Rivers Auty2020

Anti-inflammatories in Alzheimer’s disease - potential therapy or spurious correlate?
Jack Rivers-Auty1,2,3†, Alison E. Mather4,5, Ruth Peters6,7, Catherine B. Lawrence1,2, David Brough1,2.
Short title: NSAIDs in Alzheimer’s disease
Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

*for the Alzheimer’s Disease Neuroimaging Initiative
*Data used in preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging
Initiative (ADNI) database (adni.loni.usc.edu). As such, the investigators within the ADNI contributed to
the design and implementation of ADNI and/or provided data but did not participate in analysis or
writing of this report. A complete listing of ADNI investigators can be found at:
http://adni.loni.usc.edu/wp-content/uploads/how_to_apply/ADNI_Acknowledgement_List.pdf
1Division of Neuroscience and Experimental Psychology, School of Biological Sciences, Faculty of Biology,
Medicine and Health, Manchester Academic Health Science Centre, University of Manchester, AV Hill
Building, Oxford Road, Manchester, M13 9PT, U.K.
2Lydia Becker Institute of Immunology and Inflammation, University of Manchester, Manchester M13
9PT, UK
3Medical Sciences, Tasmanian School of Medicine, College of Health and Medicine, University of
Tasmania, Hobart, Tasmania, Australia
4Quadram Institute Bioscience, Norwich Research Park, Norwich, Norfolk, NR4 7UA, UK.
5University of East Anglia, Norwich Research Park, Norwich, Norfolk, NR4 7TJ, UK.
6School of Psychology, University of New South Wales, Sydney, NSW, Australia
7Neuroscience Research Australia, Barker Street, Sydney, New South Wales 2031, Australia
†to whom correspondence should be addressed: Jack Rivers-Auty. Tel: +61 (0)475 924 722; Fax: +44 161
275 5948; Email: jack.auty@UTAS.edu.au, Post: 243-14 Medical Science Precinct, 2, Liverpool Street,
University of Tasmania, Hobart, Tasmania, 7000, Australia
Search terms: Alzheimer’s disease, NSAID, inflammation, Cognitive decline
© The Author(s) (2020). Published by Oxford University Press on behalf of the Guarantors of Brain.
This is an Open Access article distributed under the terms of the Creative Commons Attribution License
(http://creativecommons.org/licenses/by/4.0/), which permits unrestricted reuse, distribution, and reproduction in any medium,
provided the original work is properly cited.
1
Keywords: Alzheimer’s disease, NSAID, inflammation, progression, anti-inflammatory
Abbreviations:
AD Alzheimer's disease
ADAS Alzheimer disease assessment scale
ADNI Alzheimer's disease neuroimaging initiative

AIC Akaike information criterion
ApoE Apolipoprotein E
CN Cognitively normal
COX Cyclooxygenase
EMCI Early mild cognitive impairment
GLMM Generalized linear mixed model
ID Identification
LMCI Late mild cognitive impairment
MCI Mild cognitive impairment
MMSE Mini-mental state examination
MRI Magnetic resonance imaging
NLRP3 NOD-like receptor family, pyrin domain containing 3
NSAID Non-steroidal anti-inflammatory
PET Positron emission tomography
2
https://mc.manuscriptcentral.com/braincom
Abstract
Epidemiological evidence suggests non-steroidal anti-inflammatories (NSAIDs) reduce the risk of

Alzheimer’s disease. However, clinical trials have found no evidence of NSAID efficacy. This
incongruence may be due to the wrong NSAIDs being tested in robust clinical trials or the
epidemiological findings being caused by confounding factors. Therefore, this study used logistic
regression and the innovative approach of negative binomial generalised linear mixed modelling to

investigate both prevalence and cognitive decline, respectively, in the Alzheimer’s Disease
NeuroImaging dataset for each commonly used NSAID and paracetamol.
Use of most NSAIDs were associated with reduced Alzheimer’s disease prevalence yet no effect on
cognitive decline was observed. Paracetamol had a similar effect on prevalence to these NSAIDs
suggesting this association is independent of the anti-inflammatory effects and that previous results
may be due to spurious associations. Interestingly, diclofenac use was significantly associated with both
reduce incidence and slower cognitive decline warranting further research into the potential therapeutic
effects of diclofenac in Alzheimer’s disease.
3
Introduction
Alzheimer’s disease (AD) is a debilitating age related dementia which is typified by initial loss of short
term memory and spatial awareness, followed by mid and long term memory loss, confusion,
personality changes, frailty, loss of motor function and death normally 5-7 years following initial
diagnosis (Wattmo et al., 2014). AD is the most prevalent form of dementia, constituting 60-80% of
dementia cases affecting an estimated 26 million people globally (Alzheimers Association, 2015).

Because of the severe social and economic costs, AD has been the focus of extensive research yet the
pathophysiology of AD remains poorly understood and disease modifying treatments continue to be
elusive. However, the role of neuroinflammation as a key etiological feature is now widely accepted due
the consensus of epidemiological, neuroimaging, preclinical and genetic evidence. Because of this, anti-
inflammatories have been thoroughly researched as putative disease modifying agents.
Non-steroidal anti-inflammatory drugs (NSAIDs), which inhibit cyclooxygenase (COX) enzymes and
subsequent prostanoid production, are the most commonly used anti-inflammatory drugs with over 110
million prescriptions annually in the USA alone (Conaghan, 2012). The prevalence of NSAIDs use makes
them an ideal candidate for epidemiological investigations into the potential therapeutic effects of anti-
inflammatories in AD. Numerous epidemiological studies in a range of ethnodemographic populations
have identified that NSAID use is associated with a lower risk of developing AD (Breitner et al., 1995,
Cote et al., 2012, Fischer et al., 2008, in 't Veld et al., 2001, Landi et al., 2003, Stewart et al., 1997,
Szekely et al., 2008, Vlad et al., 2008), and this association was suggested to be causal by numerous
intervention based studies in animal models (Yan et al., 2003, Lim et al., 2000, Weggen et al., 2001). This
led to a number of clinical trials of varying quality on NSAIDs and AD progression. Many of these trials
were short term (6-12 months) and had low numbers of patients due to the lack of private funding
available for existing drugs. Of the non-selective traditional NSAID clinical trials, Pasqualetti et al.
performed the most extensive trial with 132 patients followed for 1 year. A total of 51 and 46 patients
from the ibuprofen and placebo group completed the trial, respectively, and no differences in AD
Assessment Scale (ADAS) or the Mini-Mental State Examination (MMSE) were observed between
treatment groups (Pasqualetti et al., 2009). This is still a relatively small group of patients given the
variability of the disease and relatively short period of observation. Larger trials have been performed
with patented novel NSAIDs including trials of celecoxib with 425 subjects (Soininen et al., 2007),
rofecoxib with 692 subjects (Reines et al., 2004); and tarenfluribil with 1684 subjects (Green et al.,
2009), all were conducted for at least one year and all demonstrated no significant effect of the NSAIDs
on AD progression (Imbimbo et al., 2010). Several potential explanations for the discrepancy between
the efficacy of NSAIDs in the epidemiological and clinical research fields have been put forward: (i) The
NSAIDs effects seen in epidemiological research could be indirect through a hidden variable not
investigated by the research. (ii) NSAIDs may require a long period of administration before they can
provide a protective effect. This hypothesis is supported by epidemiological evidence; Stewart et al.
(1997) analysed a longitudinal cohort study of 1686 elderly individuals and found that the risk of AD was
only significantly decreased after more than two years of NSAID usage. (iii) The clinical trials
methodology were not optimized for the treatments in that they should have: only included early AD or
Mild Cognitively Impaired (MCI) individuals with confirmed amyloid positivity and neuroinflammation
through PET imaging, had larger numbers, and a longer period of treatment (Stewart et al., 1997,
Szekely & Zandi, 2010, Wyss-Coray, 2006). (iv) The NSAIDs selected for the high quality clinical trials
were chosen for the novelty and patentability of the drugs; as these drugs were not the focus of the
epidemiological or preclinical research it possible that there are different therapeutic profiles of
4
traditional NSAIDs and potentially the dominant mechanisms of action is independent of COX inhibition
(Stewart et al., 1997, Szekely & Zandi, 2010, Wyss-Coray, 2006). (v) While NSAIDs may reduce the risk of
developing AD, they do not slow the progression of the disease; suggesting that NSAIDs act on initiating
pathological processes of the disease not the downstream cascade of propagating mechanisms. The
latter two of these explanations (iv & v) are addressed in the present study by investigating the
association of individual NSAID use and cognitive decline (as opposed to incidence/prevalence) in MCI

and AD subjects, as measured by the MMSE and ADAS scores, in the Alzheimer’s disease Neuroimaging
Initiative (ADNI) case controlled longitudinal study dataset. From this it was found that most NSAIDs
were not associated with any change in cognitive decline including celecoxib, aspirin, ibuprofen and
naproxen. However, there was evidence that diclofenac was associated with slower cognitive decline.
Though not technically an NSAID, paracetamol (acetaminophen) was included in the analysis as a
common pain reliever which has indication overlap with NSAIDs and little mechanistic overlap.
Interestingly, paracetamol use was associated with accelerated cognitive decline. Collectively this study
concludes that the majority of NSAIDs do not affect the propagating mechanisms of AD and that the
therapeutic potential of a subset of NSAIDs including diclofenac is likely to be independent of COX
inhibition.
Methods
Data acquisition
Data used in the preparation of this article were obtained from the Alzheimer’s Disease Neuroimaging
Initiative (ADNI) database (adni.loni.usc.edu). The ADNI was launched in 2003 as a public-private
partnership, led by Principal Investigator Michael W. Weiner, MD. The primary goal of ADNI has been to
test whether serial magnetic resonance imaging (MRI), positron emission tomography (PET), other
biological markers, and clinical and neuropsychological assessment can be combined to measure the
progression of mild cognitive impairment (MCI) and early Alzheimer’s disease (AD).The subjects were
recruited from over 50 sites across the U.S. and Canada. ADNI has undergone three stages of
recruitment each with differences in the imaging and biomarker analyses, these have been named ADNI-
1, ADNI-GO and ADNI-2. Collectively these protocols have recruited 1631 adults into the study consisting
of age appropriate cognitively normal individuals, people with early or late MCI, and people with early
AD. The follow up duration of each group is specified in the protocols for ADNI-1, ADNI-2 and ADNI-GO
with 120 months as the maximum. Subjects were evaluated upon entry into the study, then at the 6 and
12 month time points, and yearly after this. For up-to-date information, see www.adni-info.org.
Data cleaning
The dataset of the medical history (RECMHIST.csv), recurrent medicines (RECCMEDS.csv) and patient
summary data (ADNIMERGE.csv) were downloaded on the 3rd of May 2018. To allow for adjustment for
relevant nuisance variables string search methods were applied to generate the explanatory variables of
headaches, arthritis, smoking, cardiovascular pathology and diabetes, where necessary (supplement
5.4). Cardiovascular pathology was defined as a subject diagnosed with hypertension or high cholesterol
(supplement 5.4.2). Terms varied widely and spelling errors were present and so manual confirmation of
correct identification was required. A similar process was required to identify recurrent oral
administration of NSAIDs (supplement 5.2). Only oral administration was included because topical
applications are not likely to reach relevant concentrations. The search terms required a mixture of
pharmacological and brand names. While a range of NSAIDs were searched for only aspirin, ibuprofen,
diclofenac, celecoxib and naproxen had sufficient numbers for analysis (supplement 5.2) and
5
paracetamol was included as a mechanistically distinct pain-reliever with similar potencies and
indications as NSAIDs. Individuals diagnosed as cognitively normal, MCI or AD were included in the
study, those with no diagnosis but reported subjective memory concerns (SMC) were removed as the
subjective nature of self-diagnosis may be a source of variability.
Distribution selection
To investigate cognitive decline over time generalised linear mixed modelling (GLMM) was applied. The

selection of distribution was thoroughly performed both prior and following model construction
investigating numerous distribution families (supplement 7.4, 7.5, 8.4, 8.5). Using graphical evaluation
and Akaike information criterion (AIC) to evaluate the appropriateness of the distribution family, the
negative binomial model was found to be the optimal distribution for both the MMSE and ADAS scores.
For the simplicity of the model, the MMSE score was converted from a count of correct answers with a
maximum of 30 into a count of incorrect answers (supplement 7.3), this has the advantage of now
having the same directional relationship with disease severity as the ADAS score, with higher values
correlating with worse cognitive performance and greater disease severity. Negative binomial models
are optimal for overdispersed Poisson (count) data, suggesting that MMSE and ADAS scores can be
modelled as a count of errors (supplement 7.5 & 8.5).
Model construction
Selecting parameterisation method
From the initial distribution analyses it was found that the variance was greater than the mean
indicating that the data were over-dispersed as a Poisson model supporting the use of negative binomial
models (supplement 7.5 & 8.5)(Bolker et al., 2011, Hardin et al., 2007). There are several
parameterisation methods which describe the relationship between the mean and the variance in the
negative binomial model (over-dispersion). The two most common (and the only methods available in
the glmmadmb package on R version 3.5.1 (R Core Team) with RStudio version 1.1.453 (Fournier DA et
al., 2012, Skaug H et al., 2013)) are the ‘nbinom1’ method, which assumes the variance = k × mean, and
‘nbinom2’ method, which assumes the variance = mean (1 + mean/k). The latter is most commonly
used, particularly in count datasets, and is derived from a Gamma/Poisson model of a heterogeneous
relationship between variance and mean (Hardin et al., 2007). The former describes a simple
proportional relationship between variance and mean and is less commonly used due to its
inflexibility(Hardin et al., 2007). Optimal parameterisation method was investigated using AIC and log-
likelihood both prior to, and following, the construction of the full models and the ‘nbinom1’
parameterisation was selected (supplement 7.6 & 8.6).
Building of initial main effect model

Building of the GLMMs followed the protocol outlined by Hosmer et al. (Hosmer et al., 2013). To
construct the negative binomial GLMM models the package glmmadmb was used on R version 3.5.1
with RStudio version 1.1.453 (Fournier DA et al., 2012, Skaug H et al., 2013). This package estimates
parameters using the maximum likelihood method with the Laplace approximation to assess the
marginal likelihood and provides coefficient summaries based on Wald approximations. The minimal
model used had the explanatory variable of time (month) included as well as subject ID as a random
effect (supplement 7.8 & 8.8). Then each biologically relevant explanatory variable was individually
added to the minimal model and then compared against the minimal model using the log likelihood
ratio test with p-value estimated using the Chi-squared distribution and AIC to investigate if the model
was significantly improved by the inclusion of the variable (supplement 7.8 & 8.8). Relevant nuisance
6
explanatory variables were first investigated, followed by pain-reliever use. These included ApoE4
genotype, age, diagnosis (control, MCI or Alzheimer’s disease), gender, education level, vascular
pathology, smoking, headaches, arthritis, diabetes and drug use (naproxen, celecoxib, diclofenac,
aspirin, ibuprofen or paracetamol). All significant variables were then included in the model and their
continued input into the model in the presence of the other explanatory variables was evaluated using
the Wald approximation statistics, Log likelihood ratio tests and AICs (supplement 7.9-7.12 & 8.9-8.12).

Variables that ceased to contribute significantly to the model were dropped and the final main effect
model of all significant variables was then constructed with p<0.05 considered as statistically significant
(supplement 7.12.8 & 8.12.8). Biologically relevant interaction terms were then investigated within the
main effect model.
Inclusion of biologically relevant two-way interaction terms

The inference of multivariable interaction terms becomes difficult, therefore, a common approach is to
investigate only biologically relevant two-way interaction terms (Hosmer et al., 2013). Similar to the
main effect analyses, each two-way interaction term was added to the main effect model in isolation
(supplement 7.13 & 8.13). Model was again assessed with the log likelihood ratio test and AIC values
(supplement 7.13 & 8.13). A final model was constructed including all significant interactions (p<0.05)
and Wald approximation statistics were scrutinized for non-significant coefficients (supplement 7.14 &
8.14). Each interaction was then dropped in isolation from the model and compared against the full
model including the variables that were dropped as main effect terms. The worsening of the model was
assessed with the log likelihood ratio test and AIC values (supplement 7.14-7.17 & 8.14-8.17).
Covariance matrices were constructed of the final model and no substantial multicollinearity was found
between included explanatory variables (supplement 7.22 & 8.22). The full model was then tested with
time (month) treated as a factor (as opposed to a continuous numerical variable) and years in education
treated as a continuous numeric variable (rather than being grouped into education levels of early,
middle, tertiary and post-graduate) (supplement 7.18-7.19 & 8.18-8.19). Treating month as a factor
introduced a substantial increase in the degrees of freedom into the model (supplement 7.18 & 8.18),
which resulted in levels of the models with insufficient data to stabilize the model resulting in issues of
model convergence, even when more simple models with fewer explanatory variables were attempted.
Treating years in education as a numeric variable worsened model fit; this is probably due to the lack of
correlation between numerical years and the (log) dependent variable (supplement 7.19 & 8.19).
Coefficient plots were generated with Laplace approximated confidence intervals to allow a quick
visualisation of the effects of significant explanatory variables on the modelled decline over time
(supplement 7.15 & 8.15). The variance/covariance matrices were used to generate Laplace plots of
modelled decline given different combinations of the significant explanatory variables (supplement 9).
To address issues of drop-out from the study for any reason, the analyses was repeated over the
timeframe of 48 months (which has limited drop-out). The results were largely unaffected, suggesting
drop-out was not having a strong bias effect on the full timeframe analyses (supplement 7.18, 8.18). We
used uncorrected p-values in the results to not limit the sensitivity of the study. The corrected p-values
(Holm-Šídák) are reported in the supplements. Example cognitive decline plots are displayed in the
results section, these relationships can be fully explored at this interactive website depicting cognitive
decline:
https://braininflammationgroup-universityofmanchester.shinyapps.io/Rivers-Auty-ADNI/
Final models for MMSE and ADNI
7
Below are the final models selected by the methods above for both MMSE and ADAS score.
MMSE decline ~ Time + Age (of participant) + Gender + Education level + Diagnosis (cognitive) + ApoE4
genotype + Diclofenac use + Aspirin use + Education level x Time + Diagnosis x Time + ApoE4 genotype x
Time + Gender x Time + Paracetamol use x Time + Diclofenac use x Time + (Patient ID as a random
variable)

ADAS decline ~ Time + Age (of participant) + Gender + Education level + Diagnosis (cognitive) + ApoE4
genotype + Headache + Diclofenac use + Ibuprofen use + Education level x Time + Diagnosis x Time +
ApoE4 genotype x Time + Gender x Time + Paracetamol use x Time + (Patient ID as a random variable)
Prevalence analysis
Baseline prevalence of AD was analysed with Chi-squared statistics and adjusted logistic regression.
Stepwise logistic regression with AIC as the selection criteria on the confounding variables was
performed to generate the base model. Each pain reliever was added to the model and the
improvement was evaluated based on Log likelihood ratio tests. P-values were adjusted using Holm-
Šídák multiple comparison adjustment method (supplement 6.4.4).
Assumption check of residuals

The Pearson residuals were extracted and plotted against the explanatory variables grouped by
individual ID. The ungrouped Pearson residuals were also plotted. No trends were observed for any
explanatory variable, indicating the appropriateness of the negative binomial models (supplement 7.21
& 8.21).
Data availability
For transparency and repeatability, the code and results of the complete analyses summarised in the
manuscript are included in full in the supplemental material. The data for the analyses presented here
are available through application to the ADNI data repository at www.adni-info.org.
Results
Baseline statistics
This study included 1619 individuals of whom 338 (21%) had AD, 560 (35%) had late mild cognitive
impairment (LMCI), 306 (19%) had early mild cognitive impairment (EMCI) and 415 (26%) were
cognitively normal (CN). The proportions of each cognitive diagnosis did significantly differ with pain-
reliever use (supplement 6.4.4). Celecoxib, diclofenac, ibuprofen, paracetamol, aspirin and naproxen use
were all associated with a substantially reduced AD prevalence compared to the no pain-reliever group
(table 1, fig. 1, supplement 6.4), this effect remained after adjusting for confounding variables
(supplement 6.4.4). This corresponded to significantly different mean baseline ADAS and MMSE scores
(table 1, supplement 6.4). Other baseline metrics were nominally similar including gender proportions,
mean age, ApoE4 status, educational attainment and diabetes prevalence (table 1, supplement 6).
Headache, arthritis and cardiovascular risk factors were elevated in the pain-reliever groups (table 1,
supplement 6). This is unsurprising as these pain-relievers are indicated for these conditions. Arthritis
prevalence was highest in more potent pain-reliever groups such as celecoxib and diclofenac (73% and
77%, respectively), compared to 28% prevalence in the no pain-reliever group (table 1, supplement
6.4.10).
8
The Effect of NSAIDs on Cognitive Scores
Aspirin, ibuprofen, naproxen and celecoxib were not found to be associated with any significant change
in cognitive decline as measure by MMSE or ADAS (table 2 & 3, fig 2, supplement 7.17 & 8.17).
Paracetamol use was associated with significantly accelerated decline in both MMSE and ADAS scores,
however, the effect size is of limited clinical relevance (table 2 & 3, fig. 2). Diclofenac was found to be
the only NSAID which was associated with reduced cognitive decline as measured by the MMSE score

(table 2 & 3, fig. 2), and this effect approached significance for the ADAS score (supplement 8.17) with
clinically meaningful effect sizes (fig. 2). The effect of diclofenac on MMSE decline remained significant
(p=0.039) after correcting for multiple comparisons (supplement 7.17).
There was some evidence of a main effect of aspirin and ibuprofen being associated with slightly
improved MMSE and ADAS scores, respectively (table 2 & 3, fig. 2, supplements 7.17 & 8.17). This
suggests that their use is associated with a mild fixed positive effect on cognitive scores, but they were
not associated with altered progression of cognitive decline (table 2 & 3, fig. 2, supplements 7.17 &
8.17). Diclofenac was associated with a significant positive effect on ADAS scores when included in the
model only as a main effect (table 2 & 3, fig. 2, supplement 7.17 & 8.17), however, this main effect was
not significant once the interaction term with time was included, suggesting the predominate effect of
diclofenac is on cognitive decline.
As expected, there were significant main effects and effects on progression conferred by cognitive
diagnosis with AD and LMCI both having worse MMSE and ADAS values and accelerated decline (table 2
& 3, fig. 3, supplements 7.16 & 8.16). However, EMCI was not associated with accelerated progression
of cognitive decline compared to the CN diagnosis, suggesting EMCI has limited prognostic utility (table
2 & 3, fig. 3, supplements 7.16 & 8.16).
Education level had a complex relationship with cognitive decline. Post-graduate level study was set as
the reference level. All other education levels had worse cognitive performance as main effects,
however, their progression slopes were less severe, with tertiary level education associating with the
slowest progression (table 2 & 3, fig. 3, supplements 7.16 & 8.16). A simplified inference of this is that
post-graduate level studies was associated with initial good performance in the cognitive tasks but
faster decline compared to tertiary, secondary and early education levels (table 2 & 3, fig. 3,
supplements 7.16 & 8.16).
Age did have a significant main effect on MMSE and ADAS scores associating with worse cognitive scores
(table 2 & 3, fig. 3, supplements 7.16 & 8.16). However, there was no evidence that age was associated
with altered progression (supplements 7.16 & 8.16).
ApoE4 genotype had a substantial gene dose main effect on MMSE and ADAS score associating with
worse cognitive performance (table 2 & 3, fig. 3, supplements 7.16 & 8.16), as well as an association
with substantially accelerated cognitive decline (table 2 & 3, fig. 3, supplements 7.16 & 8.16).
There was no discernible significant association of cardiovascular risk factors, smoking or diabetes on
MMSE or ADAS score as a main effect or altering progression, therefore, for model parsimony they were
not included in the final models (supplements 7.16 & 8.16).
Discussion
9
Here we used the innovative approach of negative binomial generalized linear modelling to analyse the
association between pain-reliever use and cognitive decline in CN, MCI and AD individuals in the ADNI
dataset. From this we found that, while pain-reliever use was associated with a lower prevalence of AD,
there were no similarly positive associations with delayed cognitive decline, with the exception of
diclofenac use. This is congruent with the decades of epidemiological evidence which suggests that
NSAID use lowers the prevalence of AD (Breitner et al., 1995, Cote et al., 2012, Fischer et al., 2008, in 't

Veld et al., 2001, Landi et al., 2003, Stewart et al., 1997, Szekely et al., 2008, Vlad et al., 2008) and the
limited number of clinical trials which have found no effect of NSAIDs on disease progression
(Alzheimer's Disease Anti-inflammatory Prevention Trial Research Group, 2013, Green et al., 2009,
Lyketsos et al., 2007, Pasqualetti et al., 2009, Scharf et al., 1999). This suggests that either the
therapeutic window of pain-relievers of AD is pre-symptomatic, acting on the initiating mechanisms of
cognitive decline and not the propagating mechanisms of AD, or there is a hidden variable which
explains the lower prevalence of AD in pain-reliever users, for example healthy user bias (Shrank et al.,
2011). Our analysis supports the latter. Healthy user bias is common in epidemiological research, it is
caused by the effect of healthier individuals seeking and using therapies such as pain-relievers, resulting
in spurious associations of therapy use and reduced disease prevalence. Our prevalence analysis shows
similar reductions in the proportion of AD diagnoses between all pain-relievers, even structurally and
functionally dissimilar compounds such as aspirin (weak COX1 and COX2 inhibitor), celecoxib (potent
COX2 inhibitor) and paracetamol (endocannabinoid modulator, unlikely to inhibit COX1 and 2 at
physiological concentrations (Klinger-Gratz et al., 2018)). The similarity in effects on AD incidence,
despite structural and functional difference supports the existence of a hidden variable such as the
healthy user bias. Furthermore, individuals taking any of the investigated drugs at the beginning of the
study were less likely to drop-out (due to very poor health, death or other). This again supports the
healthy user bias conclusion, as healthier subjects were more likely to remain in the study. This
alternative explanation of the epidemiological AD prevalence literature, is further supported by the
placebo controlled AD Anti-inflammatory Prevention Trial (ADAPT) trial (2013), which investigated the
effects of celecoxib or naproxen on AD incidence in 2528 cognitively normal elderly people and found no
positive effects, suggesting that COX inhibition is ineffective even in the presymptomatic stages of the
disease.
Unlike the other pain-relievers, diclofenac use was associated with a slower cognitive decline as
measured by MMSE scores and approached significance with ADAS scores. Though not thoroughly
researched, there is evidence that diclofenac is a promising avenue of therapeutic development for AD.
Landi et al. (2003) performed a cross-sectional study of 2708 community dwelling elderly people. They
utilized logistic regression on the proportions of those diagnosed with AD in each NSAID category and
found that diclofenac had the greatest effect on risk of AD diagnosis with an odds ratio of 0.21 (95%CI of
0.05 – 0.90). Similarly, in the present study, diclofenac had the lowest prevalence of AD of all pain-
relievers tested, suggesting a potential prophylactic effect, as well as the reported potential therapeutic
effect on disease progression. Furthermore, a small underpowered clinical trial was performed by Scharf
et al. (1999). This was a single centre trial recruiting mild to moderate AD patients defined by an MMSE
of 11 to 25. A total of 41 patients were recruited and 24 were randomly allocated to the placebo group
and 17 to the daily diclofenac treatment group and the patients were followed up for cognitive
assessment after six months of treatment (Scharf et al., 1999). Due to the lack of power of the study and
short time span, no strong inferences should be made, however, the trends largely concur with the
Landi et al. (2003) study and the present study; the placebo group declined cognitively with a mean
10
MMSE score decline of -0.86 (SD 3.21) and an ADAS increase of 1.93 (SD 5.55), while the diclofenac
MMSE score improved on average 0.41 (SD 2.69) and the ADAS stayed relatively stable with a slight
increase of 0.25 (SD 4.5)(Scharf et al., 1999). Given that the other NSAIDs tested in clinical trials did not
slow the progression of AD and were not associated with slower decline in the present study despite
also being potent inhibitors of the COX enzymes, it is fair to hypothesise that any potential effects of
diclofenac on AD are likely not through this mechanism of action. Unlike the other pain-relievers

investigated in this research, diclofenac also inhibits the release of the inflammatory cytokine
interleukine-1β (IL-1β) (26, supplement 1) by inhibiting the activation of the intracellular receptor NLRP3
(Nod-like receptor family, pyrin domain containing 3). The NLRP3 receptor in microglia has been shown
to be central to the neuroinflammatory response observed in mouse models of AD (Daniels et al., 2016)
and inhibition of the NLRP3 receptor with similar compounds has been found to be therapeutic in
several animal models of AD (Daniels et al., 2016, Dempsey et al., 2017). Furthermore, studies have
found that the genetic deletion of this receptor completely abated the AD phenotype in mouse models
(Dostert et al., 2008, Heneka et al., 2013). Therefore, NLRP3 inhibition may be the defining feature of
diclofenac. However, it should be noted that only 30 subjects were diclofenac consumers with sufficient
data for inclusion in the analysis. Therefore, any strong inference of efficacy should be avoided as future
research is needed on this promising NSAID.
The results of the analyses presented here found substantial evidence for ApoE4 causing accelerated
cognitive decline. The lipoprotein ApoE4 is a well-established risk factor for the development of AD
(Szekely et al., 2008, Notkola et al., 1998, Cornelius et al., 2004). Several studies have also found that
ApoE4 alleles are associated with accelerated cognitive decline and accelerated cortical tissue atrophy
(Bartzokis et al., 2006, Lo et al., 2011, Mielke et al., 2011, Morra et al., 2009, Rawle et al., 2018, Tilvis et
al., 2004, Whitehair et al., 2010, Young et al., 2014, Kim et al., 2017, Kanai et al., 1999, Vijayaraghavan
et al., 2016). However, a limitation common in the existing literature is the use of multi-level linear
modelling for non-Gaussian discrete cognitive scores (Kanai et al., 1999, Kim et al., 2017, Lo et al., 2011,
Mielke et al., 2011, Rawle et al., 2018, Vemuri et al., 2014, West et al., 2008, Whitehair et al., 2010),
therefore, the present study, with 744 individuals with at least one ApoE4 gene, investigating both ADAS
and MMSE measures of cognitive decline and applying discrete distribution GLMM analyses with the
negative binomial models, represents a robust and important contribution to the field.
Previous research has reported that NSAIDs only alter AD incidence in ApoE4 carriers, suggesting an
interaction between the potential therapeutic mechanism of NSAIDs and the pathological mechanisms
of ApoE4 . This was first reported in a thorough study by Szekely et al. (2008) who looked at AD
incidence in 3229 elderly people during a 10 year period. They found that NSAID use was associated
with a hazard ratio of 0.88 in non-ApoE4 carriers and 0.34 in ApoE4 carriers (compared to matched
individuals). The present study investigated a three-way interaction of NSAIDs, ApoE4 genotype and
time (month), and no significant effects were observed (supplement 7.20 & 8.20). This suggests that if
the Szekely et al. finding is due to NSAID-ApoE4 interactions, and not hidden nuisance variables present
in the analysis, then this therapeutic effect may only be useful for the prophylactic treatment to prevent
AD in ApoE4 individuals and not effective in altering the progression of the disease.
Conclusion
The present study is a thorough investigation into the effects of NSAID and paracetamol use on AD and
MCI cognitive decline. Also investigated were the effects of gender, smoking status, headaches, arthritis,
diabetes, age, vascular pathology, ApoE4 genotype and education level. Due to the discrete nature of
11
the dependent variables MMSE and ADAS scores, GLMMs were investigated and the negative binomial
distribution was found to be a robust approach which outperformed other models. Ibuprofen and
aspirin use were associated with improved cognitive performance at baseline, however, neither were
associated with an altered cognitive decline. Naproxen and celecoxib use were not associated with any
significant alterations in cognitive performance and paracetamol use was associated with accelerated
cognitive decline although this effect size was negligible. This suggests that NSAIDs and paracetamol are

not promising therapeutics for altering the progression of cognitive decline in MCI and AD individuals.
However, diclofenac use was associated with slower cognitive decline, and as this was the only NSAID to
do so, this suggests that COX inhibition is not the likely mechanism of action. Therefore, the full
interactome of diclofenac should be investigated for potential therapeutic avenues. Collectively, the
present study found interesting future avenues of research particularly the effects of paracetamol and
diclofenac on AD progression and improved the evidence for our existing understanding of factors which
effect AD such as the ApoE4 genotype by applying innovative statistical methods.
Funding
Data collection and sharing for this project was funded by the Alzheimer's Disease Neuroimaging
Initiative (ADNI) (National Institutes of Health Grant U01 AG024904) and DOD ADNI (Department of
Defense award number W81XWH-12-2-0012). ADNI is funded by the National Institute on Aging, the
National Institute of Biomedical Imaging and Bioengineering, and through generous contributions from
the following: AbbVie, Alzheimer’s Association; Alzheimer’s Drug Discovery Foundation; Araclon Biotech;
BioClinica, Inc.; Biogen; Bristol-Myers Squibb Company; CereSpir, Inc.; Cogstate; Eisai Inc.; Elan
Pharmaceuticals, Inc.; Eli Lilly and Company; EuroImmun; F. Hoffmann-La Roche Ltd and its affiliated
company Genentech, Inc.; Fujirebio; GE Healthcare; IXICO Ltd.; Janssen Alzheimer Immunotherapy
Research & Development, LLC.; Johnson & Johnson Pharmaceutical Research & Development LLC.;
Lumosity; Lundbeck; Merck & Co., Inc.; Meso Scale Diagnostics, LLC.; NeuroRx Research; Neurotrack
Technologies; Novartis Pharmaceuticals Corporation; Pfizer Inc.; Piramal Imaging; Servier; Takeda
Pharmaceutical Company; and Transition Therapeutics. The Canadian Institutes of Health Research is
providing funds to support ADNI clinical sites in Canada. Private sector contributions are facilitated by
the Foundation for the National Institutes of Health (www.fnih.org). The grantee organization is the
Northern California Institute for Research and Education, and the study is coordinated by the
Alzheimer’s Therapeutic Research Institute at the University of Southern California. ADNI data are
disseminated by the Laboratory for Neuro Imaging at the University of Southern California. AEM is a
Food Standards Agency Fellow and is supported by the Biotechnology and Biological Sciences Research
Council (BBSRC) Institute Strategic Programme Microbes in the Food Chain BB/R012504/1 and its
constituent projects BBS/E/F/000PR10348 (Theme 1, Epidemiology and Evolution of Pathogens in the
Food Chain) and BBS/E/F/000PR10351 (Theme 3, Microbial Communities in the Food Chain). JRA was a
Future Leader Fellow supported by the BBSRC fellowship grant titled Understanding how dietary zinc
and inflammation impact healthy ageing in the brain (BB/P01061X/1).
References
Alzheimer's Disease Anti-Inflammatory Prevention Trial Research Group -, Results of a follow-up study
to the randomized Alzheimer's Disease Anti-inflammatory Prevention Trial (ADAPT). Alzheimers Dement
2013; 9, 714-23.
Alzheimers Association -, Alzheimer's Association Report 2015 Alzheimer's disease facts and figures.
Alzheimers Dement 2015; 11, 332-84.
12
Bartzokis G, Lu PH, Geschwind DH, Edwards N, Mintz J, Cummings JL, Apolipoprotein E genotype and
age-related myelin breakdown in healthy individuals - Implications for cognitive decline, and dementia.
Arch. Gen. Psychiatry 2006; 63, 63-72.
Bolker BM, Brooks ME, Clark CJ, Geange SW, Poulsen JR, Stevens MHH, et al., GLMMs in action: gene-
by-environment interaction in total fruit production of wild populations of Arabidopsis thaliana Revised

version. See http://glmm.wdfiles.com/local--files/examples/Banta_2011_part1.pdf 2011.
Breitner JCS, Welsh KA, Helms MJ, Gaskell PC, Gau BA, Roses AD, et al., DELAYED-ONSET OF
ALZHEIMERS-DISEASE WITH NONSTEROIDAL ANTIINFLAMMATORY AND HISTAMINE-H2 BLOCKING-
DRUGS. Neurobiol. Aging 1995; 16, 523-30.
Conaghan PG, A turbulent decade for NSAIDs: update on current concepts of classification,
epidemiology, comparative efficacy, and toxicity. Rheumatol. Int. 2012; 32, 1491-502.
Cornelius C, Fastbom J, Winblad B, Viitanen M, Aspirin, NSAIDs, risk of dementia, and influence of the
apolipoprotein E epsilon 4 allele in an elderly population. Neuroepidemiology 2004; 23, 135-43.
Cote S, Carmichael PH, Verreault R, Lindsay J, Lefebvre J, Laurin D, Nonsteroidal anti-inflammatory drug
use and the risk of cognitive impairment and Alzheimer's disease. Alzheimers Dement 2012; 8, 219-26.
Daniels MJD, Rivers-Auty J, Schilling T, Spencer NG, Watremez W, Fasolino V, et al., Fenamate NSAIDs
inhibit the NLRP3 inflammasome and protect against Alzheimer's disease in rodent models. Nat.
Commun. 2016; 7.
Dempsey C, Rubio Araiz A, Bryson KJ, Finucane O, Larkin C, Mills EL, et al., Inhibiting the NLRP3
inflammasome with MCC950 promotes non-phlogistic clearance of amyloid-beta and cognitive function
in APP/PS1 mice. Brain. Behav. Immun. 2017; 61, 306-16.
Dostert C, Petrilli V, Van Bruggen R, Steele C, Mossman BT, Tschopp J, Innate immune activation
through Nalp3 inflammasome sensing of asbestos and silica. Science 2008; 320, 674-7.
Fischer P, Zehetmayer S, Jungwirth S, Weissgram S, Krampla W, Hinterberger M, et al., Risk factors for
Alzheimer dementia in a community-based birth cohort at the age of 75 years. Dement. Geriatr. Cogn.
Disord. 2008; 25, 501-7.
Fournier Da, Skaug Hj, Ancheta J, Ianelli J, Magnusson A, Maunder M, et al., AD Model Builder:
usingautomatic differentiation for statistical inference of highly parameterized complex nonlinear
models. Optim. Methods Softw. 2012; 27, 233-49.
Green RC, Schneider LS, Amato DA, Beelen AP, Wilcock G, Swabb EA, et al., Effect of Tarenflurbil on
Cognitive Decline and Activities of Daily Living in Patients With Mild Alzheimer Disease A Randomized
Controlled Trial. JAMA 2009; 302, 2557-64.
13
Hardin JW, Hilbe JM, Hilbe J, 2007. Generalized linear models and extensions. Stata press.
Heneka MT, Kummer MP, Stutz A, Delekate A, Schwartz S, Vieira-Saecker A, et al., NLRP3 is activated in
Alzheimer's disease and contributes to pathology in APP/PS1 mice. Nature 2013; 493, 674-+.
Hosmer JDW, Lemeshow S, Sturdivant RX, 2013. Model-Building Strategies and Methods for Logistic

Regression. In. Applied Logistic Regression. John Wiley & Sons, Inc., 89-151.
Imbimbo BP, Solfrizzi V, Panza F, Are NSAIDs useful to treat Alzheimer's disease or mild cognitive
impairment? Front. Aging Neurosci 2010; 2.
In 'T Veld BA, Ruitenberg A, Hofman A, Launer LJ, Van Duijn CM, Stijnen T, et al., Nonsteroidal
antiinflammatory drugs and the risk of Alzheimer's disease. N. Engl. J. Med. 2001; 345, 1515-21.
Kanai M, Shizuka M, Urakami K, Matsubara E, Harigaya Y, Okamoto K, et al., Apolipoprotein E4

accelerates dementia and increases cerebrospinal fluid tau levels in Alzheimer's disease. Neurosci. Lett.
1999; 267, 65-8.
Kim YJ, Seo SW, Park SB, Yang JJ, Lee JS, Lee J, et al., Protective effects of APOE e2 against disease
progression in subcortical vascular mild cognitive impairment patients: A three-year longitudinal study.
Sci Rep 2017; 7, 1910.
Klinger-Gratz PP, Ralvenius WT, Neumann E, Kato A, Nyilas R, Lele Z, et al., Acetaminophen Relieves
Inflammatory Pain through CB<sub>1</sub> Cannabinoid Receptors in the Rostral
Ventromedial Medulla. J. Neurosci. 2018; 38, 322.
Landi F, Cesari M, Onder G, Russo A, Torre S, Bernabei R, Non-Steroidal Anti-Inflammatory Drug (NSAID)
Use and Alzheimer Disease in Community-Dwelling Elderly Patients. Am. J. Geriatr. Psychiatry 2003; 11,
179-85.
Lim GP, Yang F, Chu T, Chen P, Beech W, Teter B, et al., Ibuprofen suppresses plaque pathology and
inflammation in a mouse model for Alzheimer's disease. J. Neurosci. 2000; 20, 5709-14.
Lo RY, Hubbard AE, Shaw LM, Trojanowski JQ, Petersen RC, Aisen PS, et al., Longitudinal Change of
Biomarkers in Cognitive Decline. Arch. Neurol. 2011; 68, 1257-66.
Lyketsos CG, Breitner JC, Green RC, Martin BK, Meinert C, Piantadosi S, et al., Naproxen and celecoxib
do not prevent AD in early results from a randomized controlled trial. Neurology 2007; 68, 1800-8.
Mielke MM, Leoutsakos J-M, Tschanz JT, Green RC, Tripodis Y, Corcoran CD, et al., Interaction between
vascular factors and the APOE ε4 allele in predicting rate of progression in Alzheimer's disease. J.
Alzheimers Dis. 2011; 26, 127-34.
14
Morra JH, Tu Z, Apostolova LG, Green AE, Avedissian C, Madsen SK, et al., Automated mapping of
hippocampal atrophy in 1-year repeat MRI data from 490 subjects with Alzheimer's disease, mild
cognitive impairment, and elderly controls. Neuroimage 2009; 45, S3-S15.
Notkola IL, Sulkava R, Pekkanen J, Erkinjuntti T, Ehnholm C, Kivinen P, et al., Serum total cholesterol,
apolipoprotein E epsilon 4 allele, and Alzheimer's disease. Neuroepidemiology 1998; 17, 14-20.

Pasqualetti P, Bonomini C, Dal Forno G, Paulon L, Sinforiani E, Marra C, et al., A randomized controlled
study on effects of ibuprofen on cognitive progression of Alzheimer's disease. Aging Clin. Exp. Res.
2009; 21, 102-10.
R Core Team -, R: A language and environment for statistical computing. R
Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.
Rawle MJ, Davis D, Bendayan R, Wong A, Kuh D, Richards M, Apolipoprotein-E (Apoe) ε4 and cognitive
decline over the adult life course. Transl. Psychiatry 2018; 8, 18.
Reines SA, Block GA, Morris JC, Liu G, Nessly ML, Lines CR, et al., Rofecoxib - No effect on Alzheimer's
disease in a 1-year, randomized, blinded, controlled study. Neurology 2004; 62, 66-71.
Rstudio Team -, RStudio: Integrated Development for R. RStudio, Inc., Boston, MA URL
http://www.rstudio.com 2020.
Scharf S, Mander A, Ugoni A, Vajda F, Christophidis N, A double-blind, placebo-controlled trial of

diclofenac/misoprostol in Alzheimer's disease. Neurology 1999; 53, 197-201.
Shrank WH, Patrick AR, Brookhart MA, Healthy user and related biases in observational studies of
preventive interventions: a primer for physicians. J. Gen. Intern. Med. 2011; 26, 546-50.
Skaug H, Fournier D, Nielsen A, Magnusson A, Bolker B, Generalized Linear Mixed Models using AD
Model Builder. R package version 0.7.5 2013.
Soininen H, West C, Robbins J, Niculescu L, Long-term efficacy and safety of celecoxib in Alzheimer's
disease. Dement. Geriatr. Cogn. Disord. 2007; 23, 8-21.
Stewart WF, Kawas C, Corrada M, Metter EJ, Risk of Alzheimer's disease and duration of NSAID use.
Neurology 1997; 48, 626-32.
Szekely CA, Breitner JC, Fitzpatrick AL, Rea TD, Psaty BM, Kuller LH, et al., NSAID use and dementia risk
in the Cardiovascular Health Study: role of APOE and NSAID type. Neurology 2008; 70, 17-24.
15
Szekely CA, Zandi PP, Non-steroidal anti-inflammatory drugs and Alzheimer's disease: the
epidemiological evidence. CNS Neurol. Disord. Drug Targets 2010; 9, 132-9.
Tilvis RS, Kahonen-Vare MH, Jolkkonen J, Valvanne J, Pitkala KH, Strandberg TE, Predictors of cognitive
decline and mortality of aged people over a 10-year period. J. Gerontol. A. Biol. Sci. Med. Sci. 2004; 59,
268-74.

Vemuri P, Lesnick TG, Przybelski SA, Machulda M, Knopman DS, Mielke MM, et al., Association of
lifetime intellectual enrichment with cognitive decline in the older population. JAMA Neurol 2014; 71,
1017-24.
Vijayaraghavan S, Darreh-Shori T, Rongve A, Berge G, Sando SB, White LR, et al., Association of
Butyrylcholinesterase-K Allele and Apolipoprotein E varepsilon4 Allele with Cognitive Decline in
Dementia with Lewy Bodies and Alzheimer's Disease. J. Alzheimers Dis. 2016; 50, 567-76.
Vlad SC, Miller DR, Kowall NW, Felson DT, Protective effects of NSAIDs on the development of
Alzheimer disease. Neurology 2008; 70, 1672-7.
Wattmo C, Londos E, Minthon L, Risk Factors That Affect Life Expectancy in Alzheimer's Disease: A 15-
Year Follow-Up. Dement. Geriatr. Cogn. Disord. 2014; 38, 286-99.
Weggen S, Eriksen JL, Das P, Sagi SA, Wang R, Pietrzik CU, et al., A subset of NSAIDs lower
amyloidogenic A beta 42 independently of cyclooxygenase activity. Nature 2001; 414, 212-6.
West R, Beeri MS, Schmeidler J, Hannigan CM, Angelo G, Grossman HT, et al., Better memory
functioning associated with higher total and low-density lipoprotein cholesterol levels in very elderly
subjects without the apolipoprotein e4 allele. Am. J. Geriatr. Psychiatry 2008; 16, 781-5.
Whitehair DC, Sherzai A, Emond J, Raman R, Aisen PS, Petersen RC, et al., Influence of apolipoprotein E
varepsilon4 on rates of cognitive and functional decline in mild cognitive impairment. Alzheimers
Dement 2010; 6, 412-9.
Wyss-Coray T, Inflammation in Alzheimer disease: driving force, bystander or beneficial response? Nat.
Med. 2006; 12, 1005-15.
Yan Q, Zhang JH, Liu HT, Babu-Khan S, Vassar R, Biere AL, et al., Anti-inflammatory drug therapy alters
beta-amyloid processing and deposition in an animal model of Alzheimer's disease. J. Neurosci. 2003;
23, 7504-9.
Young AL, Oxtoby NP, Daga P, Cash DM, Fox NC, Ourselin S, et al., A data-driven model of biomarker
changes in sporadic Alzheimer's disease. Brain 2014; 137, 2564-77.
16
Page 17 of 25 Manuscripts submitted to Brain Communications
Table 1: Baseline statistics of ADNI cohort by pain-reliever use.

Aspirin Celecoxib Diclofenac Ibuprofen Naproxen Paracetamol No analgesic Statistics
Diagnosis
CN 247 (29%) 19 (30%) 14 (47%) 79 (31%) 52 (28%) 119 (30%) 99 (21%)
Χ2 (18)=78.1
EMCI 174 (20%) 13 (20%) 5 (17%) 55 (22%) 45 (24%) 75 (19%) 67 (14%)
p<0.0001
LMCI 302 (35%) 25 (39%) 8 (27%) 92 (36%) 68 (37%) 140 (35%) 165 (35%)
AD 138 (16%) 7 (10%) 3 (10%) 29 (11%) 21 (11%) 66 (17%) 141 (30%)
Χ2 (6)=57.5
ADAS 16.4 (9.0)*** 15.8 (8.3) ns 12.3 (8.5)* 15.0 (8.5)*** 15.4 (8.9) ns 16.2 (8.8)ns 19.7 (10.3) p<0.0001
Χ2 (6)=48.7
MMSE 27.3 (2.6)*** 27.2 (2.6) ns 28.0 (2.4) ns 27.7 (2.2)*** 27.6 (2.4) ns 27.4 (2.5)ns 26.4 (2.8) p<0.0001
Gender
Χ2 (6)=29.5
Male 323 (38%) 31 (48%) 15 (50%) 117 (46%) 95 (51%) 204 (51%) 223 (47%)
p<0.0001
Female 538 (62%) 33 (52%) 15 (50%) 138 (54%) 91 (49%) 196 (49%) 249 (53%)
F (6)=3.7
Age 74.3 (6.7) ns 73.4 (6.9)ns 75.22 (6.4)ns 72.7 (6.8)* 72.9 (7.1) ns 74.3 (7.2) ns 73.9 (8.1) P=0.001
ApoE4
-/- 467 (54%) 35 (54%) 18 (60%) 145 (57%) 102 (55%) 231 (58%) 236 (50%) Χ2 (12)=7.5
+/- 305 (35%) 24 (38%) 10 (33%) 88 (35%) 67 (36%) 133 (33%) 184 (39%) p=0.822
+/+ 89 (10%) 5 (8%) 2 (7%) 22 (9%) 17 (9%) 36 (9%) 52 (11%)
Education
Primary 318 (37%) 21 (33%) 10 (33%) 83 (33%) 56 (30%) 128 (32%) 161 (34%)
Χ2 (18)=15.7
Secondary 251 (29%) 18 (28%) 11 (37%) 77 (30%) 59 (32%) 119 (30%) 126 (27%)
p=0.613
Tertiary 166 (19%) 12 (19%) 3 (19%) 50 (20%) 40 (22%) 92 (23%) 92 (19%)
Post-grad 126 (15%) 13 (20%) 6 (20%) 45 (18%) 31 (17%) 61 (15%) 93 (19%)
Χ2 (6)=19.2
Headache 67 (8%) 11 (17%) 4 (13%) 31 (12%) 20 (11%) 52 (13%) 33 (7%) p=0.004
Χ2 (6)=19.2
Arthritis 342 (40%) 47 (73%) 23 (77%) 127 (50%) 98 (53%) 215 (54%) 132 (28%) P<0.001
Χ2 (6)=6.9
Diabetes 92 (7%) 7 (7%) 4 (9%) 26 (7%) 21 (8%) 46 (7%) 33 (4.7%) p=0.330
Χ2 (6)=4.5
Smoker 205 (24%) 17 (27%) 8 (27%) 71 (28%) 40 (22%) 107 (27%) 127 (27%) p=0.614
Cardiovascular Χ2 (6)=19.8
risk factors 555 (64%) 44 (69%) 16 (53%) 158 (62%) 114 (61%) 273 (68%) 261 (55%) p=0.003
Total 861 64 30 255 186 400 472
All data are in N (%) except for ADAS, MMSE or Age which are expressed as mean (SD). CN=cognitively normal, EMCI and LMCI = early and late mild cognitive
impairment, AD=Alzheimer’s disease, ADAS= Alzheimer’s disease assessment scale, MMSE=mini-mental state examination. Statistical analyses are Chi-squared
test for proportions, maximum likelihood generalised linear modelling for score data, or general linear modelling for parametric data.
Manuscripts submitted to Brain Communications Page 18 of 25
Table 2: Summary of the final negative binomial GLMM with MMSE failures as the dependent variable.
Statistics of
MMSE Estimate Stand Error Z value p-value
inclusion
Intercept -0.588 0.063 -9.33 p <0.00001 -
Fixed effects
Diagnosis
CN - - - -
χ2 (3)=1147.2,
EMCI 0.708 0.066 12 p <0.00001
p<0.00001
LMCI 1.242 0.056 22.38 p <0.00001
AD 2.11 0.062 33.9 p <0.00001

Gender
χ2 (1)=2.2,
Female - - - -
p=0.138010
Male 0.096 0.040 2.36 p =0.01806
χ2 (1)=48.2,
Age 0.018 0.003 6.96 p <0.00001
p<0.00001
ApoE4
-/- - - - - χ2 (2)=60.6,
+/- 0.121 0.042 2.84 p =0.00446 p<0.00001
+/+ 0.226 0.065 3.49 p =0.00048
Education
Primary 0.352 0.057 6.14 p <0.00001
χ2 (3)=47.4,
Secondary 0.362 0.056 6.5 p <0.00001
p<0.00001
Tertiary 0.2 0.050 4.04 p =0.00005
Post-grad - - - -
χ2 (1)=4.0,
Aspirin -0.075 0.037 -2.02 p =0.04345
p=0.045500
χ2 (1)=1.8,
Paracetamol -0.093 0.045 -2.07 p =0.03882
p=0.179712
χ2 (1)=1.0,
Diclofenac 0.023 0.146 0.015 p =0.87712
p=0.317311
Interaction with time (months)
Diagnosis
CN - - - -
χ2 (3)=79.8,
EMCI -0.00332 0.0010 -3.18 p =0.00145
p<0.00001
LMCI 0.00286 0.0007 3.95 p =0.00008
AD 0.00880 0.0014 6.12 p <0.00001
Gender
χ2 (2)=8.4,
Female - - - -
p=0.003752
Male -0.00172 0.0006 -2.92 p =0.00355
ApoE4
-/- - - - - χ2 (2)=120.6,
+/- 0.00588 0.0006 9.52 p <0.00001 p<0.00001
+/+ 0.00777 0.0009 8.32 p <0.00001
Education
Primary -0.00250 0.0008 -3.06 p =0.00224
χ2 (3)=26.6,
Secondary -0.00226 0.0008 -2.79 p =0.00532
p<0.00001
Tertiary -0.00375 0.0007 -5.05 p <0.00001
Post-grad - - - -
χ2 (1)=4.8,
Paracetamol 0.00129 0.0006 2.18 p =0.02928
p=0.0284597
χ2 (1)=8.4,
Diclofenac -0.00468 0.0016 -2.89 p =0.00380
p=0.0037522
Shown are the maximum likelihood estimates with Laplace estimates of the standard error, Z- value and p-value from the
Wald approximation, as well as, the significance of inclusion of the variable in the model evaluated using the log-likelihood
ratio test chi squared. CN=cognitively normal, EMCI and LMCI = early and late mild cognitive impairment, AD=Alzheimer’s
disease, ADAS= Alzheimer’s disease assessment scale, MMSE=mini-mental state examination.
Table 3: Summary of the final negative binomial GLMM with ADAS score as the dependent variable.
Statistics of
ADAS Estimate Stand Error Z value p-value
inclusion
Intercept 2.991 0.032 92.09 p <0.00001 -
Fixed effects
Diagnosis
CN - - - -
χ2 (3)=1124.4,
EMCI 0.357 0.035 10.18 p <0.00001
p<0.00001
LMCI 0.771 0.03 25.69 p <0.00001
AD 1.248 0.035 35.84 p <0.00001
Gender

χ2 (1)=7.4,
Female - - - -
p=0.006524
Male 0.106 0.023 4.6 p <0.00001
χ2 (1)=71.4,
Age 0.0133 0.003 8.52 p <0.00001
p<0.00001
ApoE4
-/- - - - - χ2 (2)=63.6,
+/- 0.096 0.024 3.92 p <0.00001 p<0.00001
+/+ 0.149 0.039 3.86 p = 0.0011
Education
Primary 0.181 0.033 5.41 p <0.00001
χ2 (3)=26.2,
Secondary 0.148 0.032 4.66 p <0.00001
p=0.00009
Tertiary 0.105 0.028 3.72 p =0.00020
Post-grad - - - -
χ2 (1)=5.0,
Headache -0.087 0.039 -2.23 p =0.02563
p=0.025347
χ2 (1)=0.8,
Paracetamol -0.035 0.026 -1.35 p =0.17759
p=0.371093
χ2 (1)=9.8,
Ibuprofen -0.093 0.030 -3.14 p =0.00170
p=0.0017451
χ2 (1)=7.8,
Diclofenac -0.224 0.080 -2.8 p =0.00518
p=0.0052246
Interaction with time (months)
Diagnosis
CN - - - -
χ2 (3)=65.4,
EMCI -0.00162 0.0004 -3.98 p =0.00007
p<0.00001
LMCI 0.00021 0.0003 0.74 p =0.46177
AD 0.00448 0.0007 6.17 p <0.00001
Gender
χ2 (2)=67.0,
Female - - - -
p<0.00001
Male -0.00210 0.0003 -8.22 p <0.00001
ApoE4
-/- - - - - χ2 (2)=186.2,
+/- 0.00298 0.0003 11.31 p <0.00001 p<0.00001
+/+ 0.00476 0.0004 10.61 p <0.00001
Education
Primary -0.00135 0.0004 -3.74 p =0.00019
χ2 (3)=50.6,
Secondary -0.00149 0.0003 -4.29 p =0.00002
p<0.00001
Tertiary -0.00214 0.0003 -6.91 p <0.00001
Post-grad - - - -
χ2 (1)=4.8,
Paracetamol 0.00056 0.0003 -2.21 p <0.00001
p=0.0284597
Shown are the maximum likelihood estimates with Laplace estimates of the standard error, Z- value and p-
value from the Wald approximation, as well as, the significance of inclusion of the variable in the model
evaluated using the log-likelihood ratio test chi squared. CN=cognitively normal, EMCI and LMCI = early and
late mild cognitive impairment, AD=Alzheimer’s disease, ADAS= Alzheimer’s disease assessment scale,
MMSE=mini-mental state examination.
Figure 1: At baseline, use of any pain-reliever was associated with lower prevalence of Alzheimer’s disease (AD) and a
corresponding higher prevalence of cognitively normal (CN) diagnoses. A) Proportion of cognitive diagnosis between
pain reliever groups. Dotted lines show proportion divisions of No pain-reliever group. B) Pearson residuals demonstrate
the size and direction of the effect of pain-reliever use on cognitive diagnosis. C) Contribution analysis of the significant
association of pain-reliever subgroup and cognitive diagnosis reveals little difference between pain-reliever and the
largest contribution to the significant effect is the No pain-relief group having a higher prevalence of AD and a lower
prevalence of CN, compared to the other groups. CN=cognitively normal, EMCI and LMCI = early and late mild cognitive
impairment, AD=Alzheimer’s disease, ADAS= Alzheimer’s disease assessment scale, MMSE=mini-mental state
examination.
Figure 2: The effect of pain-reliever use on predicted cognitive decline as measure by MMSE (i) and ADAS (ii) scores.

These models show the predicted decline of a LMCI, Female 70 year old, and how this decline changes when pain-relief
use is included in the model. Any NSAID use (A); Aspirin (B); Celecoxib (C); Diclofenac (D); Ibuprofen (E); Naproxen (F);
Paracetamol (G); Lines are predicted value, shaded area are 95% CI. LMCI = late mild cognitive impairment, ADAS=
Alzheimer’s disease assessment scale, MMSE=mini-mental state examination.
Figure 3: The effect of significant explanatory variables on predicted cognitive decline as measure by MMSE (i) and
ADAS (ii) scores. These models show the predicted decline of a LMCI (except A), Female (except C), 70 year old, and how
this decline changes when additional variables are included. Cognitive diagnosis was predictably the most influential
variable, with Alzheimer’s disease (AD) and late mild cognitive impairment (LMCI) associated with rapid cognitive
decline; early MCI (EMCI) and cognitively normal (CN) did not have appreciatively different rates of decline (A). ApoE4
showed a very strong gene dose association with accelerated cognitive decline (B). Gender showed mild differences with
males having faster cognitive decline (C). Education had variable effects on cognitive decline with tertiary level education
associated with the slowest decline (D). Lines are predicted value, shaded area are 95% CI.
Competing interests
The authors have no competing interests to report.
A
100
1
2
3
Prevalence of diagnosis (%)
4 80
5
6
7

8 CN
9 60
10 EMCI
11
12
LMCI
13 40 AD
14
15
16
17
18 20
19
20
21
22 0
23
ol
ac
ib
n
fen
f
in
lie
24
xe
tam
ox
pir
en
-re
pro
pro
25
lec
As
lof
ce
in
26
Ibu
Na
Ce
Dic
pa
27 ra
Pa
No
28
29
30
B
31
C
EMCI
EMCI
LMCI
LMCI
32
CN
CN
AD
AD
33
34
35 Aspirin Aspirin
36
37
38 Celecoxib Celecoxib
39
40
41 Diclofenac Diclofenac
42
43
44 Ibuprofen Ibuprofen
45
46
47
Naproxen Naproxen
48
49
50
51Paracetamol Paracetamol
52
53
No pain-relief
54 No pain-relief
55
56
57 −8 0 https://mc.manuscriptcentral.com/braincom
8 0 12.5 25 37.5 50
−4 4
58
59 Pearson residual Contribution to significant effect (%)
60
A B
i 30 ii i ii
30
50
50
1
2
40
40
MMSE
MMSE
3
27
27
4
ADAS
ADAS
30
30
5

24
24
6
7
20
20
8
21
21
Non−User Non−User
9
NSAID User Apirin User
10
10
10
11
C
12
0 12 24 36 48 60
Month
72 84 96 108 120 0 12 24 36 48 60
Month
72 84 96 108 120
D 0 12 24 36 48 60
Month
72 84 96 108 120 0 12 24 36 48 60
Month
72 84 96 108 120
13
i ii i ii
30
30
50
50
14
15
16
40
40
MMSE
MMSE
27
27
17
ADAS
ADAS
18
30
30
19
24
24
20
20
20
21
21
21
22 Non−User Non−User
10
23 Celecoxib User 10 Diclofenac User
24
25 0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120
Month Month Month Month
26
E
27
F
28 i ii i ii
30
30
50
50
29
30
40
40
MMSE
MMSE
31
27
27
32
ADAS
ADAS
30
30
33
24
24
34
35
20
20
36
21
21
Non−User Non−User
37
Ibuprofen User Naproxen User
10
10
38
39 0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120
0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120
40
Month Month Month Month
G
41
42 i ii
30
50
43
44
40
45
MMSE
27
46
ADAS
47
30
24
48
49
20
50
21
51 Non−User
ParaceUser
10
52
53 https://mc.manuscriptcentral.com/braincom
0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120
54
55 Month Month
A
i 30
ii
1
80
2
3
24
MMSE
60
4
ADAS
5
18
40
7
CN
12
8 EMCI
20
9 LMCI
10 AD
6

11
12 0 12 24 36 48 60 72 0 12 24 36 48 60 72 84 96
B
13 Month Month
14 i ii
15
30
16
80
17
24
18
MMSE
60
ADAS
19
18
20
40
21
12
22 ApoE -/-
23 ApoE +/- 20
24 ApoE +/+
6
25
0
26 0 12 24 36 48 60 72 84 0 12 24 36 48 60 72 84 96
C
27
28
Month Month
29 i ii
50
30
30
31
40
32
27
MMSE
33
ADAS
30
34
24
35
36
20
37
21
Male
38 Female
10
39
40 0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120
41
D
42
Month Month
43 i ii
30
44
50
45
46
27
40
47
MMSE
ADAS
48
30
24
49
50
Primary
20
21
51 Secondary
52 Tertiary
Post-graduate
10
53
18
54
0 12 24 36 48 60 72 84 96 108 120 0 12 24 36 48 60 72 84 96 108 120
55
56 Month Month
57
58
59
60
Graphical abstract
The therapeutic potential of non-steroidal anti-inflammatories (NSAIDs) in Alzheimer’s disease is a
contentious issue due to the incongruence between epidemiological, preclinical and clinical
research. Using innovative statistical methods, we demonstrate that while most NSAIDs did not slow
the progression of cognitive impairment, diclofenac dramatically improved cognitive outcomes.

Rivers Auty2020

Uploaded by

Copyright:

Available Formats

Rivers Auty2020

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Rivers Auty2020

Uploaded by

Copyright:

Available Formats

Anti-inflammatories in Alzheimer’s disease - potential therapy or spurious correlate?

Short title: NSAIDs in Alzheimer’s disease

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

6School of Psychology, University of New South Wales, Sydney, NSW, Australia

Search terms: Alzheimer’s disease, NSAID, inflammation, Cognitive decline

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Epidemiological evidence suggests non-steroidal anti-inflammatories (NSAIDs) reduce the risk of

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Building of initial main effect model

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Inclusion of biologically relevant two-way interaction terms

Final models for MMSE and ADNI

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Assumption check of residuals

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Kanai M, Shizuka M, Urakami K, Matsubara E, Harigaya Y, Okamoto K, et al., Apolipoprotein E4

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

R Core Team -, R: A language and environment for statistical computing. R

Foundation for Statistical Computing, Vienna, Austria. URL https://www.R-project.org/.

Scharf S, Mander A, Ugoni A, Vajda F, Christophidis N, A double-blind, placebo-controlled trial of

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

The authors have no competing interests to report.

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

Downloaded from https://academic.oup.com/braincomms/advance-article/doi/10.1093/braincomms/fcaa109/5876020 by guest on 05 October 2020

You might also like