Safety of pharmacologic interventions for neuropsychiatric symptoms in dementia: a systematic review and network meta-analysis

Background Prescribing trends suggest that pharmacologic alternatives to antipsychotics are gaining in popularity, but randomized trial (RCT) data of their comparative safety is scarce. Our objective was to describe the comparative safety of pharmacologic interventions for treating neuropsychiatric symptoms in dementia. Methods We searched MEDLINE, EMBASE, CENTRAL, CINAHL, and PsycINFO, from inception to May 28, 2019, for studies of pharmacologic interventions used to treat neuropsychiatric symptoms in dementia. Dementia care partners selected fracture risk as our primary outcome. Pairs of reviewers, working independently, conducted all study screening, data abstraction, and risk of bias appraisal. We conducted Bayesian random-effects network meta-analyses (NMAs) using data from RCTs to derive odds ratios (ORs). In secondary analyses, we conducted frequentist random-effects NMAs using data from RCTs and Bayesian three-level hierarchical random-effects NMAs incorporating data from RCTs and non-randomized studies. Results Our systematic review included 209 randomized and non-randomized studies (889,378 persons with dementia). In NMAs of data from randomized trials, there were no increased odds of fracture associated with any intervention in primary analyses; however, data were sparse. We found increased odds of cerebrovascular events associated with antipsychotics (odds ratio [OR] 2.12, 95% credible interval [CrI] 1.29 to 3.62; number needed to harm [NNH] = 99) and increased odds of falls associated with dextromethorphan-quinidine (OR 4.16, 95% CrI 1.47 to 14.22; NNH = 55) compared to placebo in persons with dementia. In a subgroup of persons with Alzheimer disease, antipsychotics were associated with increased odds of fracture compared to anticonvulsants (OR 54.1, 95% CrI 1.15 to 38,300; NNH = 18). In older persons (mean age ≥ 80 years) with dementia, anticonvulsants were associated with increased odds of death compared to placebo (OR 8.36, 95% CrI 1.17 to 203.4; NNH = 35) and antipsychotics were associated with increased odds of death compared to antidepressants (OR 5.28, 95% CrI 1.06 to 3.51; NNH = 47). Conclusion Although antipsychotics were associated with greater harm than antidepressants and anticonvulsants in subgroups of persons with dementia, medications used in lieu of antipsychotics for treating neuropsychiatric symptoms in dementia, such as anticonvulsants and dextromethorphan-quinidine, were also associated with harm. Decision-making concerning treatments prescribed in lieu of antipsychotics should include potential harms. PROSPERO registration CRD42017050130.


Background
There are 50 million people worldwide living with dementia [1]. Pharmacologic (e.g. antidepressants and antipsychotics) and nonpharmacologic (e.g. recreation therapy) interventions are administered as treatment for cognitive and neuropsychiatric symptoms associated with dementia [2][3][4]. In persons with dementia, pharmacologic interventions, most notably antipsychotics, have been associated with significant risk including cerebrovascular accidents and death [5,6]. Studies show that as clinicians adopt recommendations of clinical practice guidelines and de-prescribing initiatives to taper and discontinue antipsychotics in persons living with dementia, alternative pharmacologic interventions (e.g. antidepressants, anticonvulsants) are prescribed to treat neuropsychiatric symptoms rather than nonpharmacologic alternatives, despite their demonstrated efficacy [7][8][9][10][11]. Pharmacologic interventions continue to be prescribed by clinicians for many reasons, including perceived patient, staff, and system barriers to implementation of nonpharmacologic interventions in busy clinical settings [12].
Given the high prescription rates of pharmacologic interventions for cognitive and neuropsychiatric symptoms of dementia and the paucity of head-to-head randomized trials (RCTs) that knowledge users (e.g. patients, caregivers, and clinicians) can use to inform shared decision-making, we conducted a network meta-analysis (NMA) to fill this critical knowledge gap [13]. Our objectives were to: (1) determine the comparative risks of: fracture, mortality, cerebrovascular events, and falls associated with pharmacologic interventions dispensed for treating neuropsychiatric symptoms in dementia; and (2) establish the safest pharmacologic interventions to treat neuropsychiatric symptoms in dementia.

Methods
Our protocol was registered (Prospero: CRD42017050130) and published [14]. We report our findings in accordance with the PRISMA extension statement for reporting systematic reviews incorporating NMA [8]. We discuss protocol deviations in Additional file 1 [14].

Eligibility criteria
We included RCTs and non-randomized studies (NRSs) of pharmacologic interventions used to treat neuropsychiatric symptoms in persons with any type of dementia. Pharmacologic interventions were limited to those with final approval from the US Food and Drug Administration or Health Canada, as of our literature search date. Eligible comparator groups included placebo, usual care, or another approved pharmacologic intervention.

Outcomes
We asked 12 dementia care partners (e.g. nurses, physicians, a caregiver) to rank commonly reported safety outcomes associated with the treatment of neuropsychiatric symptoms in dementia [14,15]. Risk of fracture was selected as our primary outcome. The next most highly ranked outcomes were risk of fall, cerebrovascular event, and death.
Data sources, study selection, data abstraction, and risk of bias appraisal An experienced librarian searched MEDLINE, EMBASE, CENTRAL, CINAHL, and PsycINFO for citations published in any language from inception until May 28, 2019 [14]. Our search strategy was peer reviewed by a second librarian using the Peer Review of Electronic Search Strategies checklist [16]. We searched the grey literature, reference lists of included studies, and related systematic reviews (Additional file 1). After pilot-testing, pairs of reviewers (JAW, ZG, VN, PAK, MG, and YT) independently screened all citations and full-text articles to establish eligibility for inclusion, abstracted data from included full-text articles, and appraised each study for risk of bias. Where available, we abstracted study-level estimates (e.g. hazard ratios [HR]) from NRSs that adjusted for confounders. RCTs were appraised with the Cochrane Risk of Bias Tool [17]. Cohort and casecontrol studies were appraised with the Newcastle-Ottawa quality assessment scales [18]. Other NRSs were appraised with the Cochrane Effective Practice and Organisation of Care Tool [19].

Data synthesis and analysis
We formed treatment nodes based on drug classes, irrespective of treatment dose: placebo/control, antidepressants, anticonvulsants, N-methyl-D-aspartate (NMDA) receptor antagonist (memantine), cholinesterase inhibitors, antipsychotics, anxiolytic/hypnotics, anddextromethorphanquinidine. We assessed network transitivity by visual inspection of study and patient characteristics across treatment comparisons: study size, patient age, proportion of females, study setting (e.g. nursing home or community), study duration, type of dementia, severity of dementia, and two items from the risk of bias assessment (missing data and method of randomization). We prepared network diagrams in Stata, version15.1 (StataCorp) [20].
Our primary analyses for each outcome were Bayesian random-effects NMAs, which were conducted in Open-BUGS, version 3.2.3 [21]. We conducted a Bayesian random-effects pairwise meta-analysis where there was more than one study per treatment comparison. We used random-effects models because we anticipated between-study heterogeneity. We assumed a single within-network between-study variance in each NMA model because all treatments were used in a population of persons living with dementia. Informative prior distributions were implemented for all between-study heterogeneity parameters for the outcome of mortality (τ 2l og-normal (− 4.06,1.45 2 ); and for the outcomes of fracture, fall, and cerebrovascular event (τ 2~l og-normal (− 3.02,1.85 2 )) [22]. We implemented vague prior distributions for all trial baselines and treatment differences (normal (0,1000)). Outcomes were reported as odds ratios (ORs) with 95% credible intervals (CrIs) and predictive intervals (PrIs). We ranked treatments using surface under the cumulative ranking curve (SUCRA) values and summarized treatment rankings across outcomes in a rank-heat plot [23]. The number needed to harm (NNH) and number needed to treat (NNT) were calculated as the inverse of the difference in absolute risk between two intervention groups. We assessed global consistency with a design-by-treatment interaction model and local consistency with a loop-specific approach [24,25].
In secondary analyses, we conducted frequentist random-effects network and pairwise meta-analyses for each outcome in Stata, version 15.1 (StataCorp) [20]. Frequentist approaches are most commonly reported in the medical literature and we wanted to facilitate interpretation of our findings [26]. We did not conduct frequentist analyses for subgroup or sensitivity analyses. We estimated the common within-network betweenstudy variance with the restricted maximum likelihood method. Outcomes were reported as ORs and 95% confidence intervals (CIs).
We conducted subgroup analyses based on the following effect modifiers: average age of study population ≥ vs. ≤80 years old, proportion of women ≥ vs. < 50%, Alzheimer type of dementia, severity of dementia (mild/moderate vs. moderate/severe), study setting (long-term care facility vs. community), and study duration (< 13 weeks vs. ≥13 weeks). We completed a meta-regression based on publication year. We conducted sensitivity analyses by removing studies where data were found in a secondary data source (e.g. systematic review or clinical trial registry). We also conducted sensitivity analyses by removing studies at high or unclear risk of bias based on the two components of the risk of bias assessment that were the greatest threat to the validity of study findings: incomplete outcome data and randomization procedure. We conducted a sensitivity analysis using a weakly informative prior for the between-study heterogeneity (τÑ (0,1), τ > 0) in primary analyses. We assessed for publication bias by visually inspecting comparison-adjusted funnel plots [20].

Results
We screened 19,684 article titles and abstracts and 3369 full-text articles, which resulted in 209 included studies (889,378 patients with dementia) (Fig. 1). Individual study and study-level patient characteristics are reported in Additional file 1 Tables 1 and 2. Arm-level data for RCTs included in our NMAs are reported in Additional file 1 Table 3. There was transitivity across treatment comparisons for each outcome (Additional  file 1 Tables 4a-d). In most studies, the average age of patients was greater than 70 years and women represented greater than 50% of patients (Table 1). Only 42.6% of studies reported a measure of neuropsychiatric symptom severity. Data from 49 of 150 (32.7%) RCTs included in our systematic review were retrieved from secondary sources (Additional file 1 Table 3). Bias from missing data and unclear randomization procedures posed the greatest threats to study validity (Additional file 1 Tables 5-8 and Additional file 1 Figure 1). There was no evidence of publication bias based on visual assessment of comparison-adjusted funnel plots (Additional file 1 Figures 2a-h). There was no evidence of global inconsistency according to the design-by-treatment interaction model in any of our four NMAs (Additional file  1 Tables 9b, 9e, 9 h, 9 k). In meta-regression analyses, year of study publication was not associated with any of our outcomes.

Fracture
For our primary outcome of fracture risk, 46 studies were included in our systematic review and 35 studies included in our NMAs (29 RCTs [13,410 persons with dementia] plus 6 NRSs [107,765 persons with dementia]). Fracture data from 13 RCTs were retrieved from a secondary source. In our primary analysis, the network diagram revealed a connected network with no closed loops, and between-study heterogeneity was low ( Fig. 2 and Additional file 1 Figure 3a). Inclusion of NRSs resulted in one consistent, closed network loop (Additional file 1 Figure 3b).
No treatment was associated with increased odds of fractures compared with placebo in primary or sensitivity analyses ( Table 2, Table 3, and Additional file 1 Tables 9a-c). In persons with moderate-severe dementia (OR 0.05, 95% CrI < 0.01 to 0.71; NNH = 14), with mildmoderate dementia (OR 0.01, 95% CrI < 0.01 to 0.74; NNT = 17), or enrolled in studies longer than 12 weeks (OR 0.01, 95% CrI < 0.01 to 0.8; NNT = 15), anticonvulsants  Cholinesterase inhibitors were associated with lower odds of mortality compared to placebo (Bayesian OR   (Table 2 and  Additional file 1 Tables 9 g-h). This finding among patients randomized to receive antipsychotics compared to placebo was insensitive to our choice of prior for between-study variance (OR 2.2, 95% CrI 1.32 to 3.83).
Significantly higher odds of cerebrovascular events associated with antipsychotics compared to placebo/control were not seen in NRSs (Table 3 and Additional file  1 Table 9i).

Discussion
Our systematic review and NMA described the comparative risk of fractures, mortality, cerebrovascular events, and falls associated with pharmacologic interventions implemented for treating neuropsychiatric symptoms in dementia. There were no pharmacologic interventions associated with increased odds of fracture; however, few studies reported fractures as an adverse event and there was large uncertainty for some effect estimates. Importantly, in older adults with dementia, we identify increased odds of mortality associated with anticonvulsants compared to placebo and antipsychotics compared to antidepressants. And, dextromethorphan-quinidine was associated with increased odds of falling compared to placebo, antidepressants, and antipsychotics. These estimates, based on RCT data, suggest that specific pharmacologic interventions dispensed for treating neuropsychiatric symptoms in dementia are associated with harm.
Owing to a paucity of head-to-head RCTs comparing the harms associated with pharmacologic interventions in persons with dementia, we used NMA to fill a critical knowledge gap. For example, we used indirect evidence to demonstrate that antipsychotics are associated with increased odds of death compared to antidepressants in older adults with dementia. Knowledge users (e.g. persons with dementia, caregivers, and clinicians) can personalize decision-making by using our rank-heat plot to visualize treatment risk rankings across each of the four outcomes of harm (Fig. 3).
Similar to two published NMAs of pharmacologic interventions in persons with dementia, we found that antipsychotics were associated with an increased risk of cerebrovascular events [30,31]. Although one of these NMAs included cholinesterase inhibitors, authors did not conduct analyses to assess for an association with mortality risk [30]. We found a possible association between cholinesterase inhibitor use and lower mortality risk among RCTs only; however, wide predictive intervals (that cross one) associated with this effect estimate suggest that, in a future study, we might not see this association. The strength of this finding may be limited by: the greater uncertainty in our Bayesian compared to frequentist effect estimate, the dependence of our effect estimate on unpublished data, and a number of studies at high risk of bias from missing data [32]. There were limitations to our study. First, although neuropsychiatric symptoms occur in 75% of patients with dementia each month, only 42.6% of included studies reported a measure of neuropsychiatric symptom severity (e.g. Neuropsychiatric Inventory) [33]. Therefore, our systematic review included all studies of persons with dementia who received pharmacologic interventions commonly used for neuropsychiatric symptoms. Second, as a significant amount of adverse event data were not reported in primary RCTs, we abstracted these data from clinical trial registries, systematic reviews, meta-analyses, and secondary analyses of RCTs. Third, we were not able to conduct missing data analyses to test the missing at random assumption because these data were not always available (see Additional file 1 for protocol deviations). Lastly, we did not derive effect estimates for specific treatments (e.g. risperidone, olanzapine) or doses because outcome data were too sparsely reported across treatment networks.

Conclusion
In conclusion, we demonstrated potential harms associated with pharmacologic interventions intended to treat neuropsychiatric symptoms in persons with dementia. Our findings suggest that changing prescribing trends towards pharmacologic interventions other than antipsychotics may not be appropriate given the risk for harm. Rather, we should advocate for increased use of nonpharmacologic interventions given their demonstrated efficacy [10].