Published on 19 September 2026
An apparent 90% drop in dementia treatment after antiviral use: what should we make of an effect this large?
The essentials
In a Dutch community-pharmacy dispensing database covering thirty years of collection, from 1994 to 2024, 23,887 people who had received an anti-herpetic antiviral were matched to 238,870 unexposed people and followed for three years for the possible initiation of an anti-dementia treatment. The reported hazard ratio is 0.09, 95% CI 0.07 to 0.13, an apparent 90% reduction. This figure does not read as a result, it reads as a signal. A brief antiviral exposure does not produce an effect of this size against a slow-onset, multifactorial disease. The exposed group is also more comorbid than the control group across nearly every measured drug class, while showing far fewer events, an inconsistency that points toward selection or surveillance bias. The outcome is, moreover, a surrogate marker, treatment initiation, not a diagnosis. The authors call for randomised trials, yet they also write that it is “highly unlikely” that bias or imprecision explains these findings. This analysis does not follow them on that point. No prescribing consequence follows from it.
Context
The infectious hypothesis of Alzheimer’s disease is old and regularly revived. It holds that reactivation of a herpesvirus within the central nervous system contributes to the lesion cascade. Two epidemiological studies identifiable in this publication’s bibliography reported an association between antiviral exposure and dementia risk, one Taiwanese, one Swedish, with very different magnitudes. The authors also cite a more moderate Korean estimate, but the reference they attach to it concerns proton pump inhibitors, not antivirals.
The question sometimes reaches the consulting room by another route. A patient or relative has read that an ordinary drug, prescribed for shingles or a cold sore, might protect against cognitive decline. It helps to know how to answer, and for that, to understand why an association of this size is itself a reason for caution.
The study at a glance
| Population | |
| 262,757 adults aged 50 to 80 drawn from the IADB.nl community pharmacy dispensing database of the University of Groningen, approximately 125 pharmacies in the northern Netherlands, period 1994 to 2024. Excluded were people who had received an anti-dementia drug in the year before the index date, those treated with an immunosuppressant or anticancer agent, and those with less than one year of follow-up. Mean age 65, standard deviation 9, 59% women in both groups. | |
| Exposure | |
| At least one dispensing of acyclovir, valacyclovir or famciclovir. The index date is that of this first dispensing. Only initial treatment episodes were counted. The source sets neither a minimum exposure duration nor a latency window between exposure and the start of event counting. 23,887 people exposed, of whom 18,539 received valacyclovir, 4,159 acyclovir and 1,189 famciclovir. | |
| Comparator | |
| 238,870 people with no dispensing of any anti-herpetic antiviral over the period, matched in a ratio of up to one to ten on age, sex and calendar date. | |
| Outcome | |
| Initiation of an anti-dementia treatment, defined as at least two dispensings of donepezil, rivastigmine, galantamine or memantine within one year. This is a surrogate marker of diagnosis, not a diagnosis itself. | |
| Design | |
| Retrospective incidence cohort on a dispensing database, with a three-year follow-up window. Incidence rates per 1,000 person-years, Kaplan-Meier curves and log-rank tests, multivariable Cox models adjusted for age, sex and covariates differing at baseline, time-varying Cox models where the proportional hazards assumption was violated. Analyses stratified over two periods, 1994 to 2018 and 2019 to 2024. Oxford level of evidence 3b. |
Quality control
| Point checked | Judgement |
|---|---|
| Size of the database | Considerable |
| FindingMore than 260,000 people and thirty years of collection. Dispensing data do not depend on patient recall, which rules out recall bias. | |
| Individual follow-up duration | Three years only |
| FindingThe thirty years belong to the database, not to the follow-up. Each participant is analysed over a three-year window, with a median follow-up of 946 days among the unexposed and 1,064 days among the exposed. That is short against the natural history of Alzheimer’s disease. | |
| Plausibility of the magnitude | Implausible |
| FindingFor a typically brief drug exposure, a 90% reduction in the risk of starting an anti-dementia treatment points toward a design artefact rather than a pharmacological effect. The authors themselves note that most herpes reactivations are self-limiting and that antivirals are often reserved for severe forms. | |
| Internal consistency | Comorbidity paradox |
| FindingAt baseline, the exposed group more often receives antidiabetics, antihypertensives, cholesterol-lowering drugs, antidepressants, anticholinergics, benzodiazepines and antiepileptics, and yet shows six times fewer events. Only anti-Parkinson’s drugs do not differ between the two groups. A factor linked both to the likelihood of receiving an antiviral and to that of being identified as demented would better explain this pattern than a protective effect. | |
| Nature of the outcome | Surrogate marker |
| FindingStarting an anti-dementia treatment is not the same as receiving a diagnosis. Some affected people are never treated. The authors further note that rivastigmine is also prescribed for Parkinson’s disease dementia and for the motor symptoms of Parkinson’s disease. | |
| Unmeasured confounding | Substantial |
| FindingAPOE genotype, education level, socioeconomic status, lifestyle and baseline cognitive status do not appear in a dispensing database. The database contains no cognitive assessment, no imaging and no biomarker. The authors themselves list these gaps. | |
| Molecule-specific analyses | Small numbers |
| FindingOver three years, the acyclovir group counted 6 events and the famciclovir group only 1. In the second period, no event occurred among the 801 acyclovir users or the 159 famciclovir users, which precludes any relative-risk calculation. These estimates should not be read in isolation. | |
| Model covariates | Instructive |
| FindingThe model also reports the associations of other drug classes, including a hazard ratio of 1.84 for anti-Parkinson’s drugs. These values are useful: they show that the database captures indication-based associations, which invites the same caution when interpreting the main result. | |
Results
| Result | What the source reports |
|---|---|
| Any exposure | Hazard ratio 0.09 (0.07 to 0.13), p<0.001 |
| FindingMain estimate from the Cox model adjusted for age, sex and covariates differing significantly at baseline. Over the whole follow-up, 2,495 people started an anti-dementia treatment, 2,455 in the unexposed group and 40 in the exposed group. | |
| By molecule | Valacyclovir 0.10, acyclovir 0.09, famciclovir 0.07 |
| FindingThe three estimates are of the same order, and the authors note that their confidence intervals overlap considerably. This uniformity is itself questionable: a dose- or molecule-dependent relationship would be expected from a genuine pharmacological effect, all the more so since the authors elsewhere invoke differences in bioavailability and central nervous system penetration among these three molecules. | |
| By period | 1994 to 2018: 0.14 (0.09 to 0.20). 2019 to 2024: 0.05 (0.03 to 0.10) |
| FindingThe association strengthens in the more recent period. The authors see this as noteworthy and propose several explanations, including earlier diagnosis of dementia. A rise in detection in the unexposed group alone would produce the same shift, with no pharmacological effect involved. | |
| Raw incidence rates | 0.69 versus 4.96 per 1,000 person-years |
| FindingAt three years, 0.17% of the exposed and 1.03% of the unexposed had started an anti-dementia treatment. The authors derive from this an absolute risk reduction of 0.86 percentage points and a number needed to treat of 116 to prevent one event. This calculation only holds if the association is causal, a condition not met here. | |
| Other classes in the model | Anti-Parkinson’s drugs 1.84 (1.65 to 2.05), antidepressants 1.34 (1.18 to 1.52), antiepileptics 0.72 (0.54 to 0.95) |
| FindingThe opposite direction of the association with anti-Parkinson’s drugs illustrates confounding by indication, and the authors interpret it that way themselves: parkinsonian syndromes often precede cognitive decline. The same reasoning applies to antidepressants. The other classes, antidiabetics, antihypertensives, cholesterol-lowering drugs, anticoagulants, anticholinergics and benzodiazepines, show no significant association. | |
| International comparison | Taiwan 0.09 (0.08 to 0.11), Sweden 0.89 (0.86 to 0.92) |
| FindingThese two earlier studies on antiviral exposure range from a very strong association to a reduction on the order of 11%. This spread is more compatible with differences in database design than with a stable pharmacological effect. The third estimate invoked by the authors, a Korean series at 0.76 and 0.79, refers in their bibliography to a study on proton pump inhibitors, and cannot be counted as a result concerning antivirals. | |
Critical appraisal
| Domain | Risk of bias |
|---|---|
| Confounding | Critical |
| FindingConfounding by indication and a healthy-user effect. Receiving an antiviral implies a consultation, a diagnosis and adherence, three markers of engagement with healthcare that themselves influence the later detection of dementia. | |
| Selection | High |
| FindingThe inconsistency between comorbidity burden and event incidence is the main warning sign. It suggests the two groups are not comparable on unmeasured dimensions. | |
| Outcome classification | Possible differential |
| FindingIf the two groups do not have the same probability of reaching a memory consultation and therefore a prescription, the outcome is measured differently on each side. The authors mention this possibility in one sentence, without a dedicated analysis. | |
| Follow-up time | Not explored |
| FindingThe index date is that of the first antiviral dispensing, which limits classic immortal time bias. However, the requirement of one year of registration before and after the index date, and the exclusion of follow-up shorter than one year, condition inclusion on surviving the first year. Median follow-up is longer among the exposed than the unexposed, 1,064 versus 946 days. No sensitivity analysis on these choices is reported. | |
| Adequacy of conclusions | Overreaching |
| FindingThe authors use the vocabulary of association and call for randomised trials, yet they also present matching and adjustment as strengthening the validity of causal inference, describe their data as strong evidence of a protective association, and conclude that it is highly unlikely that bias or imprecision explains these results. On dispensing data with no diagnosis attached, that last claim is not sustainable. This analysis is therefore more severe than the authors themselves on the scope of the figure. | |
Level of evidence
Oxford level of evidence 3b, retrospective cohort on a dispensing database. Confidence is high that a statistical association exists within this dataset, the size of the collection leaving little room for chance. It is very low regarding the causal interpretation of that association, and the observed magnitude is precisely what makes it suspect: an estimate can be precise and still carry very low credibility. The uncertainty here concerns first the very existence of an effect, not only its magnitude. Nothing in this work, however, rules out a real effect of smaller size, one that would still need to be demonstrated by a trial.
The colleague test
What an experienced colleague would say if shown this study in two minutes, between two consultations.
A ninety percent reduction, over three years of follow-up, on an outcome that is a prescription and not a diagnosis, deserves scrutiny before belief. And look: the exposed patients are sicker and develop less dementia. That does not describe a protective drug, it describes two groups that could not be made comparable. The subject stays worth pursuing, the herpes hypothesis is not absurd, but this particular figure is not one to quote to a patient.
Translation for practice: do not start or adjust an antiviral prescription with a cognitive-prevention intention. What to remember is a reading rule, an extraordinary effect in an observational design calls first for the bias that produced it.
What you can do with this
- Do not modify or start an antiviral prescription on the basis of an Alzheimer’s disease prevention objective. This publication is an observational study on dispensing data and establishes no preventive indication.
- Know how to answer a patient who brings in the article: the association exists in the data, and its size is exactly why it cannot be read as an effect.
- Use this case as a reading benchmark: in an observational study, an exposed group that is sicker yet less affected is a signal of bias, not good news.
- Distinguish treatment initiation from diagnosis itself when reading a study built on dispensing or reimbursement databases, in any country.
- Follow the infectious hypothesis for what it is, an active research lead whose answer will come from randomised trials, which the authors themselves call for, not from dispensing databases.
Frequently asked questions
Should an antiviral be offered to a patient worried about their memory?
No. No randomised-trial data support this practice, and the authors of this study themselves write that confirmation through randomised trials is needed to provide definitive evidence.
Does this critical reading refute the herpes hypothesis of Alzheimer’s disease?
No, and that would be a symmetrical error. A study that establishes nothing does not refute anything either. The hypothesis remains open, it simply calls for a different kind of evidence.
Why isn’t a very large sample enough?
Because size corrects for chance, not for bias. An unmeasured confounder acts the same way on three hundred people and on three hundred thousand, it is simply estimated with greater precision.
Is the result consistent with findings from other countries?
The two earlier studies on antiviral exposure identifiable in this publication’s bibliography range from a hazard ratio of 0.09 in Taiwan to 0.89 in Sweden. This spread is itself informative: it suggests that each database’s design weighs more heavily than the drug under study.
Annotated bibliography
Source study. Alghamdi A, Balafas S, Bos JHJ, van Munster BC, Rafie K, Dolga AM, Hak E. Association between the use of anti-herpetic drugs and subsequent initiation of Alzheimer’s disease drug treatment: Dutch population-based inception cohort study. BMJ Open. 2026 Jun 15 ; 16 (6) : e114033. DOI 10.1136/bmjopen-2025-114033. PMID 42297456. PMCID PMC13289104. Funding: the authors declare no specific grant for this work from any public, commercial or not-for-profit funding agency. Competing interests: the document consulted contains no statement of competing interests. Article not commissioned, externally peer reviewed. Ethics committee approval judged neither needed nor required, the data being collected in accordance with the General Data Protection Regulation. No supplementary material accompanied the document consulted: the supplementary figures and tables of this publication were not reviewed, and no data found only there is reproduced here.
Context references, antiviral exposure. Tzeng NS, Chung CH, Lin FH, et al. Anti-herpetic medications and reduced risk of dementia in patients with herpes simplex virus infections, a nationwide, population-based cohort study in Taiwan. Neurotherapeutics. 2018 ; 15 : 417 to 429. DOI 10.1007/s13311-018-0611-x. PMID 29488144. This is reference 17 of the source publication, the one from which the adjusted hazard ratio of 0.09 (0.08 to 0.11) is drawn. Lopatko Lindman K, Hemmingsson E, Weidung B, et al. Herpesvirus infections, antiviral treatment, and the risk of dementia, a registry-based cohort study in Sweden. Alzheimers Dement (N Y). 2021 ; 7 : e12119. DOI 10.1002/trc2.12119. PMID 33614892. This is reference 19, the one from which the hazard ratio of 0.89 (0.86 to 0.92) is drawn.
Context references, attributions to verify. The source publication attributes to a Korean nationwide cohort adjusted hazard ratios of 0.76 to 0.79 for antiviral exposure, referring to its reference 15: Choi HG, Kim JH, Kim JH, et al. Associations between proton pump inhibitors and Alzheimer’s disease, a nested case-control study using a Korean nationwide health screening cohort. Alzheimers Res Ther. 2022 ; 14 : 91. DOI 10.1186/s13195-022-01032-5. PMID 35773740. This reference concerns proton pump inhibitors and is a nested case-control study, not an antiviral exposure cohort. Likewise, the introduction lists among the cohorts reporting a protective antiviral effect an Italian study, its reference 18: Costa AS, Agostini S, Guerini FR, et al. Modulation of immune responses to herpes simplex virus type 1 by IFNL3 and IRF7 polymorphisms, a study in Alzheimer’s disease. J Alzheimers Dis. 2017 ; 60 : 1055 to 1063. DOI 10.3233/JAD-170520. PMID 28984602. This reference concerns genetic polymorphisms, not an antiviral treatment. Neither attribution is confirmed by the reference cited.
Editorial collections
Tags
Verified on September 1, 2026 against the full text of the publication and its supplementary material where available. This analysis underwent an independent double reading. The English version was checked for conformity on September 19, 2026, against the figures of the French version and against the source. How we verify what we publish
This analysis is intended for healthcare professionals. It does not constitute a prescribing recommendation and does not replace individual clinical judgment.
Analysis from Psychiatry Evidence Base, evidence-based psychiatry, explained with rigor.