Effect of trazodone on cognitive decline in people with dementia: Cohort study using UK routinely collected data

Evidence in mouse models has found that the antidepressant trazodone may be protective against neurodegeneration. We therefore aimed to compare cognitive decline of people with dementia taking trazodone with those taking other antidepressants.


| INTRODUCTION
There are no successful disease modifying treatments for dementia, 1 intensifying the need for the identification of new treatment approaches. The repurposing of existing medication, which has proven effective in other diseases, has potential to bring medication more rapidly to general use; the antidepressant trazodone has been proposed to have potential for repurposing in dementia. 2 Increased protein unfolding is a pathological hallmark of many neurodegenerative diseases, 3 possibly mediated through increased activation of the pancreatic endoplasmic reticulum kinase PERK/ eIF2α-P pathway of the unfolded protein response and subsequent attenuation of protein synthesis. 4 A recent study found that trazodone has effect on eIF2α-P in mouse models of neurodegenerative disease 2 ; it restored protein synthesis, was neuroprotective, restored memory deficits, prevented neurodegeneration, and prolonged survival in these models of prion disease and frontotemporal dementia. There is potential benefit of this therapeutic approach for other forms of dementia as trazodone was linked to lower levels of phosphorylated tau, a feature of Alzheimer's disease pathology. 5 Trazodone is an antidepressant with multiple therapeutic mechanisms; it is a serotonin 5-HT 2A and α1-adrenergic antagonist, a serotonin reuptake inhibitor and a histamine H 1 inverse agonist. 6 It is licenced for major depressive disorder in the UK and US and prescribed for insomnia in depression. 7 Small studies have found it to reduce neuropsychiatric symptoms of Alzheimer's disease 8 and frontotemporal dementia 9 and because of this it is prescribed in clinical practice. 10 To explore the effect of trazodone on cognitive decline in people who develop dementia we performed naturalistic cohort studies in routinely collected secondary mental healthcare data from three large clinical services. As trazodone is prescribed for non-cognitive symptoms of dementia which may reflect more rapidly progressive dementia, 11 we compared trazodone with citalopram and mirtazapine; antidepressant drugs which do not have effect on eIF2α-P 2 but are prescribed for similar indications. While there has been no evidence that mirtazapine or citalopram improve cognitive trajectory, we hypothesised that trazodone may do so. As trazodone may be used as a treatment for depression or other neuropsychiatric symptoms, and these symptoms may be risk factors for or symptoms of worse dementia trajectory, 12 we also aimed to take into account the potential confounding effect of neuropsychiatric symptoms. Our specific objectives were to: � compare the cognitive trajectory of people with dementia who were prescribed trazodone with those prescribed citalopram or mirtazapine � examine these associations in people with non-vascular dementia subtypes because the PERK/eIF2α-P pathway is not pathogenic in vascular dementias

| Study setting and data source
We conducted three cohort studies using data from three separate datasets of routinely collected clinical data from large mental health trusts in North London (Camden and Islington NHS Foundation Trust),

South London (South London and Maudsley NHS Foundation Trust)
and Oxford (Oxford Health NHS Foundation Trust), UK. We used these services' case register Clinical Record Interactive Search (CRIS) data tools, which provide pseudonymised electronic medical records for research purposes. 13,14 These healthcare providers deliver a range of psychiatric services, including dementia assessment and management, to geographic catchment areas containing 1.2 million residents in south London; 480,000 residents in north London; and 1.9 million residents in Oxfordshire, Buckinghamshire, Swindon and Wiltshire, thereby covering around 5% of the UK population. CRIS allows data to be extracted from structured fields in patients' electronic clinical records and, to enhance recognition of relevant information, unstructured text (including correspondence, discharge summaries and clinical notes). Information from the unstructured text is extracted using General Architecture for Text Engineering (GATE), a language processing software 15 ; use of GATE in the CRIS resource is detailed in the description of the data resource. 13,14 CRIS has previously been used to examine a variety of dementia-related research questions 13,16 and data from more than one database has previously been combined to allow more diverse and generalizable samples. 17 GATE software has previously been shown to have precision (akin to positive predictive value) of between 94% and 96% for correct identification of psychotropic drugs in these data. 18

| Inclusion criteria
We retrieved records from the three databases for eligible study participants of any age who had: � received diagnosis of dementia during the study window (defined as ICD-10 20 code of F01-03 or G30 entered in the structured field of the electronic medical record).
� been recorded in electronic medical record as taking trazodone, citalopram or mirtazapine at any time before or after dementia diagnosis (derived from GATE 'medication' application).
� ≥2 recorded cognitive test scores, with at least one after the initiation of trazodone.
� ≥6 weeks recorded exposure to the drug of interest: drug initiation could precede dementia diagnosis if exposure continued after diagnosis, or it could follow dementia diagnosis.

| Exposure
We ascertained exposure status by extracting information from CRIS using GATE. This bespoke application is designed to identify medications that are currently prescribed to the patient and was developed through expert annotation, which means that domain experts coded whether the medication prescription was present in a particular document based on pre-defined coding rules and/or their expert experience. 21 This searched for any reference to trazodone (and common mis-spellings or molipaxin, the UK trade name) and comparator drugs, citalopram or mirtazapine, and extracted the date of the record. Any relevant record of the drugs of interest during the study window would be detected by the GATE algorithm so drug exposure was measured longitudinally. The index date was the first record of the drug starting or continuing and participants were considered as being exposed until the date of the last record of the drug, the date of death, or the end of the study window.
Individuals recorded taking trazodone and citalopram or mirtazapine during the study window, either in combination or succession, were categorised as in the trazodone exposure group, because of our hypothesis that trazodone may benefit cognitive trajectory while mirtazapine or citalopram have not been reported do so. Those who switched medication, for example took trazodone and later took mirtazapine would have been included in the trazodone group with exposure to trazodone defined as from the first to last relevant record of trazodone. Those recorded as taking both mirtazapine and citalopram during the study window were excluded from either control groups.

| Outcomes
We extracted all MMSE 22 dates and numerator and denominator scores from included patients before and after drug prescription at any time during study window, drawn from a structured field in the source record and a further GATE information extraction. For analysis, we used up to 10 MMSE records for each participant, the baseline record being the last MMSE record before drug initiation or, if no MMSE before initiation, the first MMSE following initiation. MMSE is a 30point scale assessing global cognitive function regularly recorded by clinicians, has good psychometric properties for determining disease severity, 23 and has been used as a cognitive outcome measure in trials. 24 MMSE has shown validity in differentiating people with and without dementia in previous studies using this datasource. 25

| Covariates
We derived data on: � From the South London research site, we extracted data on clinician-rated severity of agitation and depression symptom severity at the start of exposure using the Health of the Nation Outcome Scale (HoNOS). 27 HoNOS is a clinician-rated scale with SOMMERLAD ET AL.
-3 acceptable psychometric properties which is usually completed at regular clinical assessments; depression, which may take into account symptoms of loneliness as a symptom of depression, and agitation are rated on a scale from 0 (no symptoms) to 4 (severe symptoms).
We also derived data on date of death for deceased patients from the Office of National Statistics mortality database and the NHS national spine linked using NHS national patient record numbers and additional sociodemographic data. These data have been shown to have accuracy of 94%. 28

| Analytic approach
We first examined the baseline demographic and clinical character- We compared rate of MMSE change between drug groups using linear mixed models 29 as this approach uses all available outcome data and takes account of individuals' repeated MMSE measurements being correlated. For these analyses, both the intercept and slope were fitted as random effects as individuals had different cognitive scores at baseline and different rates of cognitive change over time. Because less than 10% of our study participants had more than 10 MMSE scores, we only used up to 10 MMSE records in our analysis, with the index MMSE being the final before drug initiation or, if none was available, the first MMSE after drug initiation, and up to nine subsequent MMSEs, until the last recorded documentation of the medication.
Our pre-specified main analysis was adjusted for age, sex and the length of time (as a continuous variable) between first and last MMSE scores. We repeated our main analysis excluding those who had been diagnosed with vascular dementia, as it is not considered to be neurodegenerative, 30 so we hypothesised it as less likely to be affected by trazodone treatment.
Due to data security and governance regulations, only aggregated data could be shared across sites. We conducted meta-analysis of effect estimates from the three sites using random effects models of the weighted mean difference. 31 Data were analysed using STATA (version 12) for mixed models analysis and Comprehensive Metaanalysis (version 3) for meta-analysis.

| Sensitivity analyses
In post-hoc sensitivity analyses, we examined whether confounding by indication affected our results, by (1) conducting our primary analysis with additional adjustment for neuropsychiatric symptom severity using data from the South London site-the largest of our research sites-with HoNOS agitation and depressed mood domains included as two separate continuous variables; and (2) repeating our analysis only including people with mild dementia (MMSE ≥20), adjusted for age, time and sex and, in a separate model, additionally for agitation and depressed mood.

| Role of the funding source
Funders had no role in the study design and the collection, analysis, and interpretation of data and the writing of the article and the decision to submit it for publication.

| RESULTS
The baseline characteristics according to site and drug groups are detailed in Table 1 and study flow is in Figure 1 We  (Table 2) and mean daily dose was 101.8 mg (range 50-300 mg); for one patient, trazodone was marked as to be given 'PRN' (when necessary.)

| Comparison of trazodone to mirtazapine or citalopram in three research sites
Mean unadjusted MMSE change (Table 3) Figure 3).

| Sensitivity analyses
We included 1,236 people with all-cause dementia in South London in AD had no effect on cognition or general function 33 (mean baseline MMSE 11.4). Other associated outcomes also support no diseasemodifying benefit of trazodone in established dementia; a Cochrane review of pharmacological interventions for agitation in dementia found no difference in global impression or trial withdrawal rates between trazodone and typical antipsychotics or placebo. 34 There is mixed literature on the cognitive effect of citalopram and mirtazapine in dementia, including early preclinical evidence of neuroprotective effects. 35,36 However citalopram and mirtazapine appeared to worsen cognition in the multicentre randomised, doubleblinded, placebo-controlled trials CitAD 37 and HTA-SADD 38 respectively and there is no consistent evidence of cognitive benefit for these medications. The worse cognitive performance for patients taking trazodone compared to citalopram in our study suggests it is unlikely that there is cognitive benefit from trazodone.
Patients in our study had moderately severe dementia at baseline (mean MMSE in the three sites 13.7 to 18.1), reflecting the use of trazodone in these patients for non-cognitive symptoms of dementia which are less common in mild dementia. 39 To consider whether benefit from trazodone would only be seen in earlier disease when medication is hypothesised to be more effective as neuropathological damage is less severe, 40 we analysed only those with mild dementia.
We found those with mild dementia taking trazodone declined faster than comparators, suggesting trazodone does not confer cognitive benefit early in the illness. Three other studies support our finding as they did not find cognitive benefit or reduced dementia incidence in people without dementia. Firstly, a cohort study using UK general practice data reported higher incidence of dementia in 4,596 people aged over 50 years without dementia at baseline who were prescribed trazodone compared to those taking other antidepressants. 41 Secondly, a cross-sectional observational study examining the cognitive effects of a range of medications in participants without dementia in the UK Biobank found no cognitive benefit associated with trazodone and instead a slowing of reaction time, 42 and thirdly a small study of non-demented people found three-fold increase in incident dementia in the 15 people taking trazodone compared to non-users. 43 The coefficients for MMSE change were summarised across the whole follow-up period and show mean decline in cognitive function of between 0.5 and 1 point per successive MMSE assessment, which took place at an interval of between 4 and 7 months. The people in each of the three drug groups showed an initial improvement from the first to second assessment and this is consistent with other naturalistic studies examining cognitive function in routine practice. 16 This improvement may reflect the combined effect of simultaneous initiation of pharmacological or psycho-social cognitionenhancing treatments, improvement of cognition as a result of the antidepressants alleviating neuropsychiatric symptoms, or practice effect in MMSE performance. 44 As expected, cognitive decline was subsequently seen in all drug groups.

| Strengths and limitations
This is the largest study to examine the effect of trazodone on cognition with any comparator and with the longest follow-up, SOMMERLAD ET AL. design, data collection and analysis, decision to publish, or preparation of the manuscript.

DATA AVAILABILITY STATEMENT
The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions. Due to the data management requirements of the data source, no additional data are available for sharing.