Signals of Venous Thromboembolism Risk from Drug–Drug Interactions Involving Benzodiazepines and Other Sedative Hypnotics: Self-Controlled Case Series Analyses Authors: Fengge Wang, Lin-Chieh Meng, Kacie Bogar, Colleen M. Brensinger, Warren B. Bilker, Sascha Dublin, Todd E. H. Hecht, John R. Horn, Daniela C. Moga, Michael E. Thase, Brian L. VanderBeek, Charles E. Leonard Published: 2026 ABSTRACT Background and ObjectiveBenzodiazepines and sedative hypnotics such as zolpidem (“z-drugs”) are commonly prescribed for anxiety and sleep disorders. Epidemiologic evidence links their use to increased risk of venous thromboembolism. We investigated venous thromboembolism risk among concomitant users of individual benzodiazepines/z-drugs (examined separately) with other prescription medications to generate data-driven hypotheses about drug interactions resulting in clinically meaningful harm to inform future etiologic studies of specific drug combinations.MethodsWe conducted a series of self-controlled case series studies within a 50% random sample of US Medicaid and Medicare data. Each cohort comprised person-time exposed to a benzodiazepine/z-drug, dichotomized into focal versus referent periods based on concomitant drug use versus non-use. We used conditional Poisson regression to estimate incidence rate ratios for hospital or emergency department presentation for venous thromboembolism. We generated ratios of incidence rate ratios, leveraging negative control analyses of eye drop–concomitant drug pairs, to minimize confounding by indication for the concomitant drug. We used semi-Bayes shrinkage to minimize false positives.ResultsAmong 1590 self-controlled case series studies involving 8853 individuals with venous thromboembolism, 38 (2.4%) potential drug interaction signals were identified before calibration. After adjustment for multiple testing and negative control findings, five (0.3%) signals remained, involving gabapentin combined with eszopiclone, lorazepam, clonazepam, or alprazolam, and apixaban combined with diazepam (ratio of incidence rate ratio range: 1.92–3.57). Four (80%) involved concurrent use of gabapentin, a medication largely used to treat neuropathic pain.ConclusionsMost benzodiazepine/z-drug combinations conferred no increased venous thromboembolism risk. However, concurrent use of a benzodiazepine/z-drug with gabapentin may increase the relative rate of venous thromboembolism up to 3.5-fold. As this work was hypothesis generating, a future etiologic study should confirm this potential drug interaction.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s40261-026-01582-8. Key Points We screened for potential drug–drug interaction signals associated with venous thromboembolism risk among benzodiazepines/z-drug users.Concurrent use of benzodiazepines/z-drugs and gabapentin was associated with a potential increase in venous thromboembolism risk.These findings should be interpreted as hypothesis-generating signals, and future etiologic studies are needed to evaluate causality and underlying mechanisms. We screened for potential drug–drug interaction signals associated with venous thromboembolism risk among benzodiazepines/z-drug users.Concurrent use of benzodiazepines/z-drugs and gabapentin was associated with a potential increase in venous thromboembolism risk.These findings should be interpreted as hypothesis-generating signals, and future etiologic studies are needed to evaluate causality and underlying mechanisms. Introduction Benzodiazepines (BZDs) are used to treat anxiety, panic disorder, insomnia, and seizures, among other conditions [1]. Benzodiazepines bind to the gamma-aminobutyric acid (GABA)-A receptor and increase its affinity for GABA, thereby enhancing GABA’s main effect on central nervous system depression [2]. Benzodiazepine use is consistently high among many Western European nations, although the prevalence differs by country [3–5]. Similarly, BZDs are very commonly prescribed in the USA. According to a survey by the Substance Abuse and Mental Health Services Administration, >30 million American adults reported using BZDs in the past year [6]. Z-drugs (i.e., zolpidem) work similarly to BZDs in treating insomnia by affecting GABA but have different chemical structures. Approximately 4% of individuals aged 12 years and older reported past-year use or misuse of z-drugs, according to the US National Survey on Drug Use and Health [7]. Benzodiazepines, and to a lesser extent z-drugs, have been linked to an increased risk of venous thromboembolism (VTE) in some epidemiologic studies [8], including its components deep vein thrombosis (DVT) and pulmonary embolism (PE) [9]. The mechanisms by which BZDs increase VTE risk are incompletely understood but may relate to sedative effects and subsequent reduced mobility [9], which can lead to venous stasis and promote a prothrombotic state. Because BZD and z-drug users are often exposed to multiple concomitant medications [10, 11], such as antidepressants [12], mood stabilizers [13], opioids [14], and cardiovascular drugs [11], they may be particularly susceptible to clinically relevant drug interactions. Some of these co-medications may further impair alertness and mobility, while others may inhibit the hepatic cytochrome P450 (CYP) metabolism of certain BZDs or z-drugs (principally but not solely via CYP3A4 and CYP2C19) [2])—potentially resulting in pharmacodynamic and/or pharmacokinetic drug interactions. As an example of the latter, the BZD alprazolam is primarily metabolized via CYP3A4, so co-administration of a strong CYP3A4 inhibitor (e.g., the antidepressant nefazodone) can elevate alprazolam plasma concentrations [15], which may enhance possible BZD-associated VTE risk. While more than half a dozen epidemiologic studies have attempted to quantify VTE risk among BZD users [16], very little is known about how concomitantly used drugs may influence this risk. Given the complexity of patient profiles and treatment regimens among BZD users, studying the impact of potential BZD drug interactions on VTE risk could inform safer prescribing practices and targeted interventions, which could reduce thromboembolic complications in vulnerable groups. We conducted a longitudinal study of US Centers for Medicare and Medicaid Services (CMS) beneficiaries to identify drug pairs including a BZD or z-drug that increase VTE risk. Given the high-throughput nature of our work, we framed this study as hypothesis generating, with the goal of identifying potential BZD and z-drug drug interactions that should be prioritized for further investigation. Methods Data Source, Study Population, and Cohort Definitions This study was conducted within the Self-controlled Crossover Observational Pharmacoepidemiologic (SCOPE) framework [17], which encompasses a family of self-controlled pharmacoepidemiologic designs that use within-person comparisons to control for time-invariant confounding. Specifically, we applied an exposure-anchored self-controlled case series (SCCS) approach using healthcare enrollment and utilization data from CMS between 2016 and 2021. Medicaid and Medicare represent the two largest US government-funded health insurance initiatives and cover a diverse population [18]. Medicaid provides healthcare for >80 million individuals, including those with a low income, disabled persons, older adults, and pregnant women [19]. Medicare provides healthcare for > 60 million individuals, the vast majority of whom are older adults aged 65+ years, but also individuals of any age with a long-term disability, end-stage renal disease, and/or amyotrophic lateral sclerosis [20]. Within a 50% random sample of CMS data including both Medicare and Medicaid, we established two analytic cohorts: (1) individuals dispensed orally administered BZDs, z-drugs, or a hypnotic (suvorexant), with each active ingredient/medication analyzed independently; and (2) individuals dispensed ophthalmic or topical prostaglandin analogs/receptor agonists (“eye drops”), treated collectively as the prespecified negative control group. Medicare and Medicaid data were integrated at the individual beneficiary level using unique encrypted identifiers available within the data source (see Table S5 of the Electronic Supplementary Material [ESM]). Because of small cell sizes in certain subgroups (e.g., dual-eligible beneficiaries), detailed stratified reporting was limited in accordance with data use agreement requirements. Cohorts were constructed by identifying drug dispensings for corresponding National Drug Codes within Medicaid Analytic eXtract/Transformed Medicaid Statistical Information System Analytic Files and Medicare Part D prescription records (see Table S1 of the ESM). In this study, we limited our study of BZDs and z-drugs to oral products. Of relevance to cohorts and exposures, we use neutral terms such as cohort-defining and concomitantly used drugs (respectively) instead of object drug/precipitant drug labels, as our study was designed to be agnostic to the specific mechanism underlying any given drug pair, while being motivated by known pharmacodynamic and pharmacokinetic interaction pathways. We conducted separate SCCS analyses for each drug pair within their respective drug cohorts. The within-person nature of this epidemiologic method intrinsically adjusts for time-invariant measured and unmeasured individual characteristics, thereby effectively addressing confounding by static factors [21, 22]. Further methodologic detail is found in our prior work [23–35]. For each SCCS cohort, eligible participants were adult Medicare or Medicaid enrollees (18–100 years of age) who had continuous enrollment during a 365-day baseline period, with no history of receiving the cohort-defining drug or experiencing the study outcome of interest (VTE diagnosed in the inpatient or emergency department [ED] setting) during baseline. Within each cohort, observation intervals were determined by use of the cohort-defining drug, identified via days’ supply information recorded on prescription claims. To better reflect typical real-world adherence patterns, a 10% grace period was incorporated into the days’ supply calculation to define continuous use. Follow-up ended at the first occurrence of: discontinuation of the cohort-defining medication (accounting for the grace period), switching to a different cohort-defining medication, loss of health plan coverage, end of the study period, or death. Exposure Ascertainment Time-varying exposure, assessed at the observation day level, was defined by use versus non-use of a concomitant drug—also determined via days’ supply data from prescription claims. We did not apply a grace period for concomitantly dispensed drugs, as the potential bias from exposure misclassification could be greater than that arising from observation period misclassification within the setting of the SCCS approach [36]. Aligned with preferred SCOPE terminology, we considered exposed periods as focal time and unexposed periods as referent time. Detailed definitions of the focal and referent periods were similar to our prior work and can be found in Fig. 1. In our study design, exposure classification was anchored on BZD/z-drug treatment episodes. Time prior to BZD/z-drug initiation was not included in the risk period definition. We did not include an indeterminate/washout period following each focal exposure because of the infrequent occurrence of relevant events during these intervals. We did not proceed with further analyses if there were fewer than five concomitant drug-exposed persons in a drug pair cohort. Because comparisons are made within individuals across exposed and unexposed periods anchored to BZD/z-drug treatment episodes, the analysis isolates the incremental risk associated with concomitant use rather than the baseline risk associated with long-term therapy.Fig. 1General overview of each self-controlled case series study, including depictions of baseline, observation, focal, and referent periods. A drug episode was defined using days’ supply values on prescription claims. We allowed a modest grace period for benzodiazepine (BZD)/z-drug/nonbarbiturate sedative/hypnotic dispensings (i.e., a 10% extension of the days’ supply). If the next prescription overlapped with the previous prescription (including the grace period), it was regarded as part of the same episode. The baseline period was defined as 365 days prior to a newly initiated BZD/z-drug/nonbarbiturate sedative/hypnotic episode. The focal period was defined by exposure to both the BZD/z-drug/nonbarbiturate sedative/hypnotic and the concomitantly dispensed drug. We did not include a grace period for concomitantly dispensed drugs. The referent period was defined as observation time not otherwise defined as focal time. Of note, our study permitted prior use of a concomitant drug dispensed before the observation-defining BZD/z-drug/nonbarbiturate sedative/hypnotic episode General overview of each self-controlled case series study, including depictions of baseline, observation, focal, and referent periods. A drug episode was defined using days’ supply values on prescription claims. We allowed a modest grace period for benzodiazepine (BZD)/z-drug/nonbarbiturate sedative/hypnotic dispensings (i.e., a 10% extension of the days’ supply). If the next prescription overlapped with the previous prescription (including the grace period), it was regarded as part of the same episode. The baseline period was defined as 365 days prior to a newly initiated BZD/z-drug/nonbarbiturate sedative/hypnotic episode. The focal period was defined by exposure to both the BZD/z-drug/nonbarbiturate sedative/hypnotic and the concomitantly dispensed drug. We did not include a grace period for concomitantly dispensed drugs. The referent period was defined as observation time not otherwise defined as focal time. Of note, our study permitted prior use of a concomitant drug dispensed before the observation-defining BZD/z-drug/nonbarbiturate sedative/hypnotic episode Outcome Ascertainment The outcome was hospital or an ED encounter for an acute VTE, including the presence of DVT, PE, and other thrombotic events occurring during observation time. We identified outcomes through ED and inpatient claims, relying on primary discharge International Classification of Diseases, Tenth Revision, Clinical Modification diagnoses. Our VTE code set was driven by prior algorithms developed by the US Agency for Healthcare Research and Quality patient safety indicator and Kim et al. [37–39] (Table S2 of the ESM). Major components of our definition have been independently validated. Bikdeli et al. reported a positive predictive value of 92% for principal position discharge diagnosis PE codes in data from Mass General-Brigham hospitals [40]. Kim et al. reported a positive predictive value of 76% for a composite of PE, DVT, and other venous thromboses in ED, inpatient, and outpatient data from commercially and/or US Medicare-insured patients treated for rheumatoid arthritis [39]. A Danish study using a similar DVT algorithm reported a positive predictive value of 80% [41]. Covariates Because the SCCS study design inherently controls for measured and unmeasured time-invariant confounders, we adjusted for the following dynamic covariates to address time-varying confounding: hypnotic daily dose (categorized in quartiles); prior VTE history during the follow-up period; and follow-up duration—all measured at the observation day level. Although individuals were required to have no history of VTE during baseline, VTE events occurring during the follow-up could subsequently contribute to the time-varying VTE history covariate in later BZD/z-drug exposure periods. Thus, while cohort entry required no baseline VTE to focus on incident events, recurrent events during the follow-up were permitted and accounted for via the time-varying history indicator. Statistical Analysis We used multi-level conditional Poisson regression to calculate adjusted incidence rate ratios (IRRs) and 95% confidence intervals (CIs) comparing the rate of VTE during focal versus referent periods. To prevent model instability, analyses were excluded if they yielded fewer than 11 outcome events in the observation periods, suffered from non-convergence, or produced variance of the beta coefficient exceeding an absolute value of ten. Because of the very large number of comparisons, to minimize the risk of generating false-positive drug interaction signals, we utilized semi-Bayesian shrinkage to stabilize extreme IRR estimates by borrowing information across drug pairs, thereby reducing the likelihood of false-positive drug interaction signals. This methodology is supported by prior Bayesian approaches to multiple comparison adjustment [42–44]. To minimize confounding by indication for the co-administered drug, we incorporated a negative control cohort. Specifically, ophthalmic and topical prostaglandin (“eye drop”) exposures were selected as the prespecified negative control group because they are prescription products captured within the same healthcare utilization and dispensing framework as the study drugs but are not expected to plausibly increase VTE risk. Thus, they provide a reference for nonspecific associations driven by healthcare utilization patterns, prescribing complexity, or the underlying clinical context unrelated to BZD/z-drug pharmacology. For each BZD/z-drug and concomitant drug pair, the semi-Bayes-adjusted IRR was divided by the IRR estimated for the corresponding negative control eye drop-concomitant drug pair, with the 95% CI calculated using the delta method. This ratio-based calibration helps account for bias driven by indication or the healthcare context. The reporting of this study herein adheres to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) recommendations for observational research. To enhance transparency, the full analytic plan was registered to the Open Science Framework (https://doi.org/10.17605/OSF.IO/7EUXD) by CEL prior to initiating analyses. Institutional Review Board Approval This study was reviewed and approved by the University of Pennsylvania Institutional Review Board (protocol number 854130). Waivers of informed consent and of Health Insurance Portability and Accountability Act authorization were granted given the study’s minimal risk. Results Description of the Study Cohorts There were 22 BZD/z-drug cohorts, 14 of which met inclusion criteria, within which we conducted 1590 SCCS studies. There were 261 studies of clonazepam, 253 of lorazepam, 235 of alprazolam, 231 of zolpidem, 181 of diazepam, 172 of temazepam, 84 of eszopiclone, 51 of ramelteon, 42 of chlordiazepoxide, 31 of zaleplon, 28 of suvorexant, 10 of triazolam, 8 of clobazam, and 3 of clorazepate. These studies included a total of 8853 persons, each of whom experienced at least one VTE, consisting of 807,560 observation days. The median observation time per person ranged from 11 days for chlordiazepoxide to 62 days for eszopiclone. The majority of BZD and z-drug users were non-Hispanic White, female (except chlordiazepoxide users), and older adults (median age in years ranged from 52.6 for clobazam to 73.4 for clorazepate and temazepam) (Table 1).Table 1Characteristics of cohorts underlying the SCSSBenzodiazepine, z-drug, and nonbarbiturate sedative/hypnotic cohortsAlprazolamChlordiazepoxideClobazamClonazepamClorazepateDiazepamEszopiclonePersons, each of whom experienced ≥ 1 VTE outcome as dictated by the SCCS design N18018125136417934131VTE events N199191281541181018141Days of observation, per person Median (Q1–Q3)34 (18.0–90.0)11 (7.0–34.0)42 (34.0–92.0)55 (34.0–147.0)34 (18.0–46.0)18 (9.0–48.0)62 (34.0–175.0)Demographics Age, in years Median (Q1–Q3)69.8 (55.5–78.1)57 (45.6–70.5)52.6 (42.0–63.1)60.9 (45.3–73.7)73.4 (67.2–82.4)60.6 (47.5–71.7)67.7 (55.0–75.8) Sex, female N (%)1176 (65.3%)37 (45.7%)13 (52.0%)841 (61.7%)13 (76.5%)544 (58.2%)88 (67.2%)Benzodiazepine, z-drug, and nonbarbiturate sedative/hypnotic cohortsLorazepamRamelteonSuvorexantaTemazepamTriazolamZaleplonZolpidemPersons, each of whom experienced ≥ 1 VTE outcome as dictated by the SCCS design N2297804862317581377VTE events N2447835266618631467Days of observation, per person Median (Q1–Q3)29 (12.0–43.0)38 (33.0–110.5)34 (34.0–72.0)37 (33.0–99.0)44 (23.0–74.0)34 (31.0–91.0)34 (31.0–91.0)Demographics Age, years Median (Q1–Q3)69.5 (56.3–77.7)70.5 (59.6–79.1)69.5 (62.0–78.6)73.4 (64.5–80.3)69.1 (53.2–80.2)67.4 (54.3–76.8)69.2 (57.8–76.7) Sex, female N (%)1,486 (64.7%)48 (60.0%)32 (66.7%)341 (54.7%)n < 11b (not reported)35 (60.3%)756 (54.9%)Q quartile, SCCS self-controlled case series, VTE venous thromboembolism, z-drug non-benzodiazepine hypnoticaThe nonbarbiturate sedative/hypnotic suvorexant is a novel agent that antagonizes orexin receptors. It does not bind to or modulate gamma-aminobutyric acid receptorsbPer patient privacy and de-identification requirements, cells containing fewer than 11 patients are suppressed from reporting Table 1Characteristics of cohorts underlying the SCSSBenzodiazepine, z-drug, and nonbarbiturate sedative/hypnotic cohortsAlprazolamChlordiazepoxideClobazamClonazepamClorazepateDiazepamEszopiclonePersons, each of whom experienced ≥ 1 VTE outcome as dictated by the SCCS design N18018125136417934131VTE events N199191281541181018141Days of observation, per person Median (Q1–Q3)34 (18.0–90.0)11 (7.0–34.0)42 (34.0–92.0)55 (34.0–147.0)34 (18.0–46.0)18 (9.0–48.0)62 (34.0–175.0)Demographics Age, in years Median (Q1–Q3)69.8 (55.5–78.1)57 (45.6–70.5)52.6 (42.0–63.1)60.9 (45.3–73.7)73.4 (67.2–82.4)60.6 (47.5–71.7)67.7 (55.0–75.8) Sex, female N (%)1176 (65.3%)37 (45.7%)13 (52.0%)841 (61.7%)13 (76.5%)544 (58.2%)88 (67.2%)Benzodiazepine, z-drug, and nonbarbiturate sedative/hypnotic cohortsLorazepamRamelteonSuvorexantaTemazepamTriazolamZaleplonZolpidemPersons, each of whom experienced ≥ 1 VTE outcome as dictated by the SCCS design N2297804862317581377VTE events N2447835266618631467Days of observation, per person Median (Q1–Q3)29 (12.0–43.0)38 (33.0–110.5)34 (34.0–72.0)37 (33.0–99.0)44 (23.0–74.0)34 (31.0–91.0)34 (31.0–91.0)Demographics Age, years Median (Q1–Q3)69.5 (56.3–77.7)70.5 (59.6–79.1)69.5 (62.0–78.6)73.4 (64.5–80.3)69.1 (53.2–80.2)67.4 (54.3–76.8)69.2 (57.8–76.7) Sex, female N (%)1,486 (64.7%)48 (60.0%)32 (66.7%)341 (54.7%)n < 11b (not reported)35 (60.3%)756 (54.9%)Q quartile, SCCS self-controlled case series, VTE venous thromboembolism, z-drug non-benzodiazepine hypnoticaThe nonbarbiturate sedative/hypnotic suvorexant is a novel agent that antagonizes orexin receptors. It does not bind to or modulate gamma-aminobutyric acid receptorsbPer patient privacy and de-identification requirements, cells containing fewer than 11 patients are suppressed from reporting Characteristics of cohorts underlying the SCSS Q quartile, SCCS self-controlled case series, VTE venous thromboembolism, z-drug non-benzodiazepine hypnotic aThe nonbarbiturate sedative/hypnotic suvorexant is a novel agent that antagonizes orexin receptors. It does not bind to or modulate gamma-aminobutyric acid receptors bPer patient privacy and de-identification requirements, cells containing fewer than 11 patients are suppressed from reporting The negative control eye drop cohort included 611 individuals who experienced 630 VTE events over 33,607 observation days (median observation time = 51 days). Eye drop users were largely non-Hispanic White, female (vs male), and older adults (median age = 77.6 years) (Table S3 of the ESM). Drug Interaction Signals Before Accounting for Negative Control Cohort Findings (i.e., Based on Adjusted Incidence Rate Ratios [IRRs]) Among 1590 SCCS studies, 38 (2.4%) had an increased risk of VTE after adjusting for multiple estimation. Thirteen signals involved alprazolam, 12 involved clonazepam, four involved lorazepam, four involved zolpidem, three involved eszopiclone, and two involved diazepam. The adjusted IRRs varied between 1.30 for the lorazepam-oxycodone pair and 3.24 for the eszopiclone-codeine pair. Most signals included concomitant use of an anti-infective agent (N = 16, 42.1%), analgesic (N = 7, 18.4%), or anticoagulant (N = 3, 7.9%) (see Table S4 of the ESM). Drug Interaction Signals After Accounting for Negative Control Cohort Findings (i.e., Based on Ratios of IRRs) Among 1590 SCCS studies, five (0.3%) had an increased risk of VTE after adjusting for multiple estimation and findings from the negative control cohort, indicating a low signal yield after calibration. The remaining signals after this calibration step were: eszopiclone + gabapentin (ratio of IRRs = 3.57, 1.38–9.21), lorazepam + gabapentin (2.34, 1.20–4.57), diazepam + apixaban (2.16, 1.21–3.84), clonazepam + gabapentin (1.98, 1.03–3.83), and alprazolam + gabapentin (1.92, 1.00–3.66) (Table 2).Table 2Drug interaction pairs (N = 5) associated with an increased risk of VTE after negative control cohort and semi-Bayes shrinkage adjustmentsCohort-defining benzodiazepine/z-drugConcomitantly dispensed drugConcomitantly dispensed drug classRelative risk of VTEa, quantified via ratio of incidence rate ratiosb (95% CI)EszopicloneGabapentinAntiseizure agentc3.57 (1.38–9.21)LorazepamGabapentinAntiseizure agentc2.34 (1.20–4.57)DiazepamApixabanAnticoagulant2.16 (1.21–3.84)dClonazepamGabapentinAntiseizure agent†1.98 (1.03–3.83)AlprazolamGabapentinAntiseizure agent†1.92 (1.00–3.66)eCI confidence interval, VTE venous thromboembolismaIncidence rate ratio for cohort-defining drug + concomitantly dispensed drug vs the cohort-defining drug alonebCalculated by dividing the incidence rate ratio (see *) from the benzodiazepine/z-drug cohort by the corresponding incidence rate ratio from the negative control (i.e., eye drop) cohort. Effect estimates were intentionally attenuated via semi-Bayes shrinkage to minimize false-positive findingscWhile US Food and Drug Administration labeled indications for gabapentin are limited to focal onset seizures and postherpetic neuralgia, the drug is much more commonly used to treat neuropathic pain (other than postherpetic neuralgia) and other off-label indications (e.g., moderate-to-severe alcohol use disorder, fibromyalgia, vasomotor symptoms of menopause).dAs detailed in the discussion section, this association is likely explained by reverse causationep = 0.0484; reported given that the lower bound of the 95% CI, when rounded, was 1.00 Table 2Drug interaction pairs (N = 5) associated with an increased risk of VTE after negative control cohort and semi-Bayes shrinkage adjustmentsCohort-defining benzodiazepine/z-drugConcomitantly dispensed drugConcomitantly dispensed drug classRelative risk of VTEa, quantified via ratio of incidence rate ratiosb (95% CI)EszopicloneGabapentinAntiseizure agentc3.57 (1.38–9.21)LorazepamGabapentinAntiseizure agentc2.34 (1.20–4.57)DiazepamApixabanAnticoagulant2.16 (1.21–3.84)dClonazepamGabapentinAntiseizure agent†1.98 (1.03–3.83)AlprazolamGabapentinAntiseizure agent†1.92 (1.00–3.66)eCI confidence interval, VTE venous thromboembolismaIncidence rate ratio for cohort-defining drug + concomitantly dispensed drug vs the cohort-defining drug alonebCalculated by dividing the incidence rate ratio (see *) from the benzodiazepine/z-drug cohort by the corresponding incidence rate ratio from the negative control (i.e., eye drop) cohort. Effect estimates were intentionally attenuated via semi-Bayes shrinkage to minimize false-positive findingscWhile US Food and Drug Administration labeled indications for gabapentin are limited to focal onset seizures and postherpetic neuralgia, the drug is much more commonly used to treat neuropathic pain (other than postherpetic neuralgia) and other off-label indications (e.g., moderate-to-severe alcohol use disorder, fibromyalgia, vasomotor symptoms of menopause).dAs detailed in the discussion section, this association is likely explained by reverse causationep = 0.0484; reported given that the lower bound of the 95% CI, when rounded, was 1.00 Drug interaction pairs (N = 5) associated with an increased risk of VTE after negative control cohort and semi-Bayes shrinkage adjustments CI confidence interval, VTE venous thromboembolism aIncidence rate ratio for cohort-defining drug + concomitantly dispensed drug vs the cohort-defining drug alone bCalculated by dividing the incidence rate ratio (see *) from the benzodiazepine/z-drug cohort by the corresponding incidence rate ratio from the negative control (i.e., eye drop) cohort. Effect estimates were intentionally attenuated via semi-Bayes shrinkage to minimize false-positive findings cWhile US Food and Drug Administration labeled indications for gabapentin are limited to focal onset seizures and postherpetic neuralgia, the drug is much more commonly used to treat neuropathic pain (other than postherpetic neuralgia) and other off-label indications (e.g., moderate-to-severe alcohol use disorder, fibromyalgia, vasomotor symptoms of menopause). dAs detailed in the discussion section, this association is likely explained by reverse causation ep = 0.0484; reported given that the lower bound of the 95% CI, when rounded, was 1.00 Discussion In summary, our study identified five drug pairs associated with an increased rate of VTE after adjusting for negative control (eye drop) cohort findings; four involved concomitant use of gabapentin with a sedative hypnotic (eszopiclone, lorazepam, clonazepam, alprazolam). While most of the concomitant use drug pairs were not associated with increased rate of VTE during clinical practice in our data-driven screening, concomitant use of BZDs and gabapentin might be associated with excessive rate of hospital presentation for VTE. In this series of pharmacoepidemiologic studies, we initially identified 38 potential drug combinations associated with an increased risk of VTE. After adjustment for a negative control (eye drop cohort) analysis, most of the drug pair signals became nonsignificant, suggesting confounding by indication for the concomitant drug. This pattern highlights the importance of calibrating high-throughput screening results to distinguish true interaction signals from associations driven by the underlying clinical context or indication. For example, among those signals, many drug pairs included concomitant use of an anti-infective medication. Infection itself is an strong independent risk factor for VTE because it is often associated with systemic inflammation that compromises endothelial integrity, activates neutrophils, and downregulates natural anticoagulant pathways [45, 46]. Following negative control analysis with an eye drop comparator, five drug pairs (5/1590, 0.3%) had increased rates of VTE, four of which involved concomitant use of gabapentin. Gabapentin is an antiseizure agent widely used for neuropathic pain and other neurological conditions [47]. It is considered a GABA analog but has no direct GABAergic action [47]. Gabapentin is absorbed in the intestines through an active transport process and lacks the ability to interact with CYP enzymes, so no clinically relevant pharmacokinetic drug interactions have been reported [48]. Both BZDs and gabapentin have been found individually to increase the risk of VTE [49–51]; however, to our knowledge, this study is the first to identify a signal consistent with a potential drug interaction signal involving concomitant use of BZD and gabapentin on VTE risk. Gabapentin may have pharmacodynamic interactions, and co-administration with a BZD/z-drug may amplify central nervous system effects such as excessive sedation, dizziness, and reduced mobility through additive or otherwise amplified central nervous system effects [48]. Furthermore, gabapentin itself has been associated with increased risks of DVT and PE, which may increase the rate of VTE through additive adverse effects when co-administered with BZD/z-drugs [49, 51]. In addition, the concomitant use of both drugs may indicate the presence of complex psychiatric or neurological conditions (residual confounding) that were not adjusted for in the analysis because of the high-throughput screening nature of our study. For the remaining drug interaction pair, i.e., diazepam + apixaban, we are unaware of a direct pharmacodynamic [52] or pharmacokinetic mechanism [53] that could result in VTE. Diazepam is a BZD used primarily for anxiety and seizure disorders [54], while apixaban is a direct oral anticoagulant used to prevent and treat VTE [52]. The observed signal for this drug pair was likely driven by our method of outcome identification, which may have resulted in reverse causation. In this study, we limited VTE events to those resulting in an ED visit or hospitalization. Therefore, VTE diagnosed in an outpatient setting did not meet the outcome definition. However, some patients may have had an outpatient encounter for VTE and begun an oral anticoagulant prior to their hospital encounter. While the onset date for the VTE likely preceded the anticoagulant start, we defined the outcome date as the date of the hospital admission or ED visit for VTE. Consequently, upon hospital or emergency admission, they might have already been exposed to the anticoagulant, potentially increasing the observed VTE incidence during anticoagulant-exposed periods. This study has several advantages. To our knowledge, it is the first to present real-world evidence on a potential drug–drug interaction between BZDs and gabapentin affecting VTE risk. Second, our study used a 50% random sample of claims data from national insurance programs Medicare and Medicaid, representing publicly insured individuals across diverse race/ethnicity, sex, age, states, and vulnerable groups. The findings from this study could be generalizable to insured Americans, as CMS covers approximately one in four residents in the USA [55]. Third, this study implemented comprehensive statistical methods: our analysis used the SCCS study design, which controlled for time-invariant confounding factors within person by comparing the incidence of VTE during focal (exposed) and referent (unexposed) periods for each individual. Additionally, we adjusted for time-varying confounders measured at each observation day. This approach aimed to control for both fixed and dynamic confounders, increasing the internal validity of our findings. The results were further refined by applying semi-Bayes shrinkage to reduce false-positive signals resulting from multiple comparisons, as well as by benchmarking identified signals against a negative control cohort to address confounding by indication for the concomitant drug. This study has limitations. First, because of the observational design, there remains a potential for residual confounding. Reverse causation and random error may also impact the findings. The clinical conditions prompting initiation of concomitant medications may themselves be associated with short-term VTE risk and could therefore contribute to or exaggerate apparent drug interaction signals. Although the SCCS framework inherently controls time-invariant confounders, residual time-varying confounding related to underlying clinical status, disease severity, or treatment indication may remain. However, as this data-driven work was intended to identify possible drug interaction signals rather than to establish definitive causal relationships, the presence of residual confounding, reverse causality, or random error may be less critical at this stage. Second, because of the nature of claims data, exposure misclassification may be present because the drug exposures in our study were defined by drug dispensing rather than actual ingestion. Similarly, outcome misclassification could have occurred as the VTE outcome was defined by diagnosis code in claims using International Classification of Diseases, Tenth Revision, Clinical Modification, which may not be perfectly accurate. Although our work drew on several validation studies to generate the algorithm used to identify those conditions, our composite outcome definition has not yet been validated. Third, this study defined the VTE outcome based on only ED and inpatient VTE diagnoses, omitting VTE diagnoses in outpatient settings. Because we excluded all VTEs that were seen only in outpatient settings, our findings may be less generalizable. In addition, because of the focus on ED visits and hospital stays, we omitted the outpatient VTE encounters, where concomitant BZD/z-drug and anticoagulant use occurred before subsequent ED or hospital presentation. However, we expect this bias to be most pronounced for medications used to treat VTE, that is, anticoagulants, for which it may cause false-positive signals. We expect it to have minimal impact on other drug pairs. Fourth, because our screening-oriented analysis prioritized specificity of concurrent exposure to identify putative interaction signals, it may be less sensitive to delayed risk windows; however, as sedation-related effects are expected to begin during periods of active co-exposure, clinically meaningful signals would be expected to manifest preferentially during these intervals. Fifth, this screening focused on evaluating drug interaction signals involving only two medications; interactions among three or more drugs were not assessed in this study. Conclusions Benzodiazepines and z-drugs have been associated with VTE. Our study used a data-driven method to generate an evidence-based list of potential drug interactions involving BZDs or z-drugs that were associated with hospital presentation for VTE. Most notably, the majority of BZD/z-drug + concomitantly used drug combinations conferred no excess risk of hospital presentation for VTE. However, concurrent use of a BZD/z-drug with gabapentin may be an important exception, and increase the risk of VTE. Future research should be conducted to further investigate this finding. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (PDF 169 KB) Supplementary file1 (PDF 169 KB)