Hearing loss and loneliness in the development of impaired cognition and dementia in the US All of Us cohort: prospective analyses using electronic health records Authors: Shuai Yang, David A Sbarra, Jingyue Wu, Yann C Klimentidis Published: 2026 ABSTRACT AbstractIntroductionHearing loss (HL) and loneliness are independently associated with cognitive decline, yet the extent to which loneliness mediates the HL–cognition relationship remains uncertain. We examined prospective associations among HL, loneliness and incident dementia and impaired cognition (IC) and tested whether loneliness partially mediates the effect of HL on these outcomes.MethodsUsing the National Institutes of Health All of Us Controlled Tier V.8 dataset, we fit adjusted Cox proportional hazards models with distinct baselines for HL (first electronic health record (EHR) visit) and loneliness (social determinants of health survey). Loneliness was measured with the UCLA Loneliness Scale-8 (ULS-8) and analysed as the ULS-8 mean. We conducted four-way decomposition mediation analyses to partition total, direct and indirect effects allowing for exposure–mediator interaction. Outcomes (incident dementia and IC) were ascertained from EHRs after the relevant baseline exposure. Sensitivity analyses examined associations in racial/ethnic subgroups, removed the ≥1 year EHR follow-up requirement, accounted for death as a competing risk and used time-varying HL in Cox models. Covariates included age, sex, race, education, income, smoking and alcohol use.ResultsThe EHR observation base included 317 108 adults (mean overall follow-up, 11.8 years); 50 849 (16.0%) had EHR-documented HL and 130 266 (41.1%) had a scoreable ULS-8 loneliness measure. We observed 5663 incident dementia and 21 117 incident IC events. HL was associated with higher risk of incident dementia (HR=1.38, p<0.001) and IC (HR=1.67, p<0.001). Loneliness (per 1-point increase in ULS-8 mean) was associated with higher incident dementia (HR=1.41, p=0.004) and IC (HR=1.48, p<0.001) risk. Mediation analyses indicated small indirect effects of HL via loneliness. The proportion mediated was approximately 2.5% for dementia (pm=0.025, p=0.081) and 2.0% for IC (pm=0.020, p<0.001). Subgroup analyses suggested stronger HL–dementia associations among black/African American participants and stronger HL–IC associations among Asian and black/African American participants, while results from sensitivity analyses were largely confirmatory.ConclusionsIn this large, diverse cohort, HL and loneliness were independently associated with incident dementia and IC. Loneliness mediated only a small fraction of the HL effect, suggesting that most of the observed HL–cognitive outcome association likely operates through pathways other than loneliness. These findings support further evaluation of hearing rehabilitation and social-connection supports as complementary strategies for cognitive health, while suggesting that addressing HL directly may yield broader benefits than targeting loneliness alone. WHAT IS ALREADY KNOWN ON THIS TOPIC While hearing loss and loneliness are known independent risk factors for dementia, it remains unclear to what extent hearing loss drives cognitive decline specifically by causing social isolation and loneliness. While hearing loss and loneliness are known independent risk factors for dementia, it remains unclear to what extent hearing loss drives cognitive decline specifically by causing social isolation and loneliness. WHAT THIS STUDY ADDS Data from a large, diverse US cohort show that both hearing loss and loneliness independently predict incident dementia and impaired cognition, but mediation analyses demonstrate that loneliness explains only a very small fraction (roughly 2.0–2.5%) of the hearing loss effect. Data from a large, diverse US cohort show that both hearing loss and loneliness independently predict incident dementia and impaired cognition, but mediation analyses demonstrate that loneliness explains only a very small fraction (roughly 2.0–2.5%) of the hearing loss effect. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY Because hearing loss appears to be associated with cognition largely through pathways other than loneliness, healthcare policies and practices should evaluate scalable approaches to hearing assessment and rehabilitation—such as early screening and hearing device uptake—alongside social connection resources to support cognitive health. Because hearing loss appears to be associated with cognition largely through pathways other than loneliness, healthcare policies and practices should evaluate scalable approaches to hearing assessment and rehabilitation—such as early screening and hearing device uptake—alongside social connection resources to support cognitive health. Introduction Dementia imposes profound personal and societal costs, and among modifiable midlife risks, hearing loss (HL) stands out as a leading and actionable target. The 2024 Lancet Commission on dementia prevention estimated that untreated midlife HL may account for about 7% of dementia cases worldwide—a greater proportion than other established risk factors like hypertension or obesity.1 Several mechanisms are proposed to explain how hearing impairment might contribute to cognitive decline and dementia, including increased cognitive load (ie, the brain needs to work harder to process degraded auditory input, diverting resources from memory and thinking), accelerated brain atrophy due to reduced sensory stimulation and social disengagement.2–5 In particular, HL may contribute to loneliness or social isolation, as difficulty hearing can prompt withdrawal from conversations and activities, reducing cognitive stimulation and resilience. Consistent with this pathway, observational studies show that HL is associated with higher odds of loneliness and social isolation in older adults.5 Population-based studies further support this association: greater HL is linked to higher odds of social isolation in older women,6 and in a nationally representative US cohort, hearing impairment was associated with a 28% greater odds of social isolation over 8 years.7 Medicare beneficiaries with functional HL have higher odds of limited social engagement.8 Together, these findings suggest that social disconnection is a plausible pathway linking HL with cognitive decline and overall brain health. Loneliness, the subjective experience of social isolation, is widely recognised as an important psychosocial factor impacting cognitive health.910 A recent meta-analysis of longitudinal studies (n>600 000) found that feeling lonely is associated with about a 31% higher risk of developing dementia.11 Loneliness is also linked to faster cognitive decline and broader adverse health outcomes11–15 and appears to influence dementia risk independently of objective social isolation,16 suggesting that the emotional experience of disconnection may have a distinct impact on the brain. These lines of evidence support a plausible interplay between HL and loneliness in the pathogenesis of dementia. Hearing impairment may reduce social engagement and precipitate or exacerbate loneliness, which in turn could be one pathway through which HL influences cognitive decline—a ‘cascade’ model in which sensory deficits contribute to social isolation and subsequently to cognitive impairment.17 Empirical evidence of this mediation pathway, however, is limited. Prior studies show that HL is associated with increased loneliness and social isolation5–8 and that both HL and loneliness independently elevate dementia risk3111618; but whether, and to what extent, loneliness may explain the HL–dementia link remains unclear. A recent systematic review found insufficient evidence that social isolation significantly mediates the association between HL and cognitive decline,18 highlighting the need for direct analyses of this question in large longitudinal datasets. The present study leverages data from the US National Institutes of Health (NIH)-funded All of Us (AoU) Research Program, a large, nationally diverse cohort, to investigate inter-relationships between HL, loneliness and dementia risk. Over 80% of AoU participants are from groups historically underrepresented in biomedical research,19 providing an opportunity to examine how these associations vary across race/ethnicity and socio-economic subgroups. Linkage to electronic health records (EHRs) for a large proportion of participants allows more precise temporal ordering of exposures and outcomes, enhancing the validity of causal inferences.19 Although our work was not preregistered and should be considered exploratory, we were guided by three aims to examine whether: (1) HL diagnosed before the loneliness survey is associated with greater subsequent loneliness; (2) HL and loneliness independently predict incident dementia and impaired cognition (IC); and (3) loneliness mediates the association between HL and these cognitive outcomes. By addressing these questions in a large, diverse EHR-based cohort, this study aims to clarify the role of HL and loneliness in late-life cognitive health and to inform potential interventions, such as hearing rehabilitation and social engagement strategies, for dementia prevention. Methods Study design and sample We conducted an observational cohort study using the Controlled Tier Dataset V.8 from the AoU Research Program, a large, diverse US volunteer cohort launched in 2018 by the NIH. AoU aims to enrol at least one million adults from diverse communities across the USA, with data from linked EHRs, surveys, physical measurements and biospecimens.19 The inferential target of this study was US adults who could plausibly be represented by adults participating in a large, diverse volunteer cohort with linked EHR and survey data. The analytical samples, however, were restricted to AoU participants with sufficient linked EHR data to ascertain HL and incident cognitive diagnoses; analyses involving loneliness were further restricted to participants who completed the Social Determinants of Health survey containing the loneliness items. Estimates should therefore be interpreted as most directly applicable to EHR-linked AoU participants meeting these analytical criteria. Generalisability to the US adult population may be limited by volunteer participation, differential EHR linkage, survey completion, healthcare utilisation and other sources of selection. Participants were followed through available EHR data for incident cognitive outcomes. We distinguished two anchors: the HL-baseline, defined as the first EHR visit (start of retrospective HL ascertainment), and the loneliness baseline (L-baseline), defined as the date of loneliness survey completion. In the primary analysis plan, we required at least 1 year of postbaseline EHR follow-up to ensure a minimal observation window for incident outcome ascertainment in the setting of heterogeneous EHR capture. Because this criterion uses information observed after baseline, we evaluated its impact in sensitivity analyses and report the number of participants affected in the study flow diagram. Patient and public involvement Patients and members of the public were not involved in the design, conduct, reporting or dissemination plans of the present secondary analysis. Measures Hearing loss EHR-documented HL was identified using standardised diagnosis codes in the AoU Researcher Workbench. These codes were organised using the Observational Medical Outcomes Partnership (OMOP) Common Data Model and mapped to Systematized Nomenclature of Medicine–Clinical Terms (SNOMED CT). We started from the ancestor concept ‘Hearing loss’ and retained clinically relevant active descendants, including sensorineural, conductive, mixed and presbycusis-related HL concepts. Participants with at least one qualifying condition record were classified as having EHR-documented HL; those without such records were classified as having no EHR-documented HL.1920 The full concept set is provided in online supplemental table S1. Similar diagnosis-based approaches have been used in prior EHR-based studies of HL, including studies using AoU data.2122 To our knowledge, however, this specific OMOP concept-set definition has not been formally validated against audiometric data within AoU. Because audiometric data were not uniformly available, this measure should be interpreted as clinically recorded HL rather than audiometrically confirmed hearing impairment. Loneliness Loneliness was assessed using the 8-item version of the UCLA Loneliness Scale (ULS-8), administered as part of the Social Determinants of Health (SDoH) survey. This validated questionnaire assesses subjective loneliness and social disconnection.2324 Participants rated each item on a four-level response scale, with higher values indicating greater loneliness after reverse coding of positively worded items. Because the SDoH survey allowed a ‘skip’ response option, skipped selections were recoded as missing. We included participants who answered at least four of the eight items and computed a ULS-8 mean score across non-missing items, yielding a continuous loneliness measure ranging from 1 to 4. The eight items and scoring rules are listed in online supplemental table S2. Cognitive outcomes: incident dementia and impaired cognition We examined two EHR-derived cognitive outcomes. Dementia was defined using active SNOMED CT descendants of the ‘Dementia’ ancestor concept. We also examined IC, a broader AoU/OMOP concept-based phenotype intended to capture clinically recorded non-dementia cognitive impairment or prominent cognitive symptoms in routine care. IC should not be interpreted as a single aetiological diagnosis, adjudicated mild cognitive impairment or prodromal dementia. Full concept lists are provided in online supplemental tables S3 and S4. For both outcomes, incident status required the first qualifying diagnosis after the relevant baseline (HL-baseline for HL analyses; L-baseline for loneliness and mediation analyses), with the event date taken as the date of first occurrence. Participants with any qualifying dementia or IC code before the relevant baseline were excluded. Follow-up ran from baseline to the earliest of first qualifying diagnosis, death or the last recorded EHR encounter. We treated the last recorded encounter as the end of observable follow-up because ascertainment of dementia and IC depended on recorded healthcare data; beyond that point, we could not assume continued outcome surveillance within the available EHR data. Covariates We included a range of covariates in our models to adjust for potential confounding and to examine subgroup effects. Age (in years) was included as a continuous variable, given its strong association with both HL and dementia incidence. For analyses with HL as the exposure, age at the first EHR visit was used; for analyses with loneliness as the exposure—and for mediation analyses—age at the SDoH assessment was used. We also tested age-squared terms and categorical age bands to verify the proportional hazards assumption, with final models using the linear term. Sex/gender was included as a categorical variable (female, male and non-binary group), as differences in both HL prevalence and dementia risk have been reported between men and women. Race/ethnicity was self-identified and categorised into mutually exclusive groups: white, black or African American, Asian, American Indian or Alaska Native, Native Hawaiian or other Pacific Islander, more than one population and Middle Eastern or North African. Adjusting for race/ethnicity allowed us to account for potential disparities in HL, loneliness and dementia incidence, as well as to control for population stratification in a broad sense. Socio-economic factors included years of education completed (treated as a continuous variable) and annual household income (self-reported in ordered categories, from <$25 000 to >$200 000). Education serves as a proxy for cognitive reserve and socio-economic status and is strongly associated with dementia risk, while income may confound the associations of interest through its influence on healthcare access, living conditions and psychosocial stressors. We also adjusted for health behaviours, including smoking status (never, former or current, and unknown) and alcohol use (never, former or current, and unknown). We considered adjusting for certain comorbidities (eg, hypertension, diabetes, depression) given their influence on dementia risk; however, many chronic conditions could lie on the causal pathway between HL and dementia (for instance, HL might lead to depression, which affects cognition). To avoid overadjustment, we did not include comorbidities in the main models. To clarify our assumed causal structure, we developed a directed acyclic graph (DAG) summarising the hypothesised relationships among the primary exposure, mediator, outcomes and relevant covariates (online supplemental figure S1). The DAG was used to inform covariate selection and to contextualise potential sources of residual confounding and differential ascertainment in this EHR-based study. Data analysis All analyses were conducted using R (V.4.4.0) on the AoU secure Researcher Workbench.19 We first examined participant characteristics stratified by outcome status (dementia vs no dementia; IC vs no IC) and summarised continuous variables as mean (SD) and categorical variables as n (%). These comparisons were descriptive and unadjusted, intended to characterise the analytical samples rather than estimate causal effects. Between-group differences were assessed using Wilcoxon rank-sum tests for continuous variables and Pearson’s χ² tests for categorical variables. To minimise case loss without ad hoc imputation, we labelled missing categorical covariates as ‘Unknown’ for tabulation but excluded ‘Unknown’ observations from the primary models, and analysed continuous covariates using available-case data. Internal consistency of the ULS-8 was assessed using Cronbach’s alpha based on item-level responses among participants with at least four completed items. We fit a multiple linear regression model with loneliness as the outcome and HL status as the key predictor, adjusting for all covariates (age, sex, race, education, income, smoking, alcohol). The linear model yields the adjusted mean difference in ULS-8 loneliness for participants with versus without HL (β_HL), establishing the exposure–mediator association needed for mediation analysis. We used Cox proportional hazards models for time to incident dementia and IC. Because HL and loneliness were ascertained at different baselines—HL from the first EHR visit and loneliness from the SDoH survey—we fit separate models for each exposure (HL→dementia, HL→IC, loneliness→dementia, loneliness→IC) to respect exposure timing, avoid mixing risk sets and maximise sample size. As sensitivity analyses, we fit time-updated Cox models treating HL as a time-varying covariate (online supplemental methods S1); loneliness remained baseline-only due to the single assessment.25 Each Cox model was adjusted for all covariates. Participants with a record of dementia or IC at or before the respective baseline times were excluded from the corresponding models. We report HRs and 95% CIs for the associations of HL (yes vs no) and of loneliness (per unit; ULS-8 mean) with the outcomes. Proportional hazards assumptions were assessed using scaled Schoenfeld residuals for variables included in the Cox models. No meaningful violations were observed for the primary exposure variables based on residual diagnostics. To determine if the associations differed by demographic or socio-economic factors, we tested interaction terms of HL and loneliness with racial groups, income and education. Two additional sensitivity analyses were conducted to evaluate the robustness of the primary Cox models. First, we repeated all primary models without requiring more than 1 year between the first and last recorded EHR visit. Second, we fit Fine-Gray competing-risk models treating death before dementia or IC as a competing event. These sensitivity models used the same exposure, follow-up and covariate definitions as the primary analyses, with preoutcome death coded as the competing event. To assess whether loneliness mediates the HL→outcome association while allowing exposure–mediator interaction, we used four-way decomposition via CMAverse (cmest) package in R.2627 HL (binary) was the exposure and loneliness (mean-centred continuous ULS-8) was the mediator. The outcomes were dementia and IC in separate analyses, and covariates matched those used in the aforementioned Cox models. We obtained HR scale estimates for the total effect, pure natural direct effect, pure natural indirect effect, mediated interaction and reference interaction, and summarised mediation by the proportion mediated (pm). All statistical tests were two-sided, with p<0.05 considered statistically significant. No formal multiple comparison correction was applied for the primary hypotheses, but results were interpreted with caution and an eye to consistency with prior research. Because covariates were included for adjustment rather than as exposures of substantive interest, their model coefficients are not interpreted in the main text. Reporting followed Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines.28 We summarise the key study design and modelling framework in figure 1, including the HL Cox, the loneliness Cox and the four-way decomposition models. Analytical workflow and model timelines for HL, loneliness and cognitive outcomes in the All of Us cohort: (A) hearing loss Cox model; (B) loneliness Cox model; and (C) four-way decomposition mediation model. Covariates: age, sex, race, education, income, smoking, alcohol | Sensitivity analyses: time-varying HL, no ≥1-year EHR follow-up requirement and Fine-Gray competing-risk models. EHR, electronic health record; HL, hearing loss; IC, impaired cognition; PH, proportional hazards; SDoH, Social Determinants of Health; ULS-8, UCLA Loneliness Scale-8. Results Participant characteristics We constructed a participant flow diagram to clarify the derivation of the revised analytical samples (online supplemental figure S2). Among 633 548 AoU participants, 514 998 had at least one recorded EHR visit, and 317 108 had more than 1 year between their first and last recorded EHR visits, forming the EHR observation base. The mean first-to-last EHR follow-up in this base was 11.8 years. After excluding 189 participants with dementia recorded on or before the first EHR visit, the incident dementia analytical sample included 316 919 participants. After excluding 518 participants with IC recorded on or before the first EHR visit, the incident IC analytical sample included 316 590 participants. In the incident dementia analytical sample, 50 815 (16.0%) had EHR-documented HL. Analyses evaluating loneliness as an exposure were restricted to participants with a baseline SDoH loneliness measure. Among the EHR observation base, 131 731 participants had at least one ULS-8 loneliness-item response; of these, 130 266 (98.9%) answered ≥4 items and were scored on the ULS-8. The ULS-8 demonstrated good internal consistency (Cronbach’s alpha=0.87). Over follow-up, we identified 5663 (1.8%) participants with incident dementia and 21 117 (6.7%) with incident IC; 2894 experienced both outcomes, either concurrently or sequentially. Missingness for key analytical variables is summarised in online supplemental table S5. To characterise potential selection into the EHR observation base, we also compared participants included in the EHR observation base with those not included (online supplemental table S6). Table 1 shows descriptive, unadjusted characteristics of participants who did versus did not develop dementia. Dementia cases were older (52.98±15.95 vs 44.02±16.64 years; p<0.001), more often male (51% vs 38%; p<0.001) and more likely to have HL (38% vs 16%; p<0.001). Mean loneliness scores were higher among cases (2.06±0.69 vs 1.90±0.65; p<0.001). Smoking was more prevalent among cases (51% vs 41% former/current smoking; p<0.001), and cases were more likely to report never drinking alcohol (13% vs 10%; p<0.001). Dementia cases more often had 12 or fewer years of education (34.3% vs 27.6%; p<0.001) and lower income, with 62.4% reporting annual household income of $42 500 or lower versus 47.9% of non-cases (p<0.001). Table 1The characteristics of participants who did versus did not develop dementia during follow-upCharacteristicNNo dementia n=311 256*Dementia n=5663*P value†Hearing loss316 919<0.001 No262 615 (84)3489 (62) Yes48 641 (16)2174 (38)Gender312 164<0.001 Female190 029 (62)2737 (49) Male115 694 (38)2808 (51) Non-binary887 (0.3)<20 Unknown4646109Race263 759<0.001 White173 793 (67)3300 (71) American Indian or Alaska Native3627 (1.4)100 (2.1) Asian8675 (3.3)75 (1.6) Black or African American57 973 (22)900 (19) Middle Eastern or North African1733 (0.7)25 (0.5) More than one population13 040 (5.0)254 (5.5) Native Hawaiian or Other Pacific Islander262 (0.1)<20 Unknown52 1531007Education years309 420<0.001 1401 (0.1)21 (0.4) 42426 (0.8)112 (2.0) 6.56467 (2.1)223 (4.1) 1017 454 (5.7)342 (6.2) 1257 232 (18.8)1189 (21.6) 1479 738 (26.2)1484 (27.0) 1671 759 (23.6)1042 (18.9) 1868 443 (22.5)1087 (19.8) Unknown7336163Income (dollar)254 140<0.001 500039 295 (15.7)778 (18.7) 17 50035 020 (14.0)906 (21.7) 30 00021 154 (8.5)443 (10.6) 42 50024 249 (9.7)476 (11.4) 62 50033 380 (13.4)527 (12.6) 87 50026 317 (10.5)321 (7.7) 125 00033 040 (13.2)382 (9.2) 175 00015 641 (6.3)149 (3.6) 225 00021 875 (8.8)187 (4.5) Unknown61 2851494Smoking307 140<0.001 Never178 977 (59)2679 (49) Former/current122 669 (41)2815 (51) Unknown9610169Drinking309 850<0.001 Never30 364 (10)702 (13) Former/current273 967 (90)4817 (87) Unknown6925144Age at first EHR visit316 91944.02 (16.64)52.98 (15.95)<0.001Loneliness ULS-8 mean score130 2201.90 (0.65)2.06 (0.69)<0.001 Unknown182 4934206*Mean (SD); n (%).†Wilcoxon rank-sum test; Pearson’s χ² test.EHR, electronic health record; ULS-8, UCLA Loneliness Scale-8. The characteristics of participants who did versus did not develop dementia during follow-up Mean (SD); n (%). Wilcoxon rank-sum test; Pearson’s χ² test. EHR, electronic health record; ULS-8, UCLA Loneliness Scale-8. Online supplemental table S7 presents analogous descriptive, unadjusted data for IC. IC cases were older (50.45±15.20 vs 43.72±16.67 years; p<0.001), had higher HL prevalence (40% vs 14%; p<0.001) and higher mean loneliness scores (2.04±0.69 vs 1.90±0.65; p<0.001). They were slightly less likely to be female (61% vs 62%; p=0.023) and more likely to be white (76% vs 66%; p<0.001). Fewer IC cases had 12 or fewer years of education (24.0% vs 28.0%; p<0.001), but a larger proportion reported income of $42 500 or lower (50.5% vs 48.0%; p<0.001). Smoking was more common among IC cases (46% vs 41% former/current smoking; p<0.001), whereas differences in drinking status were small despite statistical significance (p=0.003). The association between HL and loneliness We examined the association between EHR-documented HL and loneliness using linear regression adjusted for age, sex/gender, race/ethnicity, education, income, smoking and alcohol use (online supplemental table S8). Participants with HL reported slightly higher loneliness scores than those without HL (β on ULS-8 mean=0.035, SE=0.005, p<0.001). Because the remaining variables were included for adjustment rather than as exposures of substantive interest, we do not interpret their coefficients. Overall, HL was associated with greater loneliness, fulfilling one prerequisite for the mediation analysis and supporting the hypothesis that loneliness may lie on one pathway from HL to cognitive outcomes. HL and loneliness as predictors of dementia In Cox proportional hazards models, HL was associated with a significantly increased risk of dementia. In the fully adjusted model, HL was linked to a 38% higher hazard of dementia (HR=1.38, 95% CI 1.28 to 1.48, p<0.001). Loneliness was also associated with increased dementia risk: per 1-point increase in ULS-8 mean, HR=1.41 (95% CI 1.12 to 1.77, p=0.004), corresponding to approximately 1.04 per 1-point on the ULS-8 total. Primary exposure estimates are summarised in table 2; full covariate-adjusted model outputs are provided in online supplemental table S17 for transparency. Table 2Primary adjusted associations of HL and loneliness with incident dementiaAnalysisPrimary exposureN/eventsHR (95% CI)P valueDementia (HL baseline)HL (yes vs no)208 608/33091.38 (1.28 to 1.48)<0.001Dementia (loneliness baseline)Loneliness (ULS-8 mean)86 387/1791.41 (1.12 to 1.77)0.004Note: HRs are from Cox proportional hazards models adjusted for age, sex/gender, race/ethnicity, education, income, smoking and alcohol use. Only primary exposure estimates are shown in the main manuscript; adjustment covariate coefficients are not interpreted as substantive effects. Full covariate-adjusted model outputs are provided in online supplemental table S17Supplementary Table S17.HL, hearing loss; ULS-8, UCLA Loneliness Scale-8. Primary adjusted associations of HL and loneliness with incident dementia Note: HRs are from Cox proportional hazards models adjusted for age, sex/gender, race/ethnicity, education, income, smoking and alcohol use. Only primary exposure estimates are shown in the main manuscript; adjustment covariate coefficients are not interpreted as substantive effects. Full covariate-adjusted model outputs are provided in online supplemental table S17Supplementary Table S17. HL, hearing loss; ULS-8, UCLA Loneliness Scale-8. In interaction analyses, the association between HL and dementia was significantly modified by race but not by income and education (online supplemental table S9). Compared with white participants, the HL–dementia association was stronger among black participants (HR for interaction=1.73, p<0.001), with no significant interactions for other racial groups. Loneliness showed no clear evidence of effect modification by race/ethnicity, income or education for dementia (income interaction p=0.686; education interaction p=0.793; online supplemental table S10). HL and loneliness as predictors of IC We observed similar patterns for IC. HL was associated with a substantially increased hazard of IC (HR=1.67, 95% CI 1.61 to 1.73, p<0.001). Loneliness was also associated with higher IC risk: per 1-point increase in ULS-8 mean, HR=1.48 (95% CI 1.36 to 1.62, p<0.001), corresponding to approximately 1.05 per 1-point on the ULS-8 total. Primary exposure estimates are summarised in table 3; full covariate-adjusted model outputs are provided in online supplemental table S18 for transparency. Table 3Primary adjusted associations of HL and loneliness with incident ICAnalysisPrimary exposureN/eventsHR (95% CI)P valueIC (HL baseline)HL (yes vs no)208 348/14 0321.67 (1.61 to 1.73)<0.001IC (loneliness baseline)Loneliness (ULS-8 mean)81 592/11541.48 (1.36 to 1.62)<0.001Note: HRs are from Cox proportional hazards models adjusted for age, sex/gender, race/ethnicity, education, income, smoking and alcohol use. Only primary exposure estimates are shown in the main manuscript; adjustment covariate coefficients are not interpreted as substantive effects. Full covariate-adjusted model outputs are provided in online supplemental table S18Supplementary Table S18.HL, hearing loss; IC, impaired cognition; ULS-8, UCLA Loneliness Scale-8. Primary adjusted associations of HL and loneliness with incident IC Note: HRs are from Cox proportional hazards models adjusted for age, sex/gender, race/ethnicity, education, income, smoking and alcohol use. Only primary exposure estimates are shown in the main manuscript; adjustment covariate coefficients are not interpreted as substantive effects. Full covariate-adjusted model outputs are provided in online supplemental table S18Supplementary Table S18. HL, hearing loss; IC, impaired cognition; ULS-8, UCLA Loneliness Scale-8. Although HL was strongly associated with IC, the magnitude of this association varied by race: compared with white participants, it was larger among Asian (interaction HR=1.44, p=0.010) and black participants (interaction HR=1.51, p<0.001), with no significant interactions for income or education (p=0.698 and p=0.229, respectively; online supplemental table S11). For loneliness, the association with IC showed evidence of heterogeneity for participants reporting more than one population (interaction HR=2.19, 95% CI 1.45 to 3.32, p<0.001) and by education (interaction HR=0.95 per year, 95% CI 0.92 to 0.99, p=0.018), but not by income (p=0.845; online supplemental table S12). Given sparse events in some strata, these interaction findings should be interpreted cautiously. Consistent with these estimates, unadjusted Kaplan-Meier curves displayed as cumulative incidence (1—survival probability) showed higher cumulative incidence of dementia and IC among participants with HL or higher loneliness (see online supplemental figures S3 and S4), and forest plots provide the corresponding adjusted HRs with 95% CIs (see online supplemental figures S5 and S6). Sensitivity analyses As a sensitivity analysis, we re-fit Cox models with HL coded as time-varying. Estimates were similar to the baseline-exposure models (online supplemental results SR1 and online supplemental tables S13 and S14). Results were materially unchanged when the ≥1 year EHR follow-up requirement was removed (online supplemental table S15). For dementia, the HR for HL was 1.378 (95% CI 1.280 to 1.483) in the primary model and 1.375 (95% CI 1.278 to 1.480) in the model without the ≥1 year requirement. For IC, the corresponding HRs were 1.670 (95% CI 1.612 to 1.730) and 1.666 (95% CI 1.608 to 1.726). Results for loneliness were also similar after removing this requirement: HR=1.408 (95% CI 1.119 to 1.773) versus 1.375 (95% CI 1.093 to 1.730) for dementia, and HR=1.483 (95% CI 1.356 to 1.623) versus 1.454 (95% CI 1.330 to 1.589) for IC. Fine-Gray competing-risk models treating death before dementia or IC as a competing event produced results consistent with the primary Cox models (online supplemental table S16). HL remained associated with dementia (subdistribution HR (sHR) = 1.394, p<0.001) and IC (sHR=1.682, p<0.001). Loneliness also remained associated with dementia (sHR=1.411, p=0.004) and IC (sHR=1.481, p<0.001). Mediation analysis: loneliness as a mediator of HL effect Full four-way outputs (including interaction components) are provided in table 4. Because we did not prespecify hypotheses about interaction between HL and loneliness, we restrict main-text reporting to mediation components. The total effect of HL on incident dementia was significant (Rte=1.83, p<0.001) and was driven primarily by the direct effect (Rcde=1.81, p<0.001; ERcde(prop) = 0.966). The indirect pathway via loneliness was small and not statistically significant for the total natural indirect effect (Rtnie=1.011, p=0.069), yielding an overall pm of 2.5% (pm=0.025, p=0.081). For incident IC, the total effect was likewise significant (Rte=2.00, p<0.001) and largely direct (Rcde=1.96, p<0.001), while the indirect effect through loneliness was small but statistically significant (Rtnie=1.010, p<0.001), yielding an overall pm of 2.0% (pm=0.020, p<0.001). Table 4(a). Four-way decomposition of the effect of HL on dementia via loneliness (N=86 387; events=179); (b) Four-way decomposition of the effect of HL on impaired cognition via loneliness (N=81 592; events=1154)ComponentDescriptionEstimate (95% CI)P valueModel 1: HL on dementia via loneliness RcdeControlled direct effect (HR)1.81 (1.33 to 2.47)<0.001 RpndePure natural direct effect (HR)1.81 (1.33 to 2.45)<0.001 RtndeTotal natural direct effect (HR)1.80 (1.33 to 2.45)<0.001 RpniePNIE (HR)1.013 (1.001 to 1.025)0.037 RtnieTotal natural indirect effect (HR)1.011 (0.999 to 1.024)0.069 RteTotal effect (HR)1.83 (1.34 to 2.48)<0.001 ERcdeExcess relative hazard due to CDE0.797 (0.595 to 1.000)<0.001 ERintrefExcess relative hazard due to reference interaction0.008 (−0.346 to 0.362)0.964 ERintmedExcess relative hazard due to mediated interaction0.008 (−0.017 to 0.032)0.541 ERpnieExcess relative hazard due to PNIE0.013 (0.001 to 0.025)0.038 ERcde (prop)Proportion of excess hazard from CDE0.966 (0.668 to 1.263)<0.001 ERintref (prop)Proportion from reference interaction0.010 (−0.294 to 0.314)0.949 ERintmed (prop)Proportion from mediated interaction0.009 (−0.019 to 0.038)0.525 ERpnie (prop)Proportion from PNIE0.015 (−0.003 to 0.033)0.095 pmOverall proportion mediated0.025 (−0.003 to 0.052)0.081 intOverall proportion due to interaction0.019 (−0.289 to 0.327)0.903 peOverall proportion eliminated0.034 (−0.263 to 0.332)0.820Model 2: HL on impaired cognition via loneliness RcdeControlled direct effect (HR)1.96 (1.73 to 2.22)<0.001 RpndePure natural direct effect (HR)1.98 (1.75 to 2.23)<0.001 RtndeTotal natural direct effect (HR)1.98 (1.75 to 2.24)<0.001 RpniePNIE (HR)1.009 (1.004 to 1.014)<0.001 RtnieTotal natural indirect effect (HR)1.010 (1.004 to 1.016)<0.001 RteTotal effect (HR)2.00 (1.77 to 2.26)<0.001 ERcdeExcess relative hazard due to CDE0.940 (0.852 to 1.028)<0.001 ERintrefExcess relative hazard due to reference interaction0.037 (−0.120 to 0.193)0.645 ERintmedExcess relative hazard due to mediated interaction0.011 (0.002 to 0.020)0.013 ERpnieExcess relative hazard due to PNIE0.009 (0.004 to 0.014)<0.001 ERcde (prop)Proportion of excess hazard from CDE0.943 (0.854 to 1.032)<0.001 ERintref (prop)Proportion from reference interaction0.037 (−0.053 to 0.127)0.423 ERintmed (prop)Proportion from mediated interaction0.011 (0.003 to 0.019)0.010 ERpnie (prop)Proportion from PNIE0.009 (0.004 to 0.014)0.001 pmOverall proportion mediated0.020 (0.009 to 0.031)<0.001 intOverall proportion due to interaction0.048 (−0.043 to 0.139)0.301 peOverall proportion eliminated0.057 (−0.032 to 0.146)0.211Rcde=controlled direct effect (HR): effect of HL on the outcome if loneliness were held at the reference value (see Methods for the chosen reference).Rpnde=pure natural direct effect (HR): HL effect not operating through loneliness and without exposure–mediator interaction.Rtnde=total natural direct effect (HR): HL effect not operating through loneliness but allowing interaction at the natural (counterfactual) mediator level.Rpnie=PNIE (HR): mediation through loneliness without interaction.Rtnie=total natural indirect effect (HR): mediation through loneliness with interaction (ie, includes mediated interaction).Rte=total effect (HR).ERcde, ERintref, ERintmed, ERpnie=excess relative hazard (on the HR scale) attributable to, respectively, the CDE, reference interaction, mediated interaction and PNIE. For example, ERcde=Rcde − 1 and ERpnie=Rpnie − 1. (1) INTref (reference interaction) captures HL×loneliness interaction when loneliness is fixed at the reference value. (2) INTmed (mediated interaction) is the part of interaction operating through the HL-induced change in loneliness.ERcde (prop), ERintref (prop), ERintmed (prop), ERpnie (prop)=proportions of the excess hazard due to each component: component ER divided by (Rte − 1).pm (proportion mediated) = (ERpnie+ERintmed)/(Rte − 1).int (proportion due to interaction) = (ERintref+ERintmed)/(Rte − 1).pe (proportion eliminated) = (ERpnie+ERintmed + ERintref)/(Rte − 1), that is, the fraction of excess hazard removed if both the mediator effect and exposure–mediator interaction were eliminated (holding HL fixed).CDE, controlled direct effect; HL, hearing loss; PNIE, pure natural indirect effect. (a). Four-way decomposition of the effect of HL on dementia via loneliness (N=86 387; events=179); (b) Four-way decomposition of the effect of HL on impaired cognition via loneliness (N=81 592; events=1154) Rcde=controlled direct effect (HR): effect of HL on the outcome if loneliness were held at the reference value (see Methods for the chosen reference). Rpnde=pure natural direct effect (HR): HL effect not operating through loneliness and without exposure–mediator interaction. Rtnde=total natural direct effect (HR): HL effect not operating through loneliness but allowing interaction at the natural (counterfactual) mediator level. Rpnie=PNIE (HR): mediation through loneliness without interaction. Rtnie=total natural indirect effect (HR): mediation through loneliness with interaction (ie, includes mediated interaction). Rte=total effect (HR). ERcde, ERintref, ERintmed, ERpnie=excess relative hazard (on the HR scale) attributable to, respectively, the CDE, reference interaction, mediated interaction and PNIE. For example, ERcde=Rcde − 1 and ERpnie=Rpnie − 1. (1) INTref (reference interaction) captures HL×loneliness interaction when loneliness is fixed at the reference value. (2) INTmed (mediated interaction) is the part of interaction operating through the HL-induced change in loneliness. ERcde (prop), ERintref (prop), ERintmed (prop), ERpnie (prop)=proportions of the excess hazard due to each component: component ER divided by (Rte − 1). pm (proportion mediated) = (ERpnie+ERintmed)/(Rte − 1). int (proportion due to interaction) = (ERintref+ERintmed)/(Rte − 1). pe (proportion eliminated) = (ERpnie+ERintmed + ERintref)/(Rte − 1), that is, the fraction of excess hazard removed if both the mediator effect and exposure–mediator interaction were eliminated (holding HL fixed). CDE, controlled direct effect; HL, hearing loss; PNIE, pure natural indirect effect. Discussion In this large, demographically diverse US volunteer cohort with linked EHR and survey data (mean age=44.2 years, SD=16.7), both HL and loneliness were associated with increased risk of IC and dementia. These findings extend prior evidence that midlife HL is a modifiable dementia risk factor1 and that loneliness is associated with worse late-life brain health.11 Consistent with the slightly larger effects for IC than for dementia, the results also suggest that psychosocial factors such as loneliness may shape cognitive ageing at earlier stages. Mediation analyses indicated that loneliness explained a small but non-zero fraction of the HL-IC association. The adjusted HRs we observed for HL (on the order of 1.38–1.67) are broadly consistent with prior prospective cohort meta-analyses linking HL with incident dementia and cognitive impairment, although direct numerical comparisons should be made cautiously because studies differ in outcome definitions, exposure ascertainment, populations and estimands.29 HL is common in later life—about 16% of our cohort had an HL diagnosis, and prevalence rises with age—and is potentially modifiable through rehabilitative interventions, notably hearing aids and cochlear implants. Recent evidence links these devices with slower cognitive decline and reduced dementia risk.3031 To the extent that the HL–dementia association is causal, addressing HL could delay or prevent a portion of cases. The Aging and Cognitive Health Evaluation in Elders (ACHIEVE) randomised trial found no overall benefit of a hearing intervention in the full sample, but among participants at higher dementia risk, a comprehensive hearing intervention (including hearing aid fitting and support) slowed cognitive decline by 48% over 3 years compared with a control group.32 Our observational findings are consistent with HL as an important clinical risk marker for later cognitive outcomes. However, because this study is observational and EHR-based, these results should not be interpreted as evidence that hearing rehabilitation prevents dementia; this question requires stronger causal evidence from intervention and quasi-experimental studies. Mechanistically, our results are consistent with the ‘cascade hypothesis’, in which HL may be linked to cognitive decline through multiple pathways. We explicitly examined one such hypothesised pathway—loneliness—and found that it explains only a small fraction of the HL effect. Using four-way decomposition, we found the pm by loneliness was ~2.5% for dementia (pm=0.025, p=0.081) and ~2.0% for IC (pm=0.020, p<0.001). Thus, HL may be associated with greater loneliness, which may then modestly contribute to cognitive risk, supporting the idea that social engagement is important for cognitive resilience. However, the mediation share was small. Although loneliness and social isolation are related but distinct constructs, this finding is broadly consistent with a recent systematic review reporting limited evidence that social isolation mediates the HL–cognition association18; in that review, only one longitudinal study explicitly tested mediation and found no significant effect. Our estimate implies that interventions solely aimed at reducing loneliness among those with HL might have a relatively minor impact on dementia incidence, if other pathways account for most of the association. That said, even a small mediated effect can be meaningful at a population level, given the large number of people with HL. Overall, this pattern supports further research on combined hearing and social-connection approaches, while recognising that loneliness is unlikely to be the primary pathway linking HL with cognitive outcomes. We confirmed that loneliness independently predicts risk of dementia and IC, consistent with growing evidence that loneliness/social isolation may harm brain health—for example, a recent meta-analysis also found that loneliness was associated with higher dementia incidence, although reported effect measures varied across studies.11 In adjusted hazard models, loneliness remained a significant predictor, potentially reflecting pathways noted previously—chronic stress, vascular dysregulation and reduced cognitive engagement—though we did not measure these mechanisms.12 Importantly, loneliness may also reflect prodromal dementia—preclinical withdrawal or isolation—complicating causal interpretation.16 Although we excluded baseline cognitive-impairment diagnoses, prodromal effects cannot be fully excluded. We used a working causal model in which HL may increase loneliness, which in turn may contribute to dementia risk. Although direct biological pathways from loneliness to auditory pathology are uncertain, recent prospective evidence indicates that loneliness predicts incident HL, suggesting potential bidirectionality.33 In our data, participants with HL reported higher loneliness at baseline, consistent with the hypothesised hearing→loneliness pathway; however, because loneliness was measured at a single time point, we cannot determine whether loneliness emerged after HL onset, was already present when HL was first recorded or reflected early/prodromal cognitive or functional changes. That said, common causes—such as personality traits or depressive symptoms—could predispose individuals both to feel lonely and to have worse or unaddressed hearing, and residual confounding may remain despite adjustment for demographic and behavioural covariates. Future work with longer postsurvey follow-up, repeated psychosocial assessments and designs that better separate timing (eg, time-updated models or quasi-experimental approaches) will be important to clarify directionality and mechanisms. Subgroup analyses and generalisability Leveraging the size and diversity of AoU, we tested effect modification by race/ethnicity, education and income in Cox models for HL and for loneliness. For dementia, the HL association was stronger among black participants, with no clear evidence of interaction by income or education. For loneliness, we found no clear evidence of effect modification by race/ethnicity, income or education for dementia. For IC, the HL association varied by race, with stronger associations among Asian and black/African American participants, but showed no interaction with income or education. Loneliness showed evidence of heterogeneity for IC among participants reporting more than one population and by education, but not by income. A stronger HL-dementia association in black participants aligns with prior cohort data from the Atherosclerosis Risk in Communities Neurocognitive Study, which showed larger dementia hazards per 10 dB worse hearing among black than white older adults, and may reflect differences in cumulative vascular and social risks as well as inequities in hearing-care access and device uptake that could amplify downstream cognitive risk.3435 Recent work also highlights measurement and cultural/contextual variability in loneliness across racial/ethnic groups, including potential differences in stigma and disclosure and in how loneliness instruments relate to cognitive outcomes, which may contribute to differences in observed associations across subgroups.3637 The apparently stronger loneliness association among participants in the ‘more than one population’ racial/ethnic category should be interpreted cautiously given small cell sizes and heterogeneity within this group. These patterns merit confirmation with additional follow-up and repeated psychosocial measures. Because AoU is a volunteer cohort and the present analyses required linked EHR data, sufficient EHR follow-up, survey completion for loneliness analyses and complete model covariates, absolute risks and some association estimates may not transport directly to the general US adult population. Strengths and limitations The major strengths of this study include its very large sample size, diverse participant pool, longitudinal EHR follow-up, participant responses to surveys of lifestyle and social factors, use of time-varying Cox models and adjustment for multiple prespecified confounders. We also used four-way decomposition to evaluate mediation while allowing exposure–mediator interaction, adding to the existing literature on potential psychosocial pathways linking HL with cognitive outcomes. These strengths must be understood in light of several limitations. First, HL status was determined from EHR diagnosis codes using a phenotype that, to our knowledge, has not been formally validated against audiometric testing within AoU. This approach captures clinically recognised HL but likely under-ascertains mild, untreated or unrecorded HL, and we could not distinguish severity, laterality, duration, age at onset or hearing aid use. Such exposure misclassification may have affected prevalence and association estimates. Because both HL and cognitive outcomes were EHR-ascertained, differential healthcare utilisation may also have influenced whether either condition was recorded. Second, dementia and IC may have been under-recorded, particularly among participants with limited healthcare access or early-stage disease.3839 The IC outcome was intentionally broad and included heterogeneous EHR-recorded cognitive diagnoses and symptoms. Although this may improve sensitivity for detecting cognitive problems documented in routine care, IC should not be interpreted as equivalent to clinically adjudicated mild cognitive impairment, prodromal dementia or a single aetiological construct. Dementia subtypes were not distinguished. In addition, censoring at the last recorded EHR encounter may have been informative if healthcare disengagement, death or loss to follow-up was related to cognitive risk. Missingness in the online SDoH survey may also have disproportionately excluded cognitively impaired participants, introducing selection bias in loneliness models. Third, loneliness was measured once, and HL status was not updated in the primary models. Because both exposures can change over time, temporal ambiguity and exposure misclassification remain possible, although a time-dependent HL sensitivity analysis suggested minimal short-term impact. Loneliness may have preceded HL, followed HL or reflected prodromal cognitive or functional changes; therefore, the mediation analyses should be considered exploratory rather than definitive evidence of a causal pathway. Studies with longer follow-up and repeated loneliness measures are needed. Fourth, despite adjustment for measured demographic, socio-economic and behavioural covariates, residual confounding remains possible because factors such as baseline cognitive performance, social network size, depressive symptoms, APOE ε4 status and comorbidity burden were not uniformly available or consistently measured before baseline. Some of these factors may also lie on the causal pathway or reflect prodromal disease, making confounding difficult to distinguish from mediation or reverse causation. Fifth, hearing-care and loneliness-reduction interventions during follow-up were not modelled. Finally, volunteer participation, differential EHR linkage, incomplete survey participation and missing covariate data may limit the generalisability of the observed associations and absolute risks despite the diversity of the AoU cohort. Conclusion In this large, diverse AoU cohort, HL and loneliness each predicted higher risk of dementia and IC. The HL effect appeared to act largely through direct pathways, with a small contribution via loneliness, while loneliness itself independently elevated risk. Looking ahead, further triangulation is needed to strengthen causal inference. Prospective cohorts with repeated measures of hearing, loneliness and cognition; quasi-experimental studies; Mendelian randomisation; and intervention trials targeting hearing rehabilitation and social connection will be important to test whether the patterns observed here reflect modifiable causal pathways. In parallel, health systems and payers can begin to evaluate scalable strategies such as integrating basic hearing screening into routine care, expanding access to affordable hearing devices (including over-the-counter options) and linking patients with HL to resources that support social engagement. Future work with longer follow-up, time-updated measures and genetic data should clarify mechanisms and identify who benefits most, enabling integrated, life-course approaches to preserve cognitive health. Supplementary material