomes such as the Health Assessment Questionnaire and RAPID3 were not available for a large enough sample through the current NLP processed free text data. We believe that patient-reported factors including their beliefs and perceptions of treatments are important for understanding primary adherence behavior and worth more research in the future.

Finally, the automatic variable selection procedure implemented via LASSO logistic regression cannot necessarily recover the true set of underlying parameters for an outcome. We used clinical judgment in selecting the initial pool of potential predictors and included demographic variables such as age, sex, and race regardless of the LASSO regression results. Although the emphasis of our regression model is more on prediction than explanation [28], the final selected variables generally do seem to make clinical sense. Some of them may be proxies for other unmeasured factors or latent constructs (e.g., incontinence as a proxy for frailty). The final set of predictors was also validated in a separate independent test sample. The relationship found between the predictors and outcome can only be interpreted as associative rather than causal.