Anticholinergic burden in middle and older age is associated with lower cognitive function, but not with brain atrophy

Aims The aim of this study is to estimate the association between anticholinergic burden, general cognitive ability and various measures of brain structural MRI in relatively healthy middle‐aged and older individuals. Methods In the UK Biobank participants with linked health‐care records (n = 163,043, aged 40–71 at baseline), of whom about 17 000 had MRI data available, we calculated the total anticholinergic drug burden according to 15 different anticholinergic scales and due to different classes of drugs. We then used linear regression to explore the associations between anticholinergic burden and various measures of cognition and structural MRI, including general cognitive ability, 9 separate cognitive domains, brain atrophy, volumes of 68 cortical and 14 subcortical areas and fractional anisotropy and median diffusivity of 25 white‐matter tracts. Results Anticholinergic burden was modestly associated with poorer cognition across most anticholinergic scales and cognitive tests (7/9 FDR‐adjusted significant associations, standardised betas (β) range: −0.039, −0.003). When using the anticholinergic scale exhibiting the strongest association with cognitive functions, anticholinergic burden due to only some classes of drugs exhibited negative associations with cognitive function, with β‐lactam antibiotics (β = −0.035, P FDR < 0.001) and opioids (β = −0.026, P FDR < 0.001) exhibiting the strongest effects. Anticholinergic burden was not associated with any measure of brain macrostructure or microstructure (P FDR > 0.08). Conclusions Anticholinergic burden is weakly associated with poorer cognition, but there is little evidence for associations with brain structure. Future studies might focus more broadly on polypharmacy or more narrowly on distinct drug classes, instead of using purported anticholinergic action to study the effects of drugs on cognitive ability.


| INTRODUCTION
Anticholinergic drugs (anticholinergics) are medicines thought to block muscarinic receptors. Their anticholinergic action is ascertained by consulting anticholinergic scales that assign potency scores to individual drugs; the combined score for an individual patient is the anticholinergic burden (AChB). Anticholinergics are commonly prescribed for a variety of conditions, 1 and their transient side effects on cognition are well-known. [2][3][4][5][6] Moreover, their long-term use in old age 7 and middle age [8][9][10][11] has been associated with an increased risk of cognitive decline and dementia. It has been hypothesized 12 that this relationship is due to central anticholinergic effects, affecting areas of the brain crucial for cognition. [13][14][15] Therefore, a relationship might exist between AChB, cognitive ability and brain structure, even within the normal spectrum of cognitive functioning.
However, the existing evidence on this relationship is mixed. Most studies on anticholinergic prescribing in adults classify cognition as the absence vs. presence of a disorder or test separate cognitive modalities in isolation. 16,17 When measured this way, studies of AChB and cognitive ability often produce discordant results. 16 There are reports of positive associations between anticholinergic use and executive function, 12,18-21 associative learning, 22 visual, 23 episodic, 24,25 and shortterm memory, 26 delayed and immediate recall, 27 language abilities, 28 visuospatial skills, 28 attention, 28 and reaction time [Correction added on 1 May 2023, after first online publication: The preceding sentence has been updated in this version.]. 28 However, some authors have found no evidence for delayed and immediate recall, 21,22,28,29 reaction time, 22 executive function, 23,27 language abilities, 27,29 working memory, 25,27 processing speed, 25 and implicit 28 and semantic 25 memory. Additionally, because anticholinergic scales sometimes include different drugs and score the same drugs differently, 30 they could represent another source of variation in reported findings.
It has been suggested that global composites of cognitive functioning might be more sensitive to subtle cognitive changes. 16 Individual test scores contain more random noise, and the results can limit generalisability and contribute to inconsistency among studies. By contrast, general cognitive ability (sometimes referred to as general intelligence or g) represents shared variation across cognitive domains, is predictive of various social outcomes, 31 health outcomes, 32,33 mortality, 34 and is referenced in widely utilized diagnostic manuals. 31 Analysing large samples on multiple anticholinergic scales can further strengthen the reliability of the results.
Past studies have demonstrated associations between cognitive ability and several measures of brain structural magnetic resonance imaging (MRI). While the effect sizes have varied depending on the sample characteristics and cognitive tests used, they have usually ranged from r = 0.2 to 0.3. 35 One review found evidence for cross-sectional and longitudinal associations between global cognition and total brain size, global grey matter and hippocampal volume. 36 An analysis conducted on the UK Biobank sample found correlations between general cognitive ability and total brain volume, functional anisotropy (FA), mean diffusivity (MD) and several regional cortical volumes, especially those in the frontal lobe. Additionally, the authors found associations for multiple subcortical structures, especially the thalamus. 37 However, little is known about the neural correlates of potential anticholinergic-related cognitive decline.
To our knowledge, 4 studies 8,12,38,39 to date have assessed the relationship between these brain measures and regular anticholinergic use. While each study reported on associations between anticholinergic use and various metrics of brain structure and function, replication studies in larger samples are required. Furthermore, research is needed to probe potential differences between anticholinergic scales and between different classes of anticholinergics when exploring associations with cognitive function and cerebral correlates.
What is already known about this subject • Long-term anticholinergic use is associated with a risk of dementia, but the evidence on the relationship with cognitive ability in healthy individuals is mixed. It is unclear if anticholinergic use is associated with measurable changes in brain structure before the onset of advanced age and dementia.
• The heterogeneity in previous studies may be due to differences in cognitive tests and anticholinergic scales used to measure the outcome and exposures, respectively, and in different effects of distinct classes of drugs.

What this study adds
• Our study suggests that while anticholinergic use according to most anticholinergic scales studied is associated with lower cognitive ability, the relationship holds only for some classes of drugs, especially β-lactam antibiotics and opioids.
• In contrast to previous studies linking anticholinergic use to changes in brain structure in individuals with dementia, we found no such relationship in healthy individuals.
In our study-conducted using the UK Biobank-we calculated a latent factor of general cognitive ability ( g) and utilized MRI-imaging measures and prescriptions linked from primary care, to study the association between AChB, g and various brain structural MRI measures. Our goals were to assess: (i) whether there existed differences between anticholinergic scales and (ii) between drug classes in the association of AChB and cognitive ability; and (I ii) whether potential associations between AChB and cognitive ability were reflected in brain MRI measures, including brain atrophy, the volume of various cortical and subcortical brain structures, and measures of white matter microstructure. Based on previous findings, we hypothesised AChB to negatively associate with g, total brain volume, and the volumes of prefrontal cortical areas, the thalamus and hippocampus. Prescriptions are complete until May 2016 and were sourced from primary care. The prescription entries contained names and dates of drugs prescribed by general practitioners and the (mostly region-specific) suppliers of the prescription data. For the variables described below, we provide specific Field IDs (and links to the descriptions page for each field) in Table S1.  Table S2). Analyses of their psychometric properties in this sample have been reported previously. 41, 42 We fitted a confirmatory factor analysis in a structural equation modelling (SEM) framework to calculate g from the cognitive tests ( Figure S1 and Table S3), yielding 2 separate values, 1 for each assessment visit. SEM has been used to calculate g in UK Biobank before 37,43 ; the proportional variance explained in our study is smaller (23% for the baseline assessment, 28% for the imaging assessment) than in prior work in UK Biobank that used fewer cognitive tests. 37 For participants for whom this was possible, g from the imaging assessment was used in our analyses.

| Brain imaging
Since 2014, UK Biobank has been enhancing the dataset with imaging data that includes brain MRI. 40,44 It consists of imaging-derived phenotypes, whose acquisition and quality control have been previously described. 45 Briefly, brain imaging data were obtained at 4 data collection sites (Cheadle, Newcastle, Reading and Bristol; all UK) using 3 identical scanners (3T Siemens Skyra), with a standard Siemens 32-channel receive head coil. Preprocessing and quality control were undertaken by the UK Biobank research team according to published procedures. 45 Our analyses included total brain volume, brain volumes of 68 cortical areas, 14 subcortical structures, FA and MD of 25 white matter tracts. The measures of brain volume were corrected for head size by multiplication with the T1-based scaling factor (UK Biobank field ID 25000). The brain regions and white matter tracts used in the study are depicted in Figure S2.

| Anticholinergic burden and drug classification
Anticholinergic scales typically score drugs on a 0-3 ordinal scale, with a higher score indicating greater anticholinergic potency. We considered 15 anticholinergic scales-13 28,46-57 were based on our previous analyses 1 while 2 scales 58,59 were identified through a recent review 7 (Table S4). Three scales 47,50,56 were modified to include newer drugs. 1,60 One scale 52 was modified so that drugs with improbable anticholinergic action were assigned an anticholinergic burden of 0.5 as was done before. 1 Using the British National Formulary (https://bnf.nice.org.uk/, last accessed on 11 March 2021), we replaced brand names with generic names. Combination prescriptions containing several anticholinergics were each separated to yield multiple prescriptions, each containing a single anticholinergic. Each prescription was then assigned a potency score from each anticholinergic scale. For analysis, the cumulative AChB was calculated by summing the AChB scores across all prescribed drugs in the sampling period.
The sampling period excluded the year preceding the UK Biobank assessment to avoid acute effects of drugs. Prescriptions of drugs with ophthalmic, otic, nasal or topical routes of administration were all assigned an anticholinergic score of 0, as previously reported. 1,[54][55][56][57] Each drug was assigned to a class in the WHO Anatomical Therapeutic Chemical (ATC) Classification system (https://www.whocc.no, last accessed on 11th March 2022) 61 that categorizes drugs in a 5-level hierarchy. In our analyses, the first (anatomical main group) and third (pharmacological subgroup) levels were used.

| Data preparation
Prescriptions issued before 2000 and after 2015 were removed due to low ascertainment and incomplete annual data, respectively. 1 Participants with a diagnosis of diseases that may affect brain structure or cognitive ability were removed. The data-cleaning process is depicted in Figure S3. Outliers for numerical variables were defined as values lying 4 or more standard deviations or interquartile ranges beyond the mean or median, whichever was most appropriate according to the distribution. The total number of prescribed drugs and the AChB scores were strongly right-skewed due to the high numbers of zero values. For these variables, zeroes were removed before identifying outliers. All outliers were removed before analysis. Model assumptions were mostly met, but some models exhibited non-normality in the distribution of residuals ( Figure S4).

| Modelling
We applied principal component analysis to tract-specific FA and MD and used the first principal component to compute the general FA and MD (gFA and gMD), accounting for 44 and 50% of the variance, respectively. The standardized loadings and proportional variance for gFA and gFA are presented in Figure S5 and Table S5. We used linear regression models to estimate the association between AChB, cognitive ability and brain structure. To compare anticholinergic scales, we first modelled the association between g and AChB for each scale separately. This was later repeated for total brain volume as the outcome.
The scale exhibiting the strongest association with g was selected for subsequent analyses. Second, we modelled the effects of AChB due to different drug classes on g and total brain volume. Finally, we computed the associations between AChB and the results from 9 cognitive tests, the volumes of 68 cortical areas, 14 subcortical areas, gFA and gMD, and FA and MD of 25 white matter tracts. We also conducted 2 sensitivity analyses. First, we repeated the analyses on the association between AChB and g including only the year preceding the UK Biobank assessment to calculate AChB. Second, we computed the association between AChB according to each scale and g, while including an interaction term between AChB and age at assessment.
Each model was corrected for potential confounders, which included age at assessment, number of years over which the cumulative AChB was calculated, number of prescribed nonanticholinergics (different for each anticholinergic scale), data supplier of prescriptions (region-specific-2 for England, and 1 each for Scotland and Wales), socioeconomic deprivation (higher values correspond to greater deprivation; range: À6.3-11.0), 62 smoking status (nonsmoker, previous smoker, current smoker), frequency of alcohol consumption (daily or almost daily; 3 or 4 times a week; once or twice a week; once to 3 times a month; only on special occasions; never), level of physical activity (strenuous; moderate; mild), 63 body mass index (kg/m 2 ), APOE-carrier status, comorbidities count before the first assessment date (total number of distinct diagnoses codes), history of mood disorders, anxiety disorders, schizophrenia, diabetes, hypercholesterolemia, hypertension and myocardial infarction before the assessment date. APOE-carrier status was defined through the APOE genotype, which is based on the nucleotides at SNP positions rs239358 and rs7412. Participants were denoted as ε2, ε3, or ε4 carriers, if they carried the ε2/ε2 or ε2/ε3 haplotype, ε3/ε3 or ε1/ε3 or ε2/ε4 haplotype, or ε3/ε4 or ε4/ε4 haplotype, respectively. Smoking status, alcohol consumption, physical activity, body mass index and genotype were ascertained at each of the 2 UK Biobank assessments; socioeconomic deprivation was ascertained during the baseline assessment.
When comparing anticholinergic scales, 2 additional models were run for which polypharmacy was the main predictor. The first of these models (Polypharmacy model) controlled for the same covariates as above, and the second (Polypharmacy plus model) further controlled for the total number of anticholinergics according to any scale. The models where a measure of brain imaging was the main outcome, were in addition to the covariates above controlled for age 2 , age*sex, age 2 *age, head position in the MRI-scanner (3 coordinates), ethnicity and assessment centre. The template for the linear models is described in Text S1. Results are presented for models before the adjustment for polypharmacy and after adjustment for polypharmacy.

| Sample
After removing outliers, among the 163 043 participants in our sample, $140 000 and $14 000 data points (exact value depended on the model) were available for analyses of cognitive ability and brain imaging, respectively. The demographic-and lifestyle variables are presented in Table 1. While the imaging sample was older, the distribution of other variables was similar to the rest of the sample (Table S6). In the period from 2000 to the year before the initial assessment, anticholinergics-depending on the anticholinergic scale-represented between 4.3 and 24.1% of prescriptions, with between 11.3 and 40.7% of participants prescribed an anticholinergic at least once (Table S7). We have previously characterized anticholinergic prescribing and its longitudinal trends in UK Biobank in detail. 1

| AChB and cognition
When polypharmacy was not included as a control variable, all the tested anticholinergic scales exhibited significant negative associations with g (Table S8). The scales by Durán et al. 52 and by Cancelli et al. 49 showed the strongest (β = À0.032, P FDR < .001) and weakest (β = À0.009, P FDR < .001) effects, respectively. When the models were additionally corrected for polypharmacy, the median effect size of AChB across scales was reduced by 31%, but associations of all anticholinergic scales except the scale by Cancelli et al. 49 (β = À4.4 Â 10 À5 , P FDR = .88) remained significant ( Figure 1A, Table S8). The scale by Durán et al. 52 retained the strongest association (β = À0.025, P FDR < .001; Table S9). When the predictors were not standardized, this effect size corresponds to an at most 0.    Table S10).
When testing for the effects of drug classes, we found only limited instances in which higher AChB was associated with lower g (Figure 2, Table S11). Among the pharmacological classes, AChB due to drugs for migraine (β = 0.015, P FDR < .001) showed positive associations with g. AChB due to most other drugs exhibited negative associations with g, with β-lactam antibiotics (β = À0.035, P FDR < .001) and opioids (β = À0.026, P FDR < .001) showing the strongest effects, corresponding to respectively 0.033 and 0.010 decreases in the unstandardized g for each increase of AChB by 1 standard deviation.

| ACB and brain-imaging measures
AChB was not associated with brain atrophy irrespective of the anticholinergic scale used (range of β = À0.004-0.017, P FDR ≥ .21). While there were minor differences between the predictive power of different scales, the CIs overlapped across scale models and polypharmacy models (Figure 3, Tables S12 and S13).

| Sensitivity analyses
When the analyses on the associations between AChB and cognitive function were repeated using only AChB in the year before the assessment as the predictor (Tables S16-S19), the results exhibited similar trends to those observed in the main analyses. Most anticholinergic scales tended to negatively associate with cognitive function, albeit the effect sizes were smaller. Additionally, AChB was associated with lower performance in 1/5 cognitive tests available for this analysis. Furthermore, AChB due to β-lactam antibiotics and opioids again exhibited the strongest negative associations with g.
When g was modelled with the inclusion of an interaction term between age at assessment and AChB, the interaction was not significant (β = 3.0 Â 10 À4 , P = .38), indicating that the observed effect

| DISCUSSION
In this study, we found that most of the 15 studied anticholinergic scales exhibited significant associations with cognitive ability. This remained the case after controlling for multiple potential confounds, including the history of certain disorders and polypharmacy. Interestingly, the size of the effect was not moderated by age-middle-aged and older adults showed similar AChB-cognitive associations. While the positive association between higher AChB and lower cognitive ability largely agrees with previous studies on the topic, past results have been mixed. 7,16 One potential source of heterogeneity between studies is different control for polypharmacy, which may alter the results considerably. In our study, the addition of polypharmacy substantially decreased the size of the observed effects and was a stronger predictor of lower cognitive ability than AChB according to any of the studied anticholinergic scales. Another source of heterogeneity may be the differential effect of distinct drug classes. We found large differences between drug classes when predicting cognitive ability, with β-lactam antibiotics exhibiting larger effects than other drug classes. Moreover, antimigraine drugs were associated with higher cognitive ability. The effect of a general anticholinergic score may thus strongly depend on the structure of the sample and the precise prescribing characteristics of the participants.
In our study, general AChB was not predictive of any measure of between anticholinergic use and brain structure. They found anticholinergic use to associate with reduced cortical volume and reduced temporal lobe thickness, 12 increased rates of brain atrophy, 8 reduced grey matter density and functional connectivity in the nucleus basalis of Meynert, 38 and reduced volumes of both hippocampi. 39 It is unclear why our results from MRI structural imaging diverge from previous findings, as the studies described above display a range of characteristics that overlap with our own, including longitudinal data, 8 control for polypharmacy, 12 and the inclusion of middle-aged participants. 8,39 One possibility is that the previous studies mostly classified the predictor (e.g., anticholinergic users vs. nonusers), while we used a continuous measure of AChB. The pitfalls of categorisation and the loss of power for true effects have been discussed before. 65 Furthermore, the size of our imaging sample ($16 000) was several times larger ($3000). As has been recently reported, 66 brain-wide association studies may require thousands of participants to minimize effect size inflation and increase replication rates. Finally, all but 1 39 of the above studies focused on cognitive disorders or decline later in life, with 1 reporting an effect for specifically those participants that later developed mild cognitive impairment. 38 It is possible that while brain atrophy occurs in ageing or dementia, subtle cellular changes in Furthermore, our models carefully incorporated several important control variables, including the history of relevant disorders, polypharmacy, and several lifestyle and demographic factors. Finally, we adopted a robust approach to measuring cognitive ability that can reduce variability common in the assessment of separate cognitive domains.
However, we recognize several limitations. First, the UK Biobank sample is on average less deprived and healthier than the UK population 67 and thus not representative. Participants in the imaging subsample exhibit even better indicators of psychological and physical health than the UK Biobank average. 68 Both factors are likely to result in an underestimate of the effects present in the population. Second, the prescriptions included in our study do not incorporate over-thecounter drugs and we also have no information on how many prescriptions were dispensed or taken by participants. Third, brain imaging was sometimes performed after the coverage for prescriptions had concluded and the drugs potentially prescribed in the intervening period were not accounted for. This probably decreased the accuracy of our AChB measure for those participants. Fourth, our study was crosssectional and did not assess longitudinal changes in cognitive function and brain structure. This prevented us from establishing the sequence of events and from assessing associations between anticholinergic use and within-person changes. Finally, because AchB correlates with the number of anticholinergics, the effects of polypharmacy due to the use concurrently of several anticholinergic drugs and intrinsic anticholinergic activity of those drugs could not be completely separated.
Both the present study, as well as previous analyses have reported polypharmacy more broadly to be associated with poorer cognitive ability 69,70 and dementia. 71 A recent medication-wide association study 72 found that among 744 medicines, 30% were associated with dementia. Additionally, previous studies have reported on differences between drug classes in the association between AchB and dementia. [9][10][11] This finding was extended in the present study of general cognitive ability in a nonpathological sample. These results support a more nuanced approach that distinguishes between different classes of drugs beyond their assumed anticholinergic effects. For drug classes for which associations with lower cognitive ability or dementia can be demonstrated, more studies are needed to determine the effects of chronic use earlier in life, the impact of discontinuation and the potential neural correlates.
In summary, in this study, we found positive associations between long-term anticholinergic use and general cognitive ability across most studied anticholinergic scales. However, the associations held only for some drug classes and there was no evidence for differences in brain structure as a function of AChB. While the significant effect sizes observed in our study were modest, for complex, multicausal outcomes-especially in a large and relatively healthy sample-this is to be expected. For example, angiotensin converting enzyme inhibitors-one of the most common drugs to treat hypertension-have been shown to reduce systolic/diastolic pressure by merely À8/À5 mmHg. 73 When considered in the longterm and on the scale of entire populations, even tiny effects can accumulate to produce substantial health and economic consequences for society. Given sufficient confidence in a drug-outcome relationship and the availability of alternative treatments, changes in prescribing represent an intervention that is relatively simple to implement. This should serve as additional motivation for further research in the field.

ACKNOWLEDGEMENTS
This research was funded in whole, or in part, by the Wellcome Trust

DATA AVAILABILITY STATEMENT
Data from UK Biobank is available to approved researchers directly from UK Biobank. The code used to clean and analyse the data is available at https://github.com/JuM24/UKB-AChB-cognition-MRI.