Association of Age-Related Hearing Loss With Cognitive Function, Cognitive Impairment, and Dementia

Search Strategy and Selection Criteria

Six a priori meta-analyses were planned across 2 levels of study design (cross-sectional and cohort) and 3 levels of outcome (cognitive function, cognitive impairment, and dementia). The inclusion criteria consisted of (1) cross-sectional and cohort studies, excluding case-control studies because of greater concern about sampling and retrospective analysis bias¹³ (all study designs have selected types of bias); (2) published studies (any language); (3) study sample 18 years or older; (4) baseline sample including the general, community-dwelling population rather than special risk groups (eg, patients with coronary heart disease); (5) individual’s peripheral hearing status (as assessed by pure-tone audiometric assessment) as the main exposure variable; (6) full inclusion of hearing loss sample (ie, no pure-tone audiometric cutoff); (7) assessment of cognitive function, cognitive impairment,¹⁴ and/or dementia as outcome(s); and (8) exposure and outcome measurements obtained by health care professionals or trained investigators (ie, not based on self-reported data).

Studies published on or before August 26, 2015, were retrieved from the following 4 electronic databases by one of us (D.G.L.): (1) PubMed, (2) the Cochrane Library, (3) EMBASE, and (4) SCOPUS. Keywords included hearing, cognition, dementia, and Alzheimer disease (eTable 1 in the Supplement). Results were updated on April 15, 2016. Cross-referencing for potentially eligible studies was conducted using retrieved studies and personal files belonging to one of us (D.G.L.).

Data Extraction and Quality Assessment

Two of us (D.G.L. and M.E.K.) independently screened for eligible studies and conducted data extraction. If consensus could not be reached, another of us (B.A.L.) acted as arbitrator for study inclusion, and another (G.E.K.) was consulted regarding data extraction. Cognitive function was subdivided into 10 domains, including episodic memory (delayed recall and immediate recall), executive functions (attention, fluency, reasoning, and working memory), global cognition, processing speed, semantic memory, and visuospatial ability.¹⁵ Among dementia studies, a secondary outcome of interest was any data that examined subgroups (eg, AD).

Data from the most recently published study were selected. Data from different studies that examined the same cohort were included if they were for different cognitive outcomes and were treated as separate studies in analysis. Priority was given to outcomes that were maximally adjusted for covariates. Two of us (D.G.L. and M.E.K.) independently assessed the quality of reporting for each study using the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) instrument.¹⁶ With use of the Cohen κ coefficient,¹⁷ agreement was excellent (κ = 0.91) before correcting discrepant items.

Statistical Analysis

We chose the Pearson r correlation coefficient as the effect size of the linear association between hearing loss and cognitive function (continuous variables). Negative scores indicated that greater hearing loss was associated with poorer cognition. Odds ratios (ORs) were chosen for cognitive impairment and dementia (categorical variables). Influence of various audiometric criteria (eg, worse vs better ear) and cognitive tests (visual vs auditory stimuli) on outcome were examined in subgroup analyses. If the required outcome metric was not reported in the study, values were calculated using available data. Random-effects, method-of-moments models that incorporate heterogeneity into the overall estimate were used to pool effect sizes from each study.¹⁸ All outcomes were converted to Fisher z values or logarithm ORs for analysis purposes and then converted back to the original metric (ie, r correlation coefficient and OR, respectively). For both meta-analyses of cognitive function, multiple tests of the same cognitive domain from the same study were collapsed into a single effect size and within-study subgroups were analyzed independently as separate effect sizes.

Heterogeneity was examined using the Q test, and P ≤ .10 was considered to be statistically significant.¹⁹ Inconsistency was examined using the I² statistic, and the following grades were applied: less than 25% indicated very low; 25% to less than 50%, low; 50% to less than 75%, moderate; and 75% or greater, large.¹⁹ Small-study effects were examined using funnel plots, and the regression-intercept approach of Egger and colleagues²⁰ provided at least 10 effect sizes were present. To examine the influence of each result on the overall findings, outcomes were analyzed by deleting each study from the model once. Cumulative meta-analysis ranked by year was used to examine the accumulation of evidence over time.²¹

We conducted subgroup and meta-regression analyses to examine heterogeneity between studies. Planned variables included (1) study characteristics, (2) participant characteristics, (3) audiometric factors, (4) cognitive measures, and (5) statistical analysis (eTable 2 in the Supplement provides a list of each planned variable). For continuous variables, we used random-effects meta-regression²² where at least 4 effect sizes were found. For categorical variables, we examined between-group differences (between-group Q value) in effect sizes using mixed effects analysis of variance–like models for meta-analysis²² if at least 3 effect sizes were available for each category. These analyses were considered to be exploratory.

Characteristics and Quality of Included Studies

The characteristics of included studies are shown in eTable 3 in the Supplement. Of the 1185 citations reviewed, 40 studies^{7,23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48,49,50,51,52,53,54,55,56,57,58,59,60,61} met the inclusion criteria, representing 34 471 participants from 12 countries (Figure 1). Of these, 36 unique studies with an estimated 20 264 unique participants were included in the meta-analyses. Study quality results are shown in Table 1 and Table 2 and eFigure 1 in the Supplement. Further details on the main analyses are found in eFigures 2 to 27 and eTables 3 to 9 in the Supplement; and further details on the small-study, influence, and cumulative analyses, in the eResults and eFigures 28 to 74 in the Supplement. Thirty-five of the 40 included studies (88%) met the criteria for at least 16 of 22 STROBE items. Further details on the main analyses and descriptions of the small-study, influence, and cumulative analyses are found in eFigures 2 to 27 and eTables 3 to 9 in the Supplement. Diagnostic criteria for each clinical outcome are shown in the Table 1 and Table 2.

Twenty-six studies with 15 620 participants were included in the cross-sectional cognitive function analysis.^{23,24,25,26,27,28,29,30,31,32,33,34,35,36,37,38,39,40,41,42,43,44,45,46,47,48} Two studies were omitted because of duplicate data.^49,50 Nine studies with 8233 participants^{7,29,36,40,46,48,51,52,53} were included in the cohort cognitive function analysis with a follow-up ranging from 2 to 23 years (mean [SD], 10.4 [6.7] years).

Five studies with 6582 participants (797 cases of 6553 included participants [12.2%])^{30,55,56,57,58} were included in the cross-sectional cognitive impairment analysis. Two studies were omitted because of duplicate data.^54,59 Three studies with 7817 participants (1395 cases of 6825 included participants [20.4%]) were included in the cohort cognitive impairment analysis with a follow-up ranging from 6 to 18 years (mean [SD], 11.7 [6.0] years).^7,40,55

Two studies with 741 participants (59 cases of 679 included participants [8.7%])^58,60 were included in the cross-sectional dementia analysis. One study assessed dementia (39 cases of 245 included participants [15.9%]),⁶⁰ and the other assessed AD (20 cases of 434 included participants [4.6%]).⁵⁸ Three studies with 3585 participants^7,29,61 (10.4%) were included in the cohort dementia analysis with a follow-up length ranging from 9 to 18 years (mean [SD], 15.0 [5.2] years). All 3 studies reported incident dementia outcomes (366 cases of 3439 included participants [10.6%]), 2 examined an AD subset (78 cases of 1491 included participants [5.2%]),^7,61 and 1 examined a vascular dementia subset (38 cases of 870 included participants [4.4%]).⁷

Hearing Loss and Cognitive Function

We found a small but statistically significant association between ARHL and all 10 cognitive domains of interest in cross-sectional studies, including global cognition (r = −0.15; 95% CI, −0.18 to −0.11), executive functions (range, r = −0.08 [95% CI, −0.12 to −0.04] to r = −0.18 [95% CI, −0.25 to −0.10), episodic memory (range, r = −0.10 [95% CI, −0.16 to −0.04] to r = −0.14 [95% CI, –0.20 to –0.09]), processing speed (r = −0.13; 95% CI, –0.18 to 0.08), semantic memory (r = −0.14; 95% CI, –0.20 to –0.08), and visuospatial ability (r = −0.11; 95% CI, –0.19 to –0.03). Similar results were observed in 7 of 8 domains in cohort studies, excluding fluency, which was not significant (r = −0.07; 95% CI, –0.14 to 0.01). These results included global cognition (r = −0.14; 95% CI, –0.19 to –0.09), executive functions (range, r = −0.06 [95% CI, –0.12 to –0.004] to r = −0.10 [95% CI, –0.20 to –0.001]), episodic memory (range, r = −0.06 [95% CI, –0.10 to –0.02] to r = −0.10 [95% CI, –0.15 to –0.05), processing speed (r = −0.08; 95% CI, –0.14 to –0.03), and semantic memory (r = −0.14; 95% CI, –0.23 to –0.05) (Figure 2, Figure 3, and eFigures 10-27 and eTables 1-9 in the Supplement). No cohort data were available for visuospatial ability or working memory. Heterogeneity was significant in most domains (Q range, 0.0-79.9). Inconsistency ranged from very low to high.

Hearing Loss and Cognitive Impairment

We found a statistically significant association between ARHL and cognitive impairment across cross-sectional (OR, 2.00; 95% CI, 1.39-2.89) and cohort studies (OR, 1.22; 95% CI, 1.09-1.36) (eFigures 2 and 3 and eTable 7 in the Supplement). Statistically significant heterogeneity (Q range, 0.1-23.7) and a large amount of inconsistency were observed in cross-sectional but not in cohort studies.

Hearing Loss and Dementia

We found a significant association between ARHL and dementia in cross-sectional (OR, 2.42; 95% CI, 1.24-4.72) and cohort (OR, 1.28; 95% CI, 1.02-1.59) studies (eFigures 4-9 and eTable 7 in the Supplement). Statistically significant heterogeneity (Q range, 0.4-6.6) and a moderate amount of inconsistency were observed in cohort but not cross-sectional studies. No statistically significant association was found between ARHL and AD for cross-sectional (OR, 1.80; 95% CI, 0.58-5.60) or cohort (OR, 1.69; 95% CI, 0.72-4.00) studies. In addition, the association between ARHL and vascular dementia was not significant (OR, 2.40; 95% CI, 0.99-5.82).

Subgroup Analyses and Meta-regression

The results of the subgroup and meta-regression analyses for cognitive outcomes are summarized below (eTables 8 and 9 in the Supplement). The respective Fisher z values (moderator analysis), slope (meta-regression), SEs, and 95% CIs for each variable are available in eTables 10 to 36 in the Supplement.

Statistical Analysis

A stronger association was generally found for studies that used correlation as the statistical model (compared with linear regression or linear mixed models) and those that reported results as significant. Studies that used age, sex, race, educational attainment, and vascular factors as covariates in their analysis typically reported weaker (sometimes significantly weaker) associations. This same trend was observed for studies that controlled for stroke, hypertension, diabetes, and current or previous smoking. Controlling for depression significantly weakened the association with cross-sectional attention. Results for premorbid IQ were mixed and nonsignificant except for cohort global cognition.

Because of a lack of data, no other a priori variables were examined. Other variables were reviewed ad hoc. A significantly weaker association was generally found for analyses that controlled for study site. These analyses were not conducted for cognitive impairment and dementia outcomes owing to lack of studies, with the exception of cross-sectional cognitive impairment studies. Year of publication, age (mean and minimum), sex, sample degree of hearing loss, proportion with hearing loss and cognitive impairment, impact factor, and STROBE were assessed (eTable 37 in the Supplement). No association was statistically significant.

Causal Mechanisms for ARHL and Cognitive Decline

The association between ARHL and cognitive decline remains unclear.⁶¹ One hypothesis is a common etiology, such as decline in the vascular system or a broader physiological decline. Age-related hearing loss has been linked with multiple indicators of functional decline and is a biomarker for frailty syndrome, which has been causally linked to dementia.⁶⁴ Other hypotheses suggest that the association may be mechanistic, for example, ARHL causing cognitive decline through impaired speech perception.⁶¹

Vascular risk factors contributed significantly to decline in global cognition and processing speed. However, the pooled effect size of studies controlling for vascular risk factors in these outcomes remained significant, suggesting other contributing factors, for example, depression, which significantly moderated the association with attention.

Of interest, the pattern of decline observed in this study was consistent with estimated cognitive outcomes based on behavioral and neuroimaging research.⁶⁵ This research reports increased recruitment of short-term memory and executive functions to aid speech perception after acquired hearing loss and concomitant decline in auditory cortex regions.⁶⁶ This situation is estimated to lead to less decline in these functions but greater decline in episodic and semantic long-term memory owing to reallocation of cognitive resources.⁶⁵ Consistent with this research, we observed that hearing loss was less associated with decline in executive functions and immediate recall compared with delayed and semantic memory and was increasingly less predictive of decline in attention and immediate recall among those with greater hearing loss. In addition, semantic memory, usually maintained in older age compared with episodic memory,⁶⁷ demonstrated a decline similar to that of episodic memory. Furthermore, the results indicated that hearing aids may benefit short-term and semantic memory.

The stronger association for low- to middle-frequency hearing loss with immediate recall and processing speed may be attributable to advanced aging as ARHL progresses from high to low frequencies.⁶⁴ Of interest, vascular dysfunction has been associated with lower-frequency hearing loss and white matter hyperintensities.⁶⁸ Alternatively, reallocation of executive functions to support accuracy in speech perception may be associated with decline in performance speed, as also observed in older adults with visual processing deficits.⁶⁹

Future Directions

Cognitive decline is influenced by multiple modifiable health factors.³ Hearing loss may be another serviceable risk factor, because it is easily diagnosed and can be treated.⁷⁰ Although associations were small, treatment may cumulatively benefit cognition as observed in intervention studies in older adults without cognitive impairment.^71,72,73 This benefit was not observed in patients with dementia, but treatment may still reduce disability.⁷⁴ Decline in lexical or semantic, episodic memory, and executive functions is used by clinicians as a marker for probable AD and vascular dementia.⁷⁵ In patients with ARHL, these domains may benefit from improved verbal communication through use of hearing aids. Additional randomized clinical trials exploring the cognitive benefits of hearing loss treatment are required, as is more research as to whether treatment, alone or as part of a wider approach to risk factors, modifies dementia outcomes. Neuroimaging studies could examine modification of cortical changes and neurocognitive compensation with hearing aid use in speech tasks. Future epidemiologic research might assess whether ARHL is associated with cognitive decline independently of neuropathologic hallmarks of dementia and whether a mediator of this association exists (eg, loneliness). Also of interest would be whether cognitive reserve moderates cognitive decline in the population with ARHL. Our results indicated a moderator effect of educational attainment, which is often used as a proxy for cognitive reserve.⁷⁶

Increasing evidence suggests that ARHL is associated with a wide range of health issues, higher disease burden, and increased risk for hospitalization,^64,77 leading to greater awareness of this condition as a critical public health concern.^70,77 In the United States, only 1 in 5 adults with hearing loss wears hearing aids, possibly owing to cost, lack of insurance coverage, or lack of knowledge of health care options, particularly for milder loss.⁷⁰ The National Academies of Sciences, Engineering, and Medicine recently outlined several recommendations to address this issue, with implications for public health services and policy.^70,78 Initiatives to expand access to treatment through screening programs, expand delivery of hearing services, and provide coverage for assistive hearing devices would be beneficial.⁷⁰ In addition, primary health care clinicians would benefit from standard guidelines for screening and referring patients with hearing loss.⁷⁰

Strengths and Limitations

To the our knowledge, this study is the first systematic review and meta-analysis of ARHL and cognitive decline that used only pure-tone thresholds as the audiometric criteria. Our strict inclusion criteria in study design and measurement allowed us to reduce conceptual heterogeneity and thus provide the most accurate quantitative measure of this association. Considerable heterogeneity remained across most outcomes. However, in any adjusted estimate of effect size for risk factors derived from aging studies, residual confounding will exist. Extensive subgroup and meta-regression analyses investigating this heterogeneity provided insight into how future studies may reduce bias and explore the potential basis of this association in experimental and clinical trials.

This study has several limitations. We could not examine whether studies controlled for etiology of hearing loss (eg, congenital or prelingual deafness). However, because of the low prevalence (<2%) of hearing loss in patients younger than 40 years, this prevalence was considered to be insignificant.⁷⁷ Some of the meta-analyses had a low number of effect sizes. We could not examine other planned moderators and covariates, such as attrition, owing to lack of data. For meta-analyses of dementia subgroups, the number of cases was small. Furthermore, because the meta-analyses were of observational studies, support for any inferences regarding the causal nature of the association is limited and cannot provide direct evidence for policy recommendations. However, our analyses of prospective studies give an indication of the temporal order of the association consistent with a causal effect. Further research is required to determine whether a causal relationship exists. Owing to the large number of statistical tests conducted, some of our findings could have been the result of chance. However, we did not want to risk missing potentially important findings that could be tested in future original studies. Finally, as is the case with any aggregate data meta-analysis, the potential for ecological fallacy exists.