Comparative effectiveness of antidepressants for depression using EHRs from two health systems

Abstract

Background

Depression is a prevalent mental health condition, and many patients are treated in primary care, yet optimal antidepressant selection remains uncertain. Real-world data can help address gaps in understanding comparative effectiveness. We examined comparative effectiveness of duloxetine and vortioxetine in patients with depression and evaluated heterogeneity in patient characteristics, comorbidities, and healthcare utilization using electronic health records (EHR) from two large health systems.

Methods

We conducted a retrospective cohort study using EHR data from Duke University Health System (DUHS) and Johns Hopkins Health System (JHHS) from 2014 to 2021. Eligible participants were adult patients prescribed duloxetine or vortioxetine. We defined a Narrow Cohort requiring a concurrent depression diagnosis and a Broad Cohort based on prescription records. Comparative effectiveness analyses employed propensity score weighting methods. The primary outcome was all-cause hospitalization within one year of cohort entry. Secondary outcomes included emergency department (ED) visits, a composite of hospitalization or ED visit, suicidal encounters, medication switching (either between the study drugs or from the study drugs to other psychiatric medications), and change in PHQ-9 score.

Results

The Narrow Cohort included 3,337 (2,937 prescribed duloxetine and 400 vortioxetine) patients in DUHS and 3,237 (2,921 prescribed duloxetine and 316 vortioxetine) in the JHHS. Across both sites, vortioxetine users were more likely to be white, privately insured, and to have prior psychiatric medication use, while duloxetine users had higher comorbidity burden. Hospitalization rates were similar between treatments in both systems (OR_DUHS: 0.84 [95% CI: 0.45, 1.58], OR_JHHS: 0.7 [0.47, 1.03]). Estimates for ED visits, composite outcomes, and suicidal encounters were also comparable. Vortioxetine users had higher odds of switching medications among DUHS patients (OR_DUHS: 1.75 [1.37, 2.23]) and JHHS patients (OR_JHHS: 1.31 [0.97, 1.77]). Findings were largely consistent between the Narrow and Broad Cohorts.

Conclusions

Comparative effectiveness estimates for duloxetine and vortioxetine using EHR data were largely consistent across two health systems, with no meaningful differences for most outcomes. Consistent findings across two cohorts support the robustness and generalizability of these results. These results suggest that harmonized multi-site EHR data can provide reliable evidence while highlighting the importance of accounting for health system context.

Clinical trial number

Not applicable.

Supplementary Information

The online version contains supplementary material available at 10.1186/s12888-026-08159-x.

Introduction

Depressive disorder, also known as depression, is a mental health condition characterized by recurrent low mood; physical symptoms such as insomnia, low energy, and fatigue; often includes impaired motivation; and can be associated with suicidal thoughts or behaviors [1]. Depression affects approximately 5% of individuals worldwide [2], with an estimated 20% of US adults experiencing depression at some point in their lives [3]. Depression can be associated with significant disability and functional impairment [4, 5]. In the United States, depression tends to be under-recognized and under-diagnosed, and patients depression are more likely to receive care in primary care settings [3].

Antidepressant medications, used alone or in combination with behavioral interventions, are a cornerstone of depression treatment. However, despite numerous randomized controlled trials (RCTs)—primarily placebo-controlled designs—and systematic reviews, there is inadequate evidence to support clinical benefit of any one antidepressant over others [3, 4, 6].

Duloxetine, approved by the FDA in 2004, is a serotonin norepinephrine reuptake inhibitor (SNRI) widely prescribed for depression as well as generalized anxiety disorder and chronic pain conditions such as diabetic neuropathy and fibromyalgia. Duloxetine’s broad range of approved indications contributes to its extensive clinical use [7, 8]. Vortioxetine, approved in 2013, is an antidepressant marketed to address both mood symptoms and cognitive dysfunction commonly associated with depression [9, 10]. Vortioxetine is often considered a second-line treatment option due to its higher cost [11, 12]. Although multiple RCTs have included both duloxetine and vortioxetine as treatment arms [13–16], studies have been limited to selective patient populations enrolled under controlled conditions for short-term treatment, leaving uncertainty about their comparative effectiveness in real-world settings.

Addressing the evidence gap in the treatment of depression requires data sources that reflect the complexity of everyday clinical practice. While RCTs provide essential efficacy data, their restrictive eligibility criteria often exclude patients with medical comorbidities, severe psychiatric symptoms, or elevated suicide risk, limiting the generalizability of findings to broader patient populations. Real-world data (RWD), particularly from electronic health records (EHRs), provide an opportunity to complement traditional trials by capturing large, heterogeneous patient populations, including those commonly underrepresented in RCTs [17]. Furthermore, by leveraging EHR data from diverse settings, researchers can generate evidence that better reflects routine practice.

In this study, we harmonize EHR data from two large academic health systems—Duke University Health System (DUHS) and Johns Hopkins Health System (JHHS)—to examine the comparative effectiveness of duloxetine and vortioxetine in patients with depression. We evaluate heterogeneity in patient characteristics, comorbidities, and hospital utilization, and address key challenges in data harmonization across systems. By applying standardized analytic methods to multi-site EHR data, this study aims to assess comparative effectiveness of the two study treatments in routine clinical care, evaluate heterogeneity across systems, and discuss the opportunities and challenges of conducting comparable analyses across multiple systems.

Methods

The use of EHRs from DUHS and JHHS for this study was approved by Duke IRB (Pro00108971) and JHU IRB (IRB0017538). Reporting follows the ISPOR Reporting Guidelines [18].

Data sources and generation of EHR cohorts

Patient cohorts were built using retrospective EHR data from DUHS and JHHS between January 1, 2014, and December 31, 2021. Data extraction, processing, and analysis were conducted independently within each health system, and harmonization rules were developed and finalized through joint meetings to ensure consistency across sites. At the time of the study, DUHS included one tertiary care hospital, two community-based hospitals, and a network of primary care and specialty clinics. The JHHS consists of six tertiary care hospitals and a large multi-specialty health care network with over 40 primary care and specialty centers in Maryland and Washington DC.

Figure 1 shows the four steps of cohort curation, and eFigure 1 illustrates an example of EHR records for a patient eligible for the study cohorts. We initially identified a large pool of eligible patients from DUHS and JHHS in Step 1 of cohort curation, based on encounter records in psychiatry or primary care settings and a broad set of mood disorder diagnoses documented either in encounter diagnoses or the problem list (see eTable 1 for corresponding ICD codes). This approach was used to account for depression being under-documented or inconsistently coded in routine clinical practice. Patients aged 18–65 who were prescribed either duloxetine or vortioxetine were included, with baseline defined as the encounter at which the first such prescription occurred (Step 2). To ensure adequate clinical history and follow-up, patients were required to have at least one medical encounter in their given health system within the 12 months preceding baseline (denoted as the burn-in period) at Step 3, and at least one follow-up encounter in the health system within 2 years after baseline at Step 4. Full cohort creation details are described in the eMethods in Supplementary Materials.

Fig. 1.

Consort diagram for the Narrow and Broad cohorts and the resulting number of patients at each step in the DUHS and JHHS. Detailed description of each step is described in eMethods of the supplementary material. ^$ See eTable 1 in Supplementary Materials for ICD code for mood disorder. ^# Baseline of the Narrow Cohort is defined as the visit with the first prescription of duloxetine or vortioxetine with diagnosis of depression. ^& See eTable 2 in Supplementary Materials for ICD code for depression. ^† Baseline of the Broad Cohort is defined as the visit with the first prescription of duloxetine or vortioxetine

Two EHR cohorts were created for this study: a Narrow Cohort (primary analytic cohort) and a Broad Cohort.

The Narrow Cohort comprises individuals prescribed either duloxetine or vortioxetine at a clinical encounter where a formal diagnosis of depression (see eTable 2 for ICD codes) was documented at the same clinical encounter. This definition enhances diagnostic specificity by requiring that treatment and depression diagnosis co-occur. Importantly, duloxetine has multiple approved indications beyond depression, including fibromyalgia, diabetic neuropathy, and chronic pain conditions [19, 20], which could otherwise introduce clinical differences between the duloxetine and vortioxetine groups. The Narrow Cohort minimizes inclusion of individuals prescribed duloxetine for non-depression indications, improving confidence that observed treatment effects reflect depression management.

The Broad Cohort, used as a secondary cohort, includes individuals prescribed either duloxetine or vortioxetine at a clinical encounter, regardless of whether a formal diagnosis of depression was recorded at that encounter. The Broad Cohort acknowledges the reality of under-documentation of psychiatric diagnoses in real-world practice. Including this broader cohort captures a larger, more heterogeneous, and potentially more representative population of individuals experiencing depression.

Given the nature of real-world clinical care as reflected in EHR data, patients may be prescribed additional psychiatric medication(s) (see eTable 5 for the list of psychiatric medications considered) beyond the study treatments at baseline, and initially prescribed treatment may change during follow-up. We analyzed all outcomes based on the initial treatment received at baseline, regardless of subsequent treatment changes.

Outcome measures

Endpoints were defined as outcomes occurring at any encounter (both inpatient and outpatient) within one year from the patient’s entry into the cohort (baseline). The primary outcome was hospitalization for any reason. Secondary outcomes included emergency department (ED) visit for any reason, a composite of hospitalization and ED visit, suicide attempt or ideation coded during a medical encounter (referred to as a suicidal encounter), and switching of study medication—defined as a change between study medications or a change from baseline medications to a different psychiatric treatment regimen. Medication switching was included as a pragmatic measure of treatment persistence in routine clinical practice. In addition, change in the Patient Health Questionnaire (PHQ-9) score [21], which measures the severity of depression, from baseline to one year of follow-up was evaluated in the DUHS cohort only; PHQ-9 data was not sufficiently available for the JHHS cohort.

Hospitalization and ED visits for any reason were used rather than restricting to psychiatric reasons to capture clinically meaningful healthcare utilization outcomes that are consistently and reliably recorded across systems. This approach is further supported by extensive evidence linking suboptimal treatment of behavioral health conditions, including depression, to increased utilization of urgent care and hospital services for medical problems [22, 23].

Baseline covariates

We extracted a total of 24 covariates that were potentially associated with the study treatment prescription and outcomes and that could be defined consistently across the two systems (see eTable 3). These include 6 demographic variables (age, sex, race, ethnicity, insurance, and the National Area Deprivation Index [24] [ADI] national rank); indicators for history of 9 comorbidities and the Elixhauser Comorbidity Index [25, 26]; prescription history of 3 medication classes (antidepressant and first and second-generation antipsychotic); and 5 categories of healthcare utilization history (experience of ED visit, hospitalization, any psychiatry department visit (including both inpatient and outpatient), and suicidal encounter; and frequency of any outpatient visit).

Demographic variables were measured at baseline. As illustrated in eFigure 1, comorbidities were identified at any time prior to baseline using diagnosis codes and problem lists to capture chronic conditions, while medication history and healthcare utilization were assessed during the burn-in period to capture recent treatment patterns most relevant to study treatment selection.

Data harmonization process

The study team, including clinicians and statisticians from both sites, met regularly to align cohort definitions, covariates, and outcomes and to ensure consistent interpretation of coding practices and clinical workflows across institutions. Statisticians at the two sites also exchanged and reviewed analytic code to ensure that cohort construction and analyses were implemented as consistently as possible.

Statistical analysis

We first compared sample characteristics descriptively between the two health systems. Then, to compare outcomes between duloxetine and vortioxetine groups, we applied propensity score weighting methods [27]. Propensity scores were estimated using two models, logistic regression and generalized boosted models (GBM), including all covariates. The model yielding better covariate balance between treatment groups, assessed using standardized mean differences (SMDs) [28], was selected for subsequent analyses. Logistic regression-based propensity scores were selected for all analyses because this model consistently outperformed GBM in achieving covariate balance across both health systems and cohorts based on SMDs.

Propensity score weights were calculated using the inverse probability of treatment (vortioxetine in the present study) weighting (IPTW) method to estimate the average treatment effect (ATE) [29]. To estimate the incidence of each binary outcome (all outcomes except the PHQ-9 score), a propensity score weighted logistic regression model was used and odds ratios and their 95% confidence intervals were reported. A propensity score weighted linear regression model was fitted for the PHQ-9 score change between 1 year and baseline, and the mean differences and their 95% confidence intervals were reported. All outcome models were adjusted for the same set of covariates used in the propensity score models. These models were applied separately to the Narrow and Broad Cohorts. Propensity score estimation and weight creation were conducted using the PSweight package in R [30], and outcome models were run using the glm package in R [31].

Results

DUHS and JHHS EHR cohorts and comparison of characteristics

The Narrow Cohort included 3,337 (2,937 prescribed duloxetine and 400 prescribed vortioxetine) patients in DUHS and 3,237 (2,921 prescribed duloxetine and 316 prescribed vortioxetine) in the JHHS (Fig. 1). The Broad Cohort comprised 9,770 patients in DUHS (8,611 prescribed duloxetine and 1,159 prescribed vortioxetine) and 9,586 patients in the JHHS (8,742 prescribed duloxetine and 844 prescribed vortioxetine). Duloxetine was prescribed to approximately 90% of patients and vortioxetine was prescribed to approximately 10% of patients at baseline in both health systems and in both cohorts.

Table 1 presents the baseline characteristics of patients in the Narrow Cohort from both DUHS and JHHS. Baseline demographic characteristics were generally comparable between the two health systems. Over 70% of patients were female, about 94% were non-Hispanic, and mean age was approximately 46 years old in both health systems. There was a higher proportion of Black patients at JHHS (31.9%) compared to DUHS (25.1%), and a greater reliance on public insurance (Medicaid, Medicare, etc.) at JHHS (40.3%) versus DUHS (32.6%). The DUHS Narrow Cohort had a higher proportion of patients with private insurance (62.5% vs. 48.0%) than JHHS. The DUHS cohort showed a slightly higher mean ADI national rank (49.5 [SD: 21.8] vs. 42.9 [SD: 26.2]), indicating greater socioeconomic disadvantage among DUHS patients.

Table 1.

Baseline characteristics of the Narrow Cohort

		DUHS			JHHS
		Duloxetine	Vortioxetine	Total	Duloxetine	Vortioxetine	Total
	N	2937	400	3337	2921	316	3237
Demographics	Age	46.5 (12.7)	44 (12)	46.2 (12.7)	45.7 (12.9)	45.0 (12.3)	45.7 (12.8)
	Sex
	Female	76.6%	69.8%	75.8%	70.1%	66.8%	69.8%
	Male	23.4%	30.2%	24.2%	29.9%	33.2%	30.2%
	Race
	White	67.8%	79.5%	69.2%	58.2%	81.6%	60.5%
	Black	26.8%	12.5%	25.1%	34.1%	12.3%	31.9%
	Other/Unknown	5.3%	8.0%	5.7%	7.8%	6.0%	7.6%
	Ethnicity
	Not Hispanic	94.5%	93.5%	94.4%	94.3%	94.3%	94.3%
	Hispanic	3.2%	3.0%	3.2%	4.0%	3.5%	3.9%
	Unknown	2.3%	3.5%	2.4%	1.7%	2.2%	1.8%
	Insurance
	Private	61.0%	73.8%	62.5%	47.7%	50.9%	48.0%
	Public	33.9%	23.0%	32.6%	41.0%	33.9%	40.3%
	Self-pay	5.1%	3.2%	4.9%	6.9%	7.6%	7.0%
	Missing	0.0%	0.0%	0.0%	4.3%	7.6%	4.7%
	Area Deprivation Index (ADI)*	50.5 (21.7)	42.2 (20.4)	49.5 (21.8)	43.6 (26.2)	36.1 (24.7)	42.9 (26.2)
Comorbidities	Diabetes mellitus	21.5%	15.0%	20.7%	18.6%	14.2%	18.2%
	Fibromyalgia	15.8%	15.2%	15.8%	9.2%	6.0%	8.9%
	Stroke	5.3%	3.5%	5.1%	4.5%	1.6%	4.2%
	Hypothyroidism	12.1%	15.5%	12.5%	11.2%	16.1%	11.7%
	Depression	71.2%	81.2%	72.4%	70.3%	78.8%	71.1%
	Schizophrenia/Schizoaffective	0.7%	0.5%	0.7%	2.2%	1.6%	2.1%
	Bipolar disorder	5.7%	9.0%	6.1%	8.0%	10.4%	8.2%
	Anxiety	58.4%	67.2%	59.5%	48.6%	54.4%	49.2%
	PTSD	16.2%	19.8%	16.6%	12.1%	15.2%	12.4%
	Elixhauser Comorbidity Index&	3.9 (3.2)	3.5 (3.1)	3.8 (3.2)	4.0 (3.3)	3.1 (2.6)	3.9 (3.2)
Medication History	Antipsychotics (first gen)	1.6%	3.0%	1.7%	6.7%	5.7%	6.6%
	Antipsychotics (second gen)	8.1%	16.5%	9.1%	14.1%	25.0%	15.1%
	Antidepressant	55.5%	67.5%	56.9%	50.9%	69.6%	52.7%
Healthcare Utilization History	Emergency department visits	25.7%	20.5%	25.1%	29.4%	18.0%	28.3%
	Hospitalizations	4.7%	4.0%	4.6%	25.4%	15.8%	24.4%
	Psychiatry department visits	17.1%	39.8%	19.8%	34.5%	55.4%	36.5%
	Suicidal encounter	3.3%	3.0%	3.3%	7.4%	5.1%	7.1%
	Num of any outpatient visits	13.6 (15.9)	15.8 (21.2)	13.9 (16.7)	6.7 (9.7)	6.9 (9.2)	6.7 (9.7)

Continuous variables presented as mean (standard deviation); categorical variables presented as %

^*The National Area Deprivation Index (ADI) national rank is a measure of socioeconomic disadvantage for geographic areas, where 1 indicates the least disadvantaged and 100 indicates the most disadvantaged

^& The Elixhauser Comorbidity Index measures patient comorbidity in administrative data by 31 comorbidity categories

Similar treatment patterns were observed across both health systems (Table 1). Patients prescribed duloxetine were more likely to be Black (DUHS: 26.8% versus 12.5%; JHHS: 34.1% versus 12.3%) and have a higher ADI (DUHS: 50.5 versus 42.2; JHHS: 43.6 versus 36.1), whereas those prescribed vortioxetine were more likely to be White (DUHS: 79.5% versus 67.8%; JHHS: 81.6% versus 58.2%) and privately insured (DUHS: 73.8% versus 61%; JHHS: 50.9% versus 47.7%).

Comorbidity profiles prior to baseline were consistent between DUHS and JHHS overall. However, within each system, patients prescribed duloxetine exhibited higher prevalence of diabetes mellitus (DUHS: 21.5% versus 15%; JHHS: 18.6% versus 14.2%), while those in the vortioxetine group had a slightly higher prevalence of depression (DUHS: 81.2% versus 71.2%; JHHS: 78.8% versus 70.3%), bipolar disorder (DUHS: 9% versus 5.7%; JHHS: 10.4% versus 8%), and anxiety (DUHS: 67.2% versus 58.4%; JHHS: 54.4% versus 48.6%). There were also differences by treatment group in the Elixhauser Comorbidity Index, with duloxetine users exhibiting slightly higher scores (DUHS: 3.9 versus 3.5; JHHS: 4.0 versus 3.1) than vortioxetine users.

Overall, antidepressant use within 1 year prior to baseline was slightly more prevalent in DUHS (56.9%) than JHHS (52.7%). Patients in the vortioxetine group had higher rates of previous second-generation antipsychotic use (DUHS: 16.5% versus 8.1%; JHHS: 25% versus 14.1%) and antidepressant use (DUHS: 67.5% versus 55.5%; JHHS: 69.6% versus 50.9%) compared to duloxetine users.

For healthcare utilization within 1 year prior to baseline, in both DUHS and JHHS, duloxetine users had more emergency department (ED) visits (DUHS: 25.7% versus 20.5%; JHHS: 29.4% versus 18%), while vortioxetine users had more psychiatry department visits (DUHS: 39.8% versus 17.1%; JHHS: 55.4% versus 34.5%). Although the mean number of outpatient visits was similar between treatment groups within each site, DUHS patients had a higher overall number of outpatient visits (mean 13.9) compared to JHHS patients (mean 6.7). JHHS had a substantially higher rate of hospitalization (JHHS total: 24.4%) than DUHS (total: 4.6%), likely reflecting the availability of inpatient psychiatric services at JHHS, which DUHS lacks.

The Broad Cohort showed similar findings (eTable 6).

Comparative effectiveness between duloxetine and vortioxetine

After weighting, nearly all covariates had SMDs below the conventional threshold of 0.1, and the small number of remaining covariates with slightly higher SMDs still showed substantial improvement compared with the unweighted data (eFigures 2 and 3, eTable 7).

Figure 2 displays the ATE estimates for clinical outcomes in the Narrow Cohort.

Fig. 2.

Average treatment effect estimates in the Narrow Cohort from each health system (a) DUHS and (b) JHHS. Odds ratios greater than 1 indicate that vortioxetine was associated with higher odds of the outcome compared to duloxetine

Vortioxetine was associated with lower odds of hospitalization compared to duloxetine in both health systems (OR_DUHS: 0.84 [95% CI: 0.45, 1.58], OR_JHHS: 0.7 [0.47, 1.03]), although the CIs included the null. For ED visits alone (OR_DUHS: 1.17 [95% CI: 0.87, 1.58], OR_JHHS: 0.81 [0.52, 1.26]) and composite of hospitalization or ED visit (OR_DUHS: 1.06 [95% CI: 0.81, 1.39], OR_JHHS: 0.81 [0.58, 1.13]), the odds ratios showed opposite directions between the two health systems, with confidence intervals including the null in both systems. For suicidal encounters, although odds ratios favored vortioxetine, the estimates were imprecise in both systems. For medication switching, higher odds were observed among vortioxetine users compared with duloxetine in DUHS (OR_DUHS: 1.75 [1.37, 2.23]), whereas the estimate in JHHS was smaller and more uncertain (OR_JHHS: 1.31 [0.97,1.77]).

Among DUHS patients with available baseline and follow-up PHQ-9 scores (n = 1,312; 1,191 in duloxetine and 121 in vortioxetine), duloxetine users demonstrated a greater reduction in PHQ-9 scores over one year compared to vortioxetine users (mean difference: 2.43 [0.80, 4.07]), suggesting potentially greater effectiveness of duloxetine over vortioxetine in improving depressive symptoms among patients with observed scores. However, this finding should be interpreted with caution due to the small number of vortioxetine users included in this analysis.

The Broad Cohort showed overall similar conclusions with some exceptions (eTable 8), including lower odds of hospitalization for patients prescribed vortioxetine in DUHS (OR_DUHS: 0.58 [0.40, 0.83]), higher odds of medication switching in vortioxetine users across both systems (OR_DUHS: 1.72 [1.45, 2.03], OR_JHHS: 1.36 [1.14,1.62]), and no clear difference in PHQ-9 score reduction between treatment groups in DUHS (mean difference: 0.53 [-0.65, 1.71]).

Discussion

In this study, we evaluated the comparative effectiveness of duloxetine versus vortioxetine using harmonized EHR data from two large health systems. Overall, treatment effect estimates for most outcomes, including the primary outcome, hospitalization for any reason, were largely consistent between systems, demonstrating the robustness and reproducibility of findings across sites. Although the point estimates for the ED visit outcome differed, their 95% CIs included the null, indicating no meaningful between-site differences. These results support the feasibility of conducting multi-site EHR-based analyses using standardized analytic approaches.

When comparing characteristics of the study cohorts across systems, patient populations were generally comparable with respect to most demographic and clinical variables. Although some differences were observed, including race, insurance type, and area-level socioeconomic status, differences were modest and consistent with known population characteristics of the two systems. Healthcare utilization patterns differed more substantially, particularly for hospitalization, which was considerably higher at JHHS than at DUHS. These between-site differences in our analysis likely reflect structural and contextual differences between the two systems, including the availability of greater inpatient psychiatric capacity at JHHS (over 124 inpatient psychiatric beds) than DUHS (42 inpatient psychiatric beds). Despite these differences, comparative effectiveness results were consistent across systems. These results suggest that between-site differences do not necessarily translate into inconsistent comparative effectiveness estimates when patient case-mix is appropriately controlled using standard statistical methods.

In addition to patient-level and system-level differences, prescribing patterns may also be influenced by structural factors such as medication cost and time on the market. Duloxetine has been available in generic form since 2013, prior to the study period, whereas vortioxetine remained under patent during the study period and is substantially more expensive. These differences may influence prescribing in routine practice and contribute to variation in treatment selection across patient characteristics and health systems, potentially reflecting variation in prior treatment exposure, disease severity, and socioeconomic constraints affecting medication access.

Leveraging data from multiple EHR systems presents substantial methodological and logistical challenges. Harmonizing data across distinct health systems requires close collaboration and iterative discussions to ensure a shared understanding of clinical workflows, documentation practices, and population differences (e.g., state-level healthcare infrastructure, demographic composition, and socioeconomic contexts). Physician input is essential for interpreting variations in care delivery, diagnostic coding, and prescribing behavior that may reflect local practice norms or patient needs. Defining cohorts consistently—particularly establishing the baseline encounter—and ensuring uniform application of diagnostic codes such as ICD-10 are critical to preserving internal validity. Furthermore, effective integration of expertise from data scientists, statisticians, and clinicians is essential to ensure that such cross-system differences are accurately captured in cohort construction and appropriately addressed in analytic approaches.

Despite these challenges, our study highlights the value of EHR data as a RWD source, and of using EHR data from multiple health systems. EHR-based analyses enable the evaluation of treatment effects in real-world patient populations receiving routine clinical care. Importantly, our study captured outcomes—such as medication switching and suicidal encounters—that have not been consistently assessed in existing RCTs comparing duloxetine and vortioxetine, and allowed longer follow-up than typical RCTs. This illustrates how EHR data can be carefully leveraged to generate clinically relevant insights that address important gaps left by traditional trial designs.

In addition, EHR data analysis allows flexibility in studying two slightly different cohorts. The Narrow Cohort was designed to increase diagnostic specificity by requiring a documented depression diagnosis, whereas the Broad Cohort captures a more heterogeneous population that reflects real-world prescribing patterns where depression may be under-documented in the EHR. In our analyses, the baseline covariate distributions in the two cohorts were similar and the treatment effect estimates were largely comparable between the Narrow and Broad Cohorts. These findings suggest that the comparative effectiveness results were reasonably robust to alternative cohort definitions. The consistency of results across these cohorts supports the applicability of our conclusions to both more strictly defined and broader real-world patient populations, which may represent distinct but clinically relevant groups in practice. As such, this analytic approach demonstrates how an EHR-based study can be used to conduct parallel analyses across different populations, thereby enhancing the generalizability of the findings.

Our analysis highlights that, in multi-site EHR studies, understanding and interpreting between-site heterogeneity is as important as applying harmonized statistical methods, and that site-specific results should be interpreted within their clinical and health-system context rather than assumed to generalize across institutions.

Limitations

This study has multiple limitations. Treatment groups were defined based on the first recorded prescription in the EHR, which does not guarantee that patients initiated or adhered to the prescribed medication, as EHRs capture prescribing behavior but not medication dispensing or compliance. Additionally, patients may have experienced varying durations of treatment exposure, which we were unable to fully account for; thus, our analysis was based on the initial treatment received at baseline, with medication switching considered as an outcome rather than a censoring event. All outcomes except for depressive symptoms were analyzed as binary events over a one-year follow-up period, without accounting for time-to-event or censoring, which may limit the precision of effect estimates. Furthermore, change in depressive symptoms was assessed using PHQ-9 scores, which were only available for a subset of patients and measured at non-standardized intervals. COVID-19-related hospitalizations and ED visits were not excluded; however, fewer than 2% of hospitalization events in the sample were associated with a COVID-19 diagnosis ICD-10 code (U07.1). Completed suicide and suicide planning were not included because they are not reliably captured in structured EHR data and require external mortality data linkage. EHR data often suffer from observability issues, such as incomplete capture or recording errors that are difficult to fully address [32]. Individual-level data could not be transferred between sites. Federated learning methods [33] were not considered, as the study involved only two participating health systems. Finally, propensity score methods adjust only for observed covariates and cannot account for unmeasured confounding.

Conclusions

This study provides real-world evidence comparing duloxetine and vortioxetine for depression using harmonized EHR data from two large health systems. Comparative effectiveness estimates were largely consistent across systems and between the Narrow and Broad cohorts, with no meaningful differences for most outcomes. These findings highlight the importance of considering health system context and support the use of harmonized multi-site EHR data to generate robust and generalizable evidence for clinical decision-making.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Drs. Hong and Stuart had full access to the DUHS and JHHS data, respectively, and take responsibility for the integrity of the data and the accuracy of the data analysis. Concept and design: Hong, Stuart. Acquisition, analysis, or interpretation of data: Hong, Zhao, Jeon, Goicoechea, Zandi, Stuart. Drafting of the manuscript: Hong. Critical review of the manuscript for important intellectual content: All authors. Statistical analysis: Hong, Zhao, Jeon, Goicoechea, Zandi, Stuart. Obtained funding: Hong, Stuart. Supervision: Hong, Stuart.

Funding

The study was funded by the Patient-Centered Outcomes Research Institute (PCORI) through PCORI Award ME-2020C3-21145 (PI: Stuart) and the National Institute of Mental Health (NIMH) through Award R01MH126856 (PI: Stuart).

Data availability

The datasets generated and analyzed during the current study are not publicly available because they consist of electronic health record data, which cannot be shared due to institutional and regulatory privacy protections. Aggregate results or additional details are available from the corresponding author upon reasonable request.

Declarations

Ethics approval and consent to participate

Ethical approval of the study was obtained from the Duke IRB (Pro00108971) and the Johns Hopkins University IRB (IRB0017538). Because this was a retrospective data analysis, informed consent from participants was not required. All study procedures were conducted in accordance with relevant ethical guidelines and regulations, adhering to the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.