Data source and cohort selection
This study has been approved by the Institutional Review Board of Indiana University (IRB no. 2011632512). The Indiana University School of Medicine IRB approved a waiver of informed consent given the retrospective cohort design. All study activities and research methods were performed in accordance with the relevant guidelines and regulations. This study employed EHR data from two health systems in Indiana. The Regenstrief Institute Data Core has access to data from the state’s Health Information Exchange (which includes clinical institutions, labs, and some insurance providers) via the Indiana Network for Patient Care (INPC) database as well as to the EHR data warehouses for Indiana University Health (IUHealth), a statewide system with 18 hospitals as well as outpatient clinics, and Eskenazi Health, the county hospital in Marion County where Indianapolis, Indiana, is located. Available data comes from physician data entry during routine patient care. The data warehouses capture all structured data from patient encounters within these health systems, while the INPC receives specific pieces of clinical data but is not as comprehensive in terms of data elements. The Data Core serves as the honest data broker for access to these data sources for research re-use. Identifiers from the Indiana Biobank are matched to the INPC on a weekly basis, which provides a link between biospecimen samples and all three data sources.
We identified 6,854 patients 18 and older who had at least one 60-day period of continuous antipsychotic use between 2006 and 2018. The earliest date for which antipsychotics were ordered or prescribed was identified as the “first antipsychotic date.” The “treatment date” was set as first day of the 60-day period, unless that period occurred within 6 months of the “first antipsychotic date,” in which case the “treatment date” was set equal to the “first antipsychotic date.”
Patients without a diagnosis of schizophrenia or BD, type I at any time during the study period were categorized as a separate cohort. Patients with diagnoses of schizophrenia and schizoaffective disorder (hereafter referred to together as schizophrenia) (ICD-9 295X or ICD-10 F20 or F25X) or BD (ICD-9 296.0X, 296.1X, 296.4X-6X, 296.7, 296.80, 296.89, or ICD-10 F31X but not F31.81) were categorized as such. We acknowledge that those with schizoaffective disorder are often separated or excluded from studies of schizophrenia, but we feel here that because of the similarities in clinical presentation it was reasonable to group them. Both are classified together in the Diagnostic and Statistical Manual of Mental Disorders as Schizophrenia Spectrum Disorders and share the same core symptoms of psychosis along with a similar age of onset and course of illness, and have long been subject to debate about the validity and clinical utility of any distinction between the diagnoses [11, 12]. Indeed, depending on the quality of the available history and collateral information to describe the course of illness, these diagnoses can be indistinguishable in their presentations at any given time. Interrater reliability of schizoaffective disorder is known to be lower than that of schizophrenia and bipolar disorder , meaning that this diagnosis is often subject to change and should be reexamined frequently by clinicians. In the American Psychiatric Association practice guideline, the treatment of schizophrenia  includes many studies that included individuals with schizoaffective disorder, noting that these data were rarely analyzed separately, with the result that there is no distinct practice guideline for schizoaffective disorder. Patients with diagnoses for both schizophrenia and BD were categorized based on their most recent diagnosis, as mental health diagnoses routinely change over time within EHRs. If the most recent encounter included diagnoses from both categories, the patients were excluded from our cohort.
For each patient we collected age as of the treatment date, race, gender, and insurance status. Using information from EHR records, we identified comorbidities of interest, prior acute care utilization, and previous medication use during the one-year period prior to the treatment date. We defined adherence to antipsychotics using the proportion of days covered (PDC). The PDC is calculated by dividing the number of days the medication is available (from prescription dates) by the number of days in the period of interest. Patients with a PDC ≥ 0.8 were considered to be adherent. Adherence reflected adherence to any antipsychotic medication: if a patient switched medications during the study period, both medications contributed to the PDC calculation. In addition to adherence, we identified medication switches and medication stoppages during the first 6 months of treatment. A medication switch was defined as a new antipsychotic prescription that was different than the initial antipsychotic medication. Medication stoppage was defined as the lack of an antipsychotic medication order or SureScripts dispensing data in the EHR after the date when the previous supply would have been depleted.
To examine acute care utilization after the initiation of antipsychotic therapy, we identified all-cause and mental health-related ED and inpatient admissions during the first six months after the treatment date (a period of observed adherence) and from 6 to 12 months after the treatment date. In addition to overall acute care utilization, ED visits and inpatient admissions were also classified by whether they were for mental health-related care based on a combination of diagnoses, service location, and provider specialty codes.
Given that weight gain has been cited as a risk factor for non-adherence to antipsychotic medications [14,15,16,17,18,19], we incorporated measures of weight gain into our analyses. Weight gain was calculated as the difference between baseline weight and follow-up weight. Baseline weight was defined as the most recent weight recorded in the EHR prior to the treatment date, not more than 90 days prior to the treatment date. Follow-up weight was defined as the weight recorded in the EHR closest to treatment date plus 90 days, not less than 60 days or more than 120 days after the treatment date. Significant weight gain was described as ≥ 7% increase in weight.
We used Chi-square tests to compare utilization across the three cohorts (schizophrenia, BD, those without either diagnosis) and by different measures of medication adherence and switching. Logistic regression assessed the relationship between utilization during three time points (from treatment date to 6 months post treatment date and from 6 to 12 months post treatment date) and by medication adherence (PDC ≥ 0.8, PDC 0.5 to 0.79, PDC < 0.5) and switching (yes versus no), adjusting for demographic and clinical characteristics. Additional sensitivity analyses were performed using mental health ED visits and inpatient admissions as the outcome. Finally, we performed sensitivity analyses including patients whose weight change could not be calculated.