The association between benzodiazepine use and greater risk of neurocognitive impairment is moderated by medical burden in people with HIV

Participants included 435 PWH enrolled in various NIH-funded observational studies within the HIV Neurobehavioral Research Program (HNRP, https://​hnrp.​hivresearch.​ucsd.​edu/​) at the University of California San Diego (UCSD). Study details have been published elsewhere (Heaton et al. 1995, 2010; Morgello et al. 2001). Our deidentified data are available to qualified external investigators upon requests submitted to hnrpresource@ucsd.edu. Exclusion criteria for the parent studies included history of non-HIV-related neurological, medical or psychiatric disorders that affect brain function (e.g., diabetes), learning disabilities, dementia diagnosis, and English not a primary language. Our inclusion criteria were on ART with successful viral suppression (HIV-1 RNA < 50 copies/ml) and availability of variables of interest and covariates. Our exclusion criteria were positive urine toxicology for illicit drugs (except marijuana) or Breathalyzer test for alcohol during study visits. Medication use including benzodiazepines and estimated duration of use (missing in n = 15) were assessed via structured, clinician-administered questionnaires. Self-report of medication use was confirmed via review of medical regimen records if available. Eighty-one participants reported active use of prescribed benzodiazepines (benzodiazepine +), and 354 participants reported no use of prescribed benzodiazepines (benzodiazepine-). The following agents were reported by participants in the benzodiazepine + group: alprazolam (n = 19), clonazepam (n = 23), diazepam (n = 6), lorazepam (n = 29), temazepam (n = 8), estazolam (n = 1). Benzodiazepine use was further categorized as long-acting (N = 30) if reported clonazepam, diazepam, or estazolam use or short-acting (N = 51) if reported alprazolam, lorazepam, temazepam use. If both short-acting and long-acting medication types were reported, the participant was categorized as a long-acting benzodiazepine user. If a participant reported use of multiple benzodiazepines (regardless of type), the longest duration was used in analyses.

As a covariate, we calculated the total number of currently prescribed medications excluding ART and benzodiazepines. Due to evidence of an effect of opioid, antipsychotic and anticholinergic medication use on cognition (Coupland et al. 2019; Cuesta et al. 2001; Dufort and Samaan 2021), we also specifically assessed use of these medication types and considered as covariates.

UCSD’s Human Research Protections Program approved all study procedures, and all participants provided written informed consent. Participants completed comprehensive neuromedical and neurobehavioral assessments and blood draw during study visits.

HIV infection was diagnosed by enzyme-linked immunosorbent assay (ELISA) with confirmation by Western blot or real time-polymerase chain reaction (RT-PCR). HIV disease characteristics were determined via a combination of self-report (e.g., estimated duration of HIV disease) and laboratory tests (e.g., CD4 + T-cell count). Nadir CD4 + T-cell count was the lowest lifetime value by self-report, study-obtained CD4 + T-cell count, or released medical records. CD4 + T-cell count was measured with flow cytometry. Plasma HIV-1 RNA level was measured by ultra-sensitive RT-PCR (Amplicor, Roche Diagnostic System) in a CLIA-certified clinical laboratory. HIV-1 RNA was considered undetectable below the lower limit of quantitation of 50 copies/mL. Hepatitis C serostatus was measured via MedMira Multiplo rapid test (MedMira Inc.).

Medical burden was quantified based on previously established methods for constructing a frailty index in PWH (Guaraldi et al. 2015; Oppenheim et al. 2018). Here, the medical burden index was calculated as the proportion of health deficits from a group of 28 health variables, including clinical laboratory assays, medical comorbidities, and HIV-specific characteristics. Each variable was coded as “1,” when a deficit was present and “0,” when absent (see Table 1, Supplemental Digital Content, which displays the medical burden index components and their criteria). Thus, medical burden scores ranged from 0 (no deficits) to 1 (all deficits). The 28 selected variables were chosen based on all available neuromedical data from the parent studies that were consistent with variables included in prior indices (Guaraldi et al. 2015; Oppenheim et al. 2018). Consistent with published guidelines for creating a frailty index (Searle et al. 2008), we excluded factors that (1) had greater than 5% missing data (i.e., unintentional weight loss, LDL cholesterol) and (2) had less than 1% of participants meeting criteria for the deficit (i.e., abnormal sodium and peripheral vascular disease).

Participants completed a standardized neurocognitive test battery of verbal fluency, working memory, speed of information processing, learning and delayed recall, executive function, and complex motor function. Specific tests are described elsewhere (Cysique et al. 2011). Raw test scores were transformed into demographically adjusted (i.e., age, education, sex, and race/ethnicity) T-scores based on normative samples of HIV participants (Heaton et al. 2004; Norman et al. 2011). T-scores for each test were also converted into deficit scores that ranged from 0 (T-score ≥ 40, no impairment) to 5 (T-score < 20, severe impairment) and then averaged across all tests to obtain a global deficit score (GDS) and within domains to obtain domain deficit scores (DDS) (Heaton et al. 1995; Carey et al. 2004; Blackstone et al. 2012). Based on pre-established cut-points, global NCI was defined as GDS ≥ 0.50 and domain-specific NCI was defined as DDS > 0.50 (Carey et al. 2004).

DSM-IV diagnoses of current and lifetime alcohol, cannabis, and illicit (amphetamine, cocaine, hallucinogens, inhalant, sedatives, opioids, and PCP) substance use disorders, major depressive disorder (MDD), and generalized anxiety disorder (GAD) were determined based on the fully structured computer-based Composite International Diagnostic Interview version 2.1 (World Health Organization 1997). The Beck Depression Inventory-version 2 (BDI-II) (Beck et al. 1996) was used to assess depressive symptoms within the past two weeks with higher scores reflecting greater affective symptoms. The 9-item tension/anxiety subscale of the Profile of Mood States (POMS; n = 367) (McNair et al. 1981) was administered to a subset of participants as a measure of anxious and nervous feeling in the past week. Given that BDI-II data were available on more participants and strongly correlated with the POMS tension/anxiety subscale (R = 0.71, p < 0.001), BDI-II was chosen to model affective distress in the primary statistical models; however, scores on the POMS tension/anxiety subscale were adjusted for in sensitivity analyses to see if results changed.

Differences in sample characteristics between benzodiazepine groups were examined using Chi-square tests for categorical variables, t-tests for normally distributed continuous variables, and Kruskal–Wallis one-way analyses of variance (ANOVA) for non-normally distributed continuous variables as determined by the Shapiro–Wilk test. In order to demonstrate the utility of separately examining the moderating roles of chronological age and medical burden, we examined the relationship between chronological age and medical burden index. Then, using multivariable logistic regression, we examined the benzodiazepine use X age and benzodiazepine use X medical burden interaction terms on odds of NCI. Significant interactions were probed by a Johnson–Neyman analysis (Johnson and Neyman 1936; Preacher et al. 2006) to empirically identify the specific range of values of the moderator (i.e., age, medical burden, or both) for which the difference in odds of NCI between benzodiazepine groups was statistically significant. Interaction terms that were not significant were removed from the model in order to examine the main effects. Demographics and relevant clinical factors that were not included in the medical burden index were considered as covariates, i.e., BDI-II score, history of MDD, history of GAD, current use of antidepressant, opiate, antipsychotic and anticholinergic medications, number of current medications (excluding benzodiazepines and ART), and current and past substance use disorders including alcohol, cannabis, cocaine, hallucinogens, inhalants, methamphetamine, opioids, PCP, sedatives, and others. Covariates that related to NCI at p ≤ 0.10 in univariate analyses were included in statistical models and were retained if significant in the multivariable model at p ≤ 0.10.

Upon detection of a significant relationship between benzodiazepine use and odds of NCI, either independently or through interactions with medical burden and/or age, we further probed the relationship by examining the following: (a) whether it was driven by long-acting versus short-acting benzodiazepines, (b) whether it was driven by specific cognitive domains, and (c) the relationship between duration of benzodiazepine use and odds of NCI among benzodiazepine users in order to test for a potential dose-dependent relationship.

The sample of 435 PWH was 87% male and 50% non-Hispanic White, 27% Hispanic with a mean age of 52.6 (SD = 12.0; age range: 18–87) and mean years of education of 14.0 (SD = 2.7). Among benzodiazepine users, estimated duration of use ranged from 2–212 months with a median of 47 months. Sample characteristics by benzodiazepine status are displayed in Table 1. Compared to benzodiazepine non-users, benzodiazepine users were older and had a higher proportion of non-Hispanic Whites. As expected, benzodiazepine users had more depressive and tension/anxiety symptoms, a higher prevalence of current and past MDD diagnoses, a higher prevalence of opioid and antipsychotic medication use and higher medical burden index scores compared to benzodiazepine non-users. In contrast, benzodiazepine groups were comparable in rates of current and past substance use disorders and in HIV disease characteristics with the exception of a longer duration of HIV disease in benzodiazepine users versus non-users. Consistent with our previous results in a sample of PWH with and without viral suppression (Saloner et al. 2019b), we found higher rates of NCI in benzodiazepine users (49.3%) versus non-users (36.1%) (X² = 4.86, p = 0.03) among ART-treated, virally suppressed PWH. Among considered covariates, BDI-II scores, lifetime MDD diagnosis, number of medications, antidepressant and antipsychotic medication use and current alcohol use disorder were associated with NCI at p ≤ 0.10 and, thus, were included in all statistical models. When examining the relationship between chronological age and medical burden index, we found a significant quadratic relationship (p = 0.004) whereby age positively related to medical burden index until age 63, as determined by the Johnson–Neyman analysis, at which point this relationship is nonsignificant (Fig. 1).

Among included covariates (BDI-II scores, history of MDD, history of GAD, current alcohol use disorder, number of medications), BDI-II scores and antipsychotic medications were retained in final statistical models due to its significance in the multivariable models (p ≤ 0.10). The benzodiazepine use X age interaction was not significant indicating no moderating role of age in the adverse effect of benzodiazepine use. There was also no main effect of age on NCI. In contrast, we detected significant benzodiazepine use X medical burden interactive effects on NCI (Table 2). In probing this interaction, the Johnson–Neyman analysis revealed that the adverse effect of benzodiazepine use on the odds of NCI is only observed when the medical burden index is greater than 0.3 reflecting presence of at-least 30% of the considered health deficits (Fig. 2). Among those with medical burden ≥ 0.3, benzodiazepine use related to an almost threefold increased likelihood of NCI (OR = 2.76, 95%CI = 1.23–6.19, p = 0.01; Fig. 2). However, benzodiazepine use did not relate to NCI among those with medical burden ≤ 0.3 (OR = 0.90, 95%CI = 0.45–1.80, p = 0.76). When adjusting for POMS tension/anxiety scores in sensitivity analysis in a participant subset (N = 367), results did not change except that the association between benzodiazepine use and higher odds of NCI was slightly attenuated in the higher medical burden group (OR = 2.51, 95%CI = 1.08–5.83, p = 0.03).

We next examined whether our findings were driven by long-acting versus short-acting benzodiazepines. Although the interaction of medical burden with either long-acting (OR:2.64, 95%CI:0.83–8.39, p = 0.10) or short-acting (OR:1.66, 95%CI:0.86–3.22, p = 0.13) benzodiazepines was not significant, both use of both types of benzodiazepines (compared to no benzodiazepine use) related to a higher odds of NCI only within those with a medical burden index ≥ 0.3 (long-acting: OR:2.83, 95%CI:0.84–9.48, p = 0.09; short-acting: 2.86, 95%CI:1.15–7.15, p = 0.02). Although this relationship was the same magnitude for short-acting and long-acting benzodiazepines, it was only significant for the short-acting type likely due to the larger sample and, thus, greater statistical power. When examining duration of benzodiazepine use among benzodiazepine users, we found no relationship between the benzodiazepine X medical burden interaction and odds of NCI (OR = 1.10, 95%CI = 0.95–1.26, p = 0.19). There was also no relationship between duration of benzodiazepine use and odds of NCI among those with high medical burden (index ≤ 0.3) (OR = 1.00, 95%CI = 0.92–1.09, p = 0.97); however, although benzodiazepine duration ranged from 2 to 212 months, variability was limited in that the majority of users (70%) reported a duration within about 3–5 years.

Significant medical burden X benzodiazepine interactions were observed for speed of information processing, verbal fluency and motor impairment (Table 3). The Johnson–Neyman analysis revealed that the adverse effect of benzodiazepine use on the odds of NCI is only observed when the medical burden index is greater than or equal to 0.32 for speed of information processing, 0.34 for motor and 0.27 for verbal fluency. These interactions revealed a significantly higher likelihood of speed of information processing (OR = 3.79, 95%CI = 1.06–13.53, p = 0.04), motor (OR = 3.34, 95%CI = 1.24–8.99, p = 0.02) and verbal fluency (OR = 2.58, 95%CI = 1.28–5.22, p = 0.008) impairment in benzodiazepine users versus non-users only among those above the respective Johnson–Neyman-derived cut-point of the medical burden index. Benzodiazepine use independently related to a higher likelihood of learning and recall impairment and working memory/attention impairment. When adjusting for POMS tension/anxiety scores in sensitivity analysis in a participant subset (N = 367), the relationships between benzodiazepine use and higher odds of processing speed (OR:2.52, 95%CI:0.63–10.1, p = 0.19), motor (OR:3.09, 95%CI:1.10–8.65, p = 0.03), and verbal fluency (OR:2.38, 95%CI:1.12–5.03, p = 0.02) impairment were attenuated although remained significant except for processing speed. Results for all cognitive domains did not change substantively when adjusting for POMS tension/anxiety scores.