Optum Integrated Claims Electronic Health Records (EHRs) from 1 January 2011 to 30 June 2017 was used for this study. This database included information on > 10 million individuals insured by UnitedHealth Group with their adjudicated claims linked to Humedica’s EHR. Data were de-identified and compliant with the requirements outlined in the Health Insurance Portability and Accountability Act.
This retrospective matched-cohort study comprised three mutually exclusive TRD, non-TRD MDD, and non-MDD cohorts. The first antidepressant pharmacy claim (on or after 1 July 2011 for TRD and non-TRD MDD patients, and randomly imputed dates on or after 1 July 2011 for non-MDD patients) was defined as the index date. Patient characteristics were evaluated in the 6 months that preceded the index date (baseline period), and study outcomes (treatment patterns, HRU, and costs) were evaluated from the index date up to 2 years post-index date, the end of continuous enrollment in a health insurance plan, or end of data availability (30 June 2017), whichever occurred first. Since the average time to become TRD was reported to be 1.3 years , a 2-year period was used to allow MDD patients the opportunity to be classified as TRD.
Patients with one or more visits to an IDN were considered IDN patients. Once patients were seen in an IDN, they remain flagged as IDN patients even if they later received care in a non-IDN setting. The IDN indicator was not determined at the patient level and was therefore not time or visit specific. Although a consensus definition of an IDN is lacking , an IDN was defined as a coordinated hospital inpatient and outpatient system that offers healthcare services in a defined geographic area in the Optum database. Multiple coordinated hospitals or IDNs could be included. IDNs that were included are primarily hospital based in Optum, but they provide more comprehensive services than solely inpatient care. Moreover, Optum defined an IDN as a provider network-level variable; providers self-identified whether they were IDNs and were required to have inpatient data to support their status as IDNs. In this study, 80.1% of the MDD cohort were associated with an IDN, whereas 75.7% were associated in the non-MDD cohort (Tables S1–2).
Patients with TRD (TRD cohort) were compared with two control cohorts: patients with MDD who did not have TRD (non-TRD MDD cohort) and patients without MDD (non-MDD cohort). Patients in both control cohorts were propensity score (PS) matched 1:1 with patients in the TRD cohort.
To be included in both the TRD and the non-TRD MDD cohorts, patients were required to meet the following criteria: (1) one or more diagnoses for MDD (International Classification of Diseases, Ninth Edition, Clinical Modification [ICD-9-CM]: 296.2x [MDD – single episode], 296.3x [MDD – recurrent episode]; ICD, Tenth Edition, CM [ICD-10-CM]: F32.x [excluding F32.8], F33.x [excluding F33.8]); (2) one or more claims for an antidepressant between 1 July 2011 and 30 June 2017 (MDD diagnosis could occur before or after the first antidepressant pharmacy claim); (3) one or more diagnoses of depression (ICD-9-CM: 296.2x, 296.3x, 300.4x, 311.x, 309.0x, or 309.1x; ICD-10-CM: F32.x, F33.x, F34.1, or F43.21) during the baseline or observation period. Both the baseline and follow-up periods were used to identify patients with depression due to the episodic nature of the disease .
Treatment-Resistant Depression (TRD) Cohort
In the TRD cohort, a pharmacy claims-based algorithm was used to identify MDD patients who were likely to have TRD within 2 years after the index date (i.e., follow-up). MDD patients were considered likely to have TRD if their MDD did not respond to two antidepressant treatment regimens, including augmentation therapy with an anticonvulsant, anxiolytic, antipsychotic, lithium, psychostimulant, or thyroid hormone medication, with adequate dose and duration (> 6 weeks) as per the American Psychiatric Association guidelines . Failure of a treatment regimen was defined as a switch of antidepressant (i.e., < 180 days after the end of the previous treatment), the addition of an antidepressant, or the initiation of an augmentation therapy. The initiation of the third treatment regimen had to occur > 6 weeks after the start of the first antidepressant treatment.
Non-TRD Major Depressive Disorder (MDD) Cohort
Patients in the non-TRD MDD cohort included MDD patients who did not meet the criteria for TRD, as previously defined, within 2 years of the index date.
Patients in the non-MDD cohort consisted of a randomly selected sample of patients without MDD (ICD-9-CM: 296.2x, 296.3x; ICD-10-CM: F32.x, F33.x [excluding F33.8]) from 1 January 2011 to 30 June 2017. Since a retrospective matched-cohort study design was used, the random selection of 500,000 non-MDD patients was considered sufficient to provide a pool of controls eligible for matching.
All patients included in this study were required to meet the following conditions: (1) no diagnosis for specific psychiatric comorbidities (i.e., psychosis [ICD-9-CM: 298.xx; ICD-10-CM: F23.x, F25.x, F44.89], schizophrenia [ICD-9-CM: 295.xx; ICD-10-CM: F20.x, F25.x], bipolar disorder/manic depression [ICD-9-CM: 296.0x, 296.1x, 296.4x, 296.5x, 296.6x, 296.7x, 296.8x; ICD-10-CM: F30.x, F31.x], dementia [ICD-9-CM: 290.xx, 294.1x; ICD-10-CM: F01.x, F02.x, F03.x]) between 1 January 2011 and 30 June 2017; (2) ≥ 18 years at the index date; (3) ≥ 6 months of continuous enrollment in a health insurance plan before and after the index date; (4) no antidepressant claims during the baseline period.
Baseline characteristics included demographics, physical and mental comorbidities, medication use, HRU, and costs were used to assess the distribution of the TRD, non-TRD, and non-MDD cohorts.
For each cohort, the proportion of patients receiving antidepressant medications during the observation period was reported. Duration of antidepressant therapy was also presented and defined as the number of days with medication available between the first antidepressant claim (i.e., index date) and the last day of supply of antidepressant with no gaps > 14 days (gaps were excluded in duration of therapy).
HRU and costs were categorized as all-cause, behavioral health-related, depression-related, and suicide-related. For each category, HRU and costs were divided into inpatient visits, inpatient days, emergency department (ED) visits, outpatient visits, and other visits. Behavioral health-related medical HRU and costs (including depression-related but not suicide-related costs) were identified using primary or secondary ICD-9-CM diagnostic codes 290.xx–319.xx and ICD-10-CM diagnostic codes F01.xxx–F99.xxx. Psychiatric pharmacy costs included the following classes of agents: anxiolytics, antidepressants, anticonvulsants/mood stabilizers, antipsychotics, and other mood stabilizers (e.g., lithium). Depression-related HRU and costs were identified using primary or secondary ICD-9-CM diagnostic codes 296.2x, 296.3x, 300.4x, 309.0x, 309.1x, 311.xx and ICD-10-CM diagnostic codes F32.x, F33.x, F34.1, or F43.21. Antidepressant pharmacy costs included the following classes of agents, selective serotonin reuptake inhibitors (SSRIs), norepinephrine-dopamine reuptake inhibitors (NDRIs), serotonin-norepinephrine reuptake inhibitors (SNRIs), serotonin modulators (i.e., nefazodone, trazodone, vilazodone, venlafaxine), tricyclics and tetracyclics, norepinephrine-serotonin modulators, monoamine oxidase inhibitors (MAOIs), and other (i.e., olanzapine-fluoxetine). Suicide-related HRU and costs were defined using primary or secondary ICD-9-CM diagnostic codes E95x or V62.84 and ICD-10-CM diagnostic codes T14.91x, X71.x-X83.x, T36.x- T65.x with a suffix of ‘2’ indicating ‘intentional self-harm’, T71.x with a suffix of ‘2’ indicating ‘intentional self-harm’ or R45.851.
Patients with TRD were matched 1:1 to non-TRD MDD and non-MDD patients using PS defined as the conditional probability of having TRD based on observable characteristics [26, 27]. The PS across patients were classified into strata, each containing an equal proportion of patients (5%). One TRD patient was matched to one non-TRD MDD and one non-MDD patient within the same stratum. The logistic regression model was used to estimate the PS, where having TRD (yes/no) was the binary dependent variable and patient characteristics were the predictors and included age, sex, race, year of the index date, geographical region, and type of healthcare plan. Baseline patient characteristics were compared using standardized differences; a covariate with a standardized difference < 10% was considered well-balanced . Rates of HRU were compared between matched cohorts using multivariable negative binomial regression (i.e., incidence rate ratios [IRRs]). Costs were expressed as per patient per year (PPPY) in $US, year 2017 values, using the Consumer Price Index for Medical care using means, standard deviations (SDs), and medians, and adjusted cost differences were calculated using multivariable ordinary least squares regression with 95% confidence intervals (CIs) and p-values obtained from nonparametric bootstraps with 499 replications. To account for censoring in the cost data, a phase-based approach has been recommended in studying cumulative costs . Specifically, mean costs per patient per month (PPPM) in the 6 months before and up to 2 years after the index date and stratified by 1-month periods were calculated as a sensitivity analysis. Since the matching was limited to demographic variables, baseline Quan-Charlson comorbidity index (Quan-CCI)  and all-cause healthcare costs were adjusted for in the multivariable models due to the remaining imbalances observed after matching. Given that TRD is complex, with multifactorial causes and effects with respect to other comorbidities, minimally matching on and controlling for potential confounders to avoid conditioning on mediators was deemed appropriate in this context. For the parameter estimates, standard errors, and Akaike information criterion from the regression models, see Tables S3–7.