We analyzed data from the University of Missouri Health Care Research Data Lake, a repository encompassing approximately 2.2 million patients with clinical encounters recorded between January 1, 2010, and December 31, 2024. The Data Lake integrates structured inpatient, outpatient, emergency department, and specialty clinic records standardized to the OMOP Common Data Model (CDM) version 5.0, enabling consistent phenotype ascertainment across heterogeneous clinical settings [63]. Cohort definitions were implemented in ATLAS version 2.12 (Observational Health Data Sciences and Informatics [OHDSI]) [64,65] and executed in R version 4.4.1 (R Foundation for Statistical Computing) using the OHDSI packages CohortMethod version 4.2.3 [66,67], FeatureExtraction version 3.7.2 [68], and EmpiricalCalibration version 3.1.3 [69,70], together with supporting dependencies. The analytic package (ATLAS JSON cohort definitions, concept sets, executable R scripts, model objects) is archived in a version-controlled public repository.

We conducted a retrospective cohort study to evaluate whether a recorded HSV-1 diagnosis is associated with subsequent incident cognitive impairment or dementia. We used a new-user, incident-exposure cohort design in which both cohorts entered at incident eligibility, had a clean baseline period, and followed parallel exclusion rules. We do not describe this as an active-comparator design because, in pharmacoepidemiology, an active comparator typically includes new users of alternative treatments for the same indication, helping control confounding by indication and healthy-user or frailty bias at the design stage [71]. Comparator entry was anchored to a health-system encounter, defined by an observation period and any recorded condition occurrence. This ensured that comparator members were health-system users with measurable baseline covariates and a defined time zero, rather than untreated general-population members. Thus, the design aligned data provenance, eligibility, and health-system engagement across groups, while control of confounding by indication and related host factors relied on the high-dimensional PS approach, as discussed in the Limitations section. Eligible observation time spanned 2010-01-01 through 2024-12-31. Each participant was required to have at least 365 days of continuous observation prior to cohort entry (baseline wash-in). Follow-up for outcome assessment began 1 day after the cohort index date and continued until the earliest of (1) qualifying outcome, (2) end of continuous observation in the Data Lake, or (3) 2024-12-31 (administrative study end). This approach ensured that follow-up periods were standardized between groups, independent of subsequent diagnostic activity. A schematic of the cohort construction and time-at-risk is provided in Multimedia Appendix 1 (Figure A: analytic workflow diagram and Figure B: cohort logic flow diagram).

The primary endpoint was incident all-cause cognitive impairment or dementia, ascertained using a curated OMOP concept set of more than 120 ICD-9-CM and ICD-10-CM codes spanning mild cognitive impairment, ADRD (vascular, frontotemporal, Lewy body), amnestic syndromes, and nonspecific cognitive deficit codes; the full listing is provided in Multimedia Appendix 5. Individuals with any code from this set during the 365-day baseline were excluded from both cohorts. The earliest qualifying postindex diagnosis defined the outcome event; both inpatient and outpatient settings were permitted.

"Covariates for PS modeling were identified using a high-dimensional feature extraction process implemented through the FeatureExtraction package version 3.7.2 (Observational Health Data Sciences and Informatics [OHDSI], coordinating center at Columbia University, New York, NY, United States) [68].". Approximately 17,000 candidate baseline covariates were considered, capturing clinical features from two distinct observational windows prior to the index date: a long-term period (day −365 through day −31) and a short-term period immediately preceding cohort entry (day −30 through day −1). These covariates represented 5 broad domains, including medication use, clinical procedures, condition history, aggregated clinical measurements, and indicators of health care usage. In addition, patient demographic characteristics encompassing age categories, sex, race, ethnicity, and calendar timing (index month and year) were explicitly included. To prevent redundancy within the large covariate set, intermediate clinical measurement windows and standardized composite indices such as Charlson and CHADS2 scores were deliberately excluded. Rare covariates were retained without imposing frequency thresholds, and regularized logistic regression was utilized to handle sparse data. No imputation was performed for missing values; race or ethnicity data classified as “unknown” or “not recorded” in the source records were retained as distinct categories. Continuous covariates were converted to categorical or binary indicator variables per default FeatureExtraction processing methods.

A PS for HSV-1 exposure was estimated using regularized logistic regression implemented in CohortMethod with a Laplace (L1) prior with variance 1 and tuned by ten-fold cross-validation. All 17,154 candidate baseline covariates were used in the model; 826 retained nonzero coefficients after shrinkage. Coefficients ranged from −6.56 to +2.62 (median −0.006; IQR −0.12 to +0.11). The full coefficient vector with OMOP covariate identifiers is available in Multimedia Appendix 6 to permit external reconstruction of the PS.

Individuals were divided into four PS strata using CohortMethod defaults (strata defined to contain equal numbers of target persons). Stratum weights were applied in downstream effect estimation; no trimming was used in the primary analysis. Baseline balance was evaluated using the CohortMethod and covariateBalance output. For each of the 17,154 covariates we computed prevalence (or mean for continuous measures) by cohort and the absolute standardized mean difference (|SMD|) before and after PS stratification. Residual imbalance was considered acceptable if |SMD| was less than 0.10. Summary diagnostics appear in Figure 1 (domain-stratified Love plot) and Figure 2 (covariate-level balance scatterplot).

Eligibility was applied sequentially using the criteria described above. The composition of the resulting analytic cohorts is summarized in Table 1 (attrition table) and illustrated in Figure 3 (cohort flow diagram). Person-time and outcome counts are reported in the Results section.

The primary association between HSV-1 diagnosis and incident cognitive impairment or dementia was estimated using a Cox proportional hazards model stratified by PS stratum. Time zero was index date plus 1 day. Individuals were followed until outcome, end of continuous observation, or 31 December 2024, whichever occurred first. Hazard ratios with 95% CIs are reported in the Results section.

We prespecified 271 negative-control outcomes (conditions with no plausible causal relationship to HSV-1, eg, psoriasis, osteoarthritis; complete list in Multimedia Appendix 7) to assess residual systematic error. Each control outcome was analyzed through the same PS-stratified Cox workflow used for the primary endpoint. Empirical null distributions were to be estimated from these controls to calibrate effect estimates following the approach of Schuemie et al [69,70]; the results of this calibration step are reported in the Results section.

Minimum detectable relative risk was calculated using CohortMethod power diagnostics, assuming 80% power and two-sided α=.05. The calculation was applied to observed person-time and event counts from the PS-adjusted analytic sample; these values are reported in the Results section in Table 2.

Analyses were conducted in R version 4.4.1 (R Foundation for Statistical Computing) using the open-source packages CohortMethod version 4.2.3 [66,67], FeatureExtraction version 3.7.2 [68], and EmpiricalCalibration version 3.1.3 [69,70], all developed and maintained by Observational Health Data Sciences and Informatics (OHDSI), together with related dependencies. All ATLAS (developed by OHDSI) JSON cohort definitions, concept set files, analysis settings, and executable code required to reproduce the results are available in a public GitHub (Microsoft Corporation) repository [72].

This secondary analysis of deidentified University of Missouri (MU) Health system EHR data was reviewed by the University of Missouri Institutional Review Board and determined to be non-human subjects research (IRB #2114987 MU); informed consent was waived. Analyses were conducted within a secure University of Missouri analytic environment compliant with (Health Insurance Portability and Accountability Act) HIPAA and institutional data governance policies. This study adheres to the (Strengthening the Reporting of Observational Studies in Epidemiology) STROBE guidelines; the completed checklist is provided in Checklist 1.

We identified 6900 candidate HSV-1 index events and 450,628 potential comparator candidates in the University of Missouri Research Data Lake for the 2010‐2024 study window. After applying the continuous prior observation requirement (at least 365 days), exclusion criteria (HIV, specified other neurotropic infections, transplant), and the 365-day baseline washout for prior cognitive diagnoses, the analytic cohorts included 6,274 HSV-1 and 379,975 non-HSV-1 (no recorded HSV-1 diagnosis) individuals (about 60:1 comparator to HSV-1). Baseline attrition was 9.1% in the HSV-1 cohort and 15.7% in the comparator cohort.

The distribution of follow-up times was right-skewed in both cohorts, reflecting a minority of individuals with extended observation periods. Median follow-up was 1111 days (IQR 413‐2,631; 90th percentile 2924; maximum 5460) in HSV-1 participants and 956 days (interquartile range 327‐3165; 90th percentile 3708; maximum 5478) in non-HSV-1 participants (Table 3). These distributions informed the prespecified administrative censoring horizons evaluated in sensitivity analyses (1 year, 5 years, and 10 years).

Among participants retained in the analytic sample, the HSV-1 cohort contributed 23,186 PY and 469 incident all-cause cognitive impairment events; the non-HSV-1 cohort contributed 1,519,827 person-years and 24,764 events. Given these observed rates and sample sizes, the study achieved 80% power to detect HRs of 1.15 or greater (two-sided α=.05). Crude incidence rates are shown in Table 2.

Unweighted crude incidence of the primary outcome was 20.23 per 1000 PY (95% CI 18.44‐22.14) in HSV-1 and 16.29 per 1000 person-years (95% CI 16.09‐16.50) in non-HSV-1 participants. The absolute rate difference was about 3.94 events per 1000 PY (one additional event for every approximately 254 PY of follow-up among HSV-1 participants compared with non-HSV-1). These crude findings were consistent in direction with the adjusted HR.

Before PS adjustment, the HSV-1 cohort had a higher proportion assigned female at birth (Table 4) and a lower proportion aged 0‐4 years than the comparator cohort, with 72.1% (4976/6900) vs 52.8% (132,122/250,000) female and 2.7% (189/6900) vs 11.3% (28,370/250,000) aged 0‐4 years. Four-stratum PS weighting reduced these imbalances. With an absolute |SMD| threshold of less than 0.10, all demographics displayed in Table 4 met balance after weighting except age 0‐4 years (|SMD| 0.18) and race (Black |SMD| 0.11; White |SMD| −0.13). These residual differences were addressed in a doubly robust sensitivity model. Global balance across approximately 17,000 covariates is shown in the domain-stratified balance plot (Figure 1), and nearly all postweighting covariates had |SMD| less than 0.10. Baseline characteristics before and after weighting are summarized in Table 4, and full covariate-level results are provided in Multimedia Appendix 3.

Figures 1 and 4 depict the preference score distribution & domain-stratified covariate balance plot.

Empirical overlap in the preference score distribution (min{PS, 1 - PS}) was visualized with density plots; trimming thresholds used in sensitivity analyses are marked at 0.10 and 0.90.

PS density plots (Figures 4 and 5) showed a high concentration of HSV-1 participants at higher exposure probabilities and non-HSV-1 participants at lower probabilities, with limited overlap between about 0.30 and 0.70, consistent with strong model discrimination (c-statistic approximately 0.92). The full distribution of preference scores in exposure and comparator cohorts is provided in Multimedia Appendix 8. However, trimming observations with preference scores below 0.10 or above 0.90 retained 89 % (5584/6274) of HSV-1 and 84 % (319,179/379,975) of non-HSV-1 participants and was evaluated as a sensitivity analysis.

Figures 2 and 5 show preference score density distributions by cohort and covariate balance scatterplot before vs after weighting.

In the PS-stratified Cox proportional hazards model, HSV-1 diagnosis was associated with a 14% higher hazard of incident all-cause cognitive impairment (HR 1.14; 95% CI 1.03‐1.25; P=.007).

Figures 6-8 show Kaplan-Meier survival curves, cumulative absolute risk difference, and annual survival probabilities.

Kaplan-Meier curves weighted by PS strata showed gradual divergence over follow-up, with clearer separation emerging over extended follow-up (Figure 6 for KM curves; Figure 7 for absolute risk difference; Figure 8 for annual survival probabilities). At 1-year postindex, cognitive disorder-free survival was about 97.7% in HSV-1 and 97.8% in non-HSV-1 participants, a difference of less than 0.1 percentage points. The gap widened to about 2.3 percentage points near year 8 and narrowed in late follow-up (about 1‐2 percentage points by year 10 (Figure 8); wider confidence limits because of attrition). The cumulative absolute risk difference curve (non-HSV-1 minus HSV-1) and annual survival probability plot illustrate these patterns.

The absolute risk difference curve (Figure 7) increased steadily from cohort entry to a peak of approximately 2.8 percentage points near day 3048 (approximately 8.3 years of follow-up), then declined in the final observational period. This pattern reflects primarily the depletion of the HSV-1 at-risk pool in the tail rather than a true convergence of cumulative risk. At day 3000, 539 HSV-1 participants remained at risk compared with 63,554 comparator participants; by day 3500, these counts were 143 and 45,025 respectively, a proportional depletion approximately 2.5 times greater in the HSV-1 group. This asymmetry arises because most HSV-1 individuals entered the cohort after 2010, anchored on their incident diagnosis, and the 2024 administrative end date therefore limits maximum observable follow-up disproportionately for those enrolled later. With fewer than 150 HSV-1 participants at risk beyond day 3500, the Kaplan-Meier estimator becomes unstable: individual outcome events produce large step-function changes in the HSV-1 survival estimate, and the derived absolute difference fluctuates accordingly. Additionally, individuals still event-free at more than 8 years represent a highly selected survivor population, which may attenuate the apparent relative difference independent of statistical noise. For these reasons, point estimates of the absolute risk difference beyond year 8 should be interpreted with caution; the 5- and 10-year administratively censored sensitivity analyses (Table 5) provide more stable comparisons over those defined horizons.

In the MU Data Lake extract, no postindex events were observed in either cohort for any of the 271 prespecified negative-control outcomes (0 events across all 271 outcomes in both the HSV-1 [n=6274] and non-HSV-1 [n=379,975] groups; full listing in Multimedia Appendix 7). Because no events occurred, an empirical null distribution could not be constructed, and effect calibration was not performed; all reported HRs are therefore uncalibrated. This finding likely reflects the rarity of the selected control conditions within the single-institution study window and the relatively modest size of the HSV-1 cohort, rather than true absence of these conditions in the broader population.

Sensitivity analyses demonstrated consistent findings across multiple analytic strategies. After applying preference-score trimming, excluding participants with preference scores below 0.10 or above 0.90, the adjusted HR was 1.12, (95% CI 1.01‐1.25). In a doubly adjusted model accounting for residual imbalance in the age 0‐4 years and race (White; Black or African American) covariates, the estimated HR was 1.13 (95% CI 1.02‐1.24). Administrative censoring at one-year postindex showed no measurable association (HR 1.00, 95% CI 0.84‐1.20), a finding consistent with the low absolute number of outcome events accruing within the first year of follow-up and indicating that the primary signal emerges over extended follow-up rather than in the immediate postindex period. Censoring at five-year and ten-year horizons yielded HRs consistent with the primary finding (HR 1.15, 95% CI 1.04‐1.28, and HR 1.15, 95% CI 1.04‐1.26, respectively), confirming that the association becomes statistically detectable only as cumulative follow-up time increases. Restricting the outcome definition to dementia attributable to Alzheimer disease produced a HR identical to the primary estimate (HR 1.14, 95% CI 1.03‐1.25), reinforcing the stability of findings across these analytic scenarios.

Among 386,249 participants followed for up to 15 years, individuals with a clinically documented HSV-1 diagnosis experienced a 14% higher hazard of incident cognitive impairment or dementia than comparator participants without coded HSV-1 diagnoses (HR 1.14, 95% CI 1.03‐1.25, P=.007). Throughout this analysis, “HSV-1 exposure” refers specifically to a coded HSV-1 diagnosis in structured EHR data, which most plausibly reflects symptomatic or clinically recognized disease rather than latent or asymptomatic infection; the clinical phenotype captured is therefore a subset of the larger infected population identifiable only by serology. The absolute incidence was approximately 4 events per 1000 PY higher in the HSV-1 cohort (Table 2). PS-weighted survival curves showed gradual divergence that became progressively more apparent and sustained across extended follow-up (Figures 6-8). This pattern is consistent with the sensitivity analyses, in which the 1-year administratively censored model yielded no measurable association, whereas the 5-year and 10-year censored models produced HRs consistent with the primary estimate, together indicating that the association accumulates over extended follow-up rather than being concentrated in the immediate postindex period. The high seroprevalence of HSV-1 in the general adult population means that, if the observed association reflects a causal relationship, even a 14% increase in individual hazard could translate to a meaningful population-attributable increment in dementia cases, a point elaborated in the implications section below.

Prior epidemiological investigations have reported heterogeneous results, likely reflecting differences in case definitions, study design, and confounding adjustment. A nested case-control study using Swedish registry data reported elevated odds of AD in individuals with prior HSV-1 infection, suggesting a temporal link between infection and diagnosis [61]. A Taiwan-based retrospective cohort study using national claims data found increased dementia incidence among patients with HSV diagnoses, though that study combined HSV-1 and HSV-2 under a single code [17]. Our HR is directionally concordant but smaller, possibly reflecting our narrower HSV-1 concept set, finer confounder adjustment via high-dimensional PS stratification, and the single health-system EHR source [61,73]. A parallel analysis by Chen et al [59] examining antiviral treatment in herpetic infections suggested a protective association, but that study did not isolate HSV-1 and did not use PS methods [73].

The biological plausibility of the observed association is supported by mechanistic evidence. Experimental models show that HSV-1 can induce tau hyperphosphorylation and amyloid-beta aggregation in neuronal cultures [11,74,75], disrupt tau exon splicing in human neuron preparations [76,77], and accelerate microglial activation in murine AD models [78,79]. A prospective study within the Multidomain Alzheimer Preventive Trial cohort detected HSV-1 DNA in peripheral blood mononuclear cells and found seropositivity associated with abnormal phosphorylated tau and Abeta42/40 ratios [2,40,61]. The temporal association between HSV-1 diagnosis and cognitive impairment in our cohort is consistent with, but does not directly test, these mechanistic hypotheses; biomarker and laboratory data that would be required to evaluate them were unavailable in this study.

The magnitude of the observed association in this study (HR 1.14) is smaller than estimates from several earlier investigations, and misclassification of exposure is the most parsimonious explanation for this discrepancy. Studies reporting HRs or odds ratios of 1.5 or greater for HSV and dementia have generally used seroprevalence data or seroconversion designs [25,60], in which the exposed group includes all seropositive individuals regardless of whether they received a clinical diagnosis code. In contrast, EHR-based studies restricted to coded diagnoses, as in the present analysis, identify only a small, clinically visible fraction of the truly exposed population. The non-HSV-1 comparator group therefore includes a large number of latently infected individuals, diluting the between-group difference and attenuating the observed HR toward the null. Rather than indicating that the true association is smaller than previously reported, our estimate is best interpreted as a lower bound of the true association under the assumption of nondifferential misclassification. This interpretation is consistent with the systematic evidence gap that Hong et al [12] identified when publishing their systematic review protocol in 2025, namely that the field lacks adequately powered, methodologically rigorous longitudinal studies using structured clinical data with explicit confounding control, a gap the present analysis is specifically designed to address.

The present findings carry several specific implications for clinical practice, health system design, and future research, even in the absence of causal confirmation. Before describing these, we emphasize that a HR of 1.14, while statistically meaningful and directionally consistent across sensitivity analyses, represents a modest effect at the individual patient level and does not, in isolation, justify any change to current clinical management of HSV-1 infection or to existing cognitive-screening recommendations. The findings are intended to guide future research rather than to inform immediate changes in care. At the clinical level, the results suggest that a coded HSV-1 diagnosis, already routinely captured in structured EHR data, may serve as a readily available indicator for an elevated likelihood of subsequent cognitive impairment that clinicians could factor into decisions about cognitive screening. This does not imply that every patient with a cold sore diagnosis requires neurological surveillance; the absolute incidence difference of approximately 4 events per 1000 PY is modest. However, in older patients with additional factors associated with dementia, including advanced age, apolipoprotein E. (APOE) epsilon4 carrier status, and cardiovascular disease burden, a history of clinically documented HSV-1 infection may justify earlier cognitive screening or enrollment in prevention programs. Health systems with OMOP-standardized data could operationalize such surveillance without additional data collection.

At the population level, the estimated seroprevalence of HSV-1 exceeding 65% among adults over 50 years means that even a HR of 1.14 translates to a nontrivial population-attributable fraction of late-life cognitive impairment. If the association is causal, a meaningful proportion of late-life cognitive decline may be linked to this common, potentially modifiable exposure. Antiviral therapies that suppress HSV-1 reactivation, including acyclovir and valacyclovir, are safe, inexpensive, and widely available. Several observational studies have suggested that treated patients have lower dementia incidence than untreated patients [59,61,80], and a registry-based cohort study reported an association between antiherpetic therapy and reduced AD incidence [61]. These reports are hypothesis-generating and are subject to confounding by indication, healthy-user effects, and adherence-related bias; they do not by themselves establish that antiviral therapy modifies dementia risk. Taken together with the mechanistic and epidemiologic evidence summarized above, our findings support continued investigation into whether HSV-1 reactivation contributes meaningfully to cognitive decline and whether antiviral strategies targeting reactivation warrant formal evaluation in appropriately designed prospective studies. The present analysis does not, on its own, provide a basis for recommending antiviral prophylaxis for dementia prevention.

This study provides an analytical template, including OMOP-standardized cohort definitions, open-source ATLAS and CohortMethod code, and a replicable PS framework, that could support replication across multiple OMOP-standardized health systems prior to trial design. Prospective studies that distinguish HSV-1 from HSV-2 serologically, quantify viral reactivation frequency, capture antiviral treatment patterns, and incorporate genetic susceptibility markers, including APOE genotype and markers of immune response such as IGDCC4 [81], are needed to determine whether antiviral therapy or immunomodulation could modify dementia trajectories.

Large, multisite studies integrating laboratory-confirmed infection data, pharmacy claims, and vital status linkage will be necessary to validate these findings and to clarify whether virologic suppression is associated with reduced dementia incidence at the population level.

In a fifteen-year EHR cohort, patients with any HSV-1 diagnosis code experienced a 14% higher hazard of incident cognitive impairment, including dementia subtypes, than comparator patients with no recorded HSV-1 diagnosis (adjusted HR 1.14, 95% CI 1.03‐1.25). Absolute incidence differed by about four events per 1000 PY (Table 2), and PS-weighted survival curves diverged gradually, with separation emerging over extended follow-up (Figures 6-8). Although the estimated association is smaller than some earlier reports, it is consistent with estimates from recent longitudinal studies that applied extensive adjustment. Measurement limitations, incomplete death data, and potential residual confounding preclude causal inference; nevertheless, the consistent direction across sensitivity analyses supports a modest association within the observed data environment. A HR of 1.14 represents a modest effect on the individual level, and these findings alone do not support changes to current clinical practice for HSV-1 management or cognitive screening. Given the high prevalence of HSV-1, if the observed association is causal, even this modest relative excess could contribute meaningfully to population dementia burden. Future studies should integrate laboratory-confirmed HSV-1 typing, antiviral treatment histories, and genetic data to clarify whether virologic suppression is associated with reduced dementia incidence. If a causal relationship is confirmed, sustained antiviral suppression and future HSV-1 vaccines could become useful tools for reducing dementia incidence at the population level.