Trial design

Multicenter, parallel-group, unmasked, active-controlled randomized trial with a 1:1 allocation ratio.

Participants, eligibility criteria, and settings

The ACHIEVE study is based within the Atherosclerosis Risk in Communities (ARIC) study, an ongoing longitudinal study in four US field sites (Forsyth County, NC; Jackson, MS; Minneapolis, MN; and Washington County, MD) of 15,792 adults aged 45–64 years were recruited from a random community sample in 1987–89 (S1 File) [20]. Participants continue to be followed as part of the ARIC-Neurocognitive Study (ARIC-NCS) [20–21].

ACHIEVE participants (N = 977) were recruited from the four ARIC sites between 2018–2019; 238 participants were from ARIC-NCS, and 739 participants were de novo volunteers from the community. Inclusion criteria were: 70–84 years old; community-dwelling; audiometric hearing loss [better-hearing ear pure-tone average (PTA) ≥30 and <70 dB hearing level (dB HL)]; Mini-Mental State Exam ≥23 for high school degree or less, and ≥25 for some college or more; Word Recognition in Quiet score ≥60% correct (better-hearing ear); fluent English-speaker; and remaining in the area during the study. Exclusion criteria were: self-reported difficulty in ≥2 activities of daily living [22]; prior dementia diagnosis; presenting vision worse than 20/63 (~14-point font) with usual correction [23]; medical contraindication to hearing treatment; untreatable conductive hearing impairment; unwillingness to regularly wear hearing aids; and self-reported hearing aid use in the past year [17,18,24]. We excluded participants missing baseline SPPB (n = 10), grip strength (n = 5), education (n = 1), and BMI (n = 5). Our analytic sample was n = 956.

The recruitment process for the trial has been previously described in detail [24]. A soft-launch of recruitment efforts began at a limited number of study sites in November 2017 and full-scale efforts at all sites in January-February 2018. The original target sample size (N = 850) was achieved in July 2019. After interim Data Safety and Monitoring Board (DSMB) analysis of drop-in, drop-out and missing data, and prior to recruitment closure, with National Institute of Aging approval, investigators were authorized to extend the recruitment window, which closed on October 27, 2019 with a final sample of N = 977. The trial concluded as scheduled at the end of 3 years. The Institutional Review Boards (IRBs) at each collaborating site, including the 4 field centers and the coordinating center (University of Mississippi Medical Center IRB, University of Minnesota IRB, Wake Forest University IRB, Johns Hopkins Bloomberg School of Public Health IRB, University of North Carolina IRB) approved the trial, and participants provided written informed consent. An independent data and safety monitoring board oversaw study progress, adverse events, and changes to the study protocol and statistical analysis plan. We followed EQUATOR reporting guidelines (S2 File) [25].

Interventions

For the hearing intervention, participants completed four 1-hour sessions with an audiologist spaced every 1–3 weeks post-randomization. Intervention included bilateral hearing aids fitted to prescriptive targets using real-ear measures, paired with hearing-assistive technologies (e.g., devices to stream smartphones and television), systematic orientation and instruction in device use, and provision of hearing toolkit materials for self-management and communication strategies [26–27]. Re-instruction was provided during booster visits every 6 months [17].

For the health education control intervention, participants completed four 1-hour sessions with a certified health educator spaced every 1–3 weeks post-randomization with subsequent booster sessions every 6 months. The intervention included administration of the 10 Keys^TM to Healthy Aging program [28], an evidence-based, interactive, health education program for older adults on topics relevant to chronic disease and disability prevention [17].

Outcome measures

Physical function, including the SPPB and grip strength, were pre-specified secondary outcomes of the trial. SPPB [19] and grip strength were measured at baseline and each annual in-person follow-up visit. The SPPB is a series of physical performance tests (chair stands, balance, 4-meter walk) to assess lower extremity function in older adults that predicts mortality [29], falls [30], hospitalizations [31] and incident disability [19,32]. For chair stands, participants were asked to fold arms over chest and rise to standing from a seated position in a chair. If able to complete one, participants were timed as they completed five in a row, as quickly as possible without stopping, keeping arms folded over chest. Standing static balance scores are based on the participant’s ability to hold 3 balance tasks (side-by-side, semi-tandem, full tandem) for 10 seconds each. To maximize time allotment for testing, participants started with feet in a semi-tandem position. If able to hold for the full 10 seconds, it was assumed that they would also be able to hold the easier side-by-side position and were scored as having completed it. They then attempted the full tandem stand. If participants were unable to hold the semi-tandem stand, they attempted the side-by-side position and received a score of 0 seconds for the full tandem position.

Standardized SPPB scoring is available, ranging from 0 to 12 [19,33]. However, given that our population was high-functioning, and that ceiling effects in high-functioning populations can lead to biased effect estimates [34], for analysis, we rescaled SPPB scores according to published guidelines developed in the Health Aging and Body Composition study [35]. To account for non-normality, we converted testing times to rates for chair stands (chair stands/s) and 4-meter walk (m/s), assigning a score of 0 when a participant was not able to perform the test. For standing balance, we summed the time for the 3 balance positions. Component scores were then divided by the maximal performance possible based on external data from other studies (1 chair stand/s; 2 m/s walk time; 30 seconds for balance) to derive 3 component ratio scores, each ranging from 0 to 1. We then summed the component ratio scores to get a continuous rescaled SPPB summary score ranging from 0 to 3, with higher scores indicating better performance.

Grip strength was measured in kg with a Jamar handheld dynamometer in the dominant hand. After one practice trial, participants completed two trials, squeezing as hard as possible, with a 15–20 second rest between trials. The two trials were averaged for analysis, modeled continuously.

Component scores were modeled as speed for chair stands (number of chairs stands/s) and the 4-meter walk (m/s), with higher scores indicating faster completion rates. Because time to complete the standing balance trials was heavily skewed, we modeled balance as a binary variable taking a value of 0 if unable to hold all 3 positions for the full time, and 1 if able.

We standardized continuous outcome scores to baseline values by subtracting baseline means and dividing by baseline standard deviation (SD) in our analytic sample. Baseline distributions for outcome measures are therefore mean zero, SD of one, with subsequent values representing change relative to baseline.

Minimal detection effect size calculations

Given that physical function was a secondary outcome, the ACHIEVE study was not powered to detect differences in rates of SPPB change over time. To guide the interpretation of the results of the present study, we conducted post-hoc estimation of the minimal detection effect size estimates. These estimates suggest, at the alpha = 0.05 level, our study can detect a mean difference in SPPB score change comparing intervention to control of at least 0.36 points with 80% power and 0.42 points with 90% power.

Interim analyses and stopping guidelines

There were no interim analyses for efficacy.

Randomization (random number generation, allocation concealment, implementation)

Randomization was determined by an allocation schedule developed by the Data Coordinating Center at the University of North Carolina (Chapel Hill, NC, USA) and completed within the CDART, the web-based data management system. Participants were randomized in a 1:1 ratio to either the hearing or the control intervention To ensure balance between the treatment groups, participants were randomized in permuted order blocks of varying sizes within strata defined by severity of hearing loss, defined as mild (PTA ≥ 30dB and <40dB) or moderate (PTA ≥ 40 dB and <70dB), recruitment source (ARIC or de novo participant), and by field site. Block size was not revealed to field center staff as this would have allowed them to determine the final treatment assignment of a block before ascertaining eligibility and obtaining consent. [17].

Masking

Participants and study staff collecting outcome data were unmasked to treatment assignment. Participants were masked to study hypothesis. ACHIEVE investigators and staff (except coordinating center staff and one statistician) were masked to trial data until the primary findings were reported [17].

Statistical analysis

Participant characteristics were compared by intervention assignment and recruitment source. Our primary analysis was intention-to-treat and restricted to available cases, estimating the effect of the hearing intervention on 3-year change in physical function outcomes (rescaled SPPB score and grip strength) using two-level linear mixed effects models with an unstructured covariance matrix for the random effects (intercepts and slopes). We restricted to data from baseline and Year 3 given large amounts of missing outcome data during the intervening years due to COVID-19 pandemic shutdowns of the study clinics (Fig 1). Conversions to account for non-normality of some outcomes are described in the Outcome Measures section of the Methods. Time was modeled continuously, and an interaction term between time and intervention assignment was the regression coefficient used to test if the rate of change in the outcome differed by intervention assignment. Restricted maximum likelihood with a Kenward-Roger correction was used to generate parameter estimates and 95% confidence intervals (CI); these methods were chosen over others (e.g., maximum likelihood methods) as they can reduce bias in smaller sample sizes [36]. To control for possible residual confounding, models adjusted for baseline age, sex, race, field site, education, recruitment source, BMI, hearing loss severity, and interaction terms between time and all covariates except education. Participants self-reported age (years, modeled continuously), sex (male, female), educational attainment (less than high school, high school or equivalent, greater than high school) and race (Black, White, Asian, American Indian or Alaska Native) at baseline. Recruitment source was recorded as ARIC-NCS or de novo. Body mass index (BMI; kg/m²) was calculated using measured height and weight and modeled continuously. Audiometric thresholds at 0.5, 1, 2, and 4 kHz were averaged, and a binary variable was created to indicate greater hearing loss severity (PTA ≥ 40 dB HL vs. < 40 dB HL) in the better-hearing ear. Model fit was adequate and assumptions for linearity met as assessed using standard fit indices, including the AIC and BIC and by visualizing the measures and model residuals over time. For grip strength, given documented sex differences in cutpoints for clinically relevant weakness [37], we included an interaction term between sex and baseline grip strength, and a 3-way interaction between sex, intervention assignment, and time, to estimate the hearing intervention effect by sex.

Download:

• PNG
  larger image
• TIFF
  original image

Fig 1. Consort Flow Diagram for Analysis of Physical Function, Aging and Cognitive Health Evaluation in Elders (ACHIEVE) Study, 2018−19 to 2021−22.

Abbreviations: ARIC, Atherosclerosis Risk in Communities; PF, Physical Function. Participants must have completed all components of the physical function battery (chair stands, balance, walking speed, grip strength) to have complete physical function data.

https://doi.org/10.1371/journal.pone.0347500.g001

In additional analyses, we modeled the SPPB components separately. Linear mixed models to estimate rate differences by intervention group for chair stands and walking speed are the same as those described above, including adjustment covariates. We estimated the marginal odds ratio of inability to hold all 3 positions over follow-up using Generalized Estimating Equations (GEE) [38] with an unstructured correlation matrix and robust standard errors. Adjustment factors were the same as above.

Given that findings for the primary ACHIEVE trial outcome (3-year cognitive decline) differed by recruitment source [17], and hypotheses that the effect of hearing loss on balance may be stronger in individuals with poorer cognition [39], we conducted a sensitivity analysis that was pre-specified in the statistical analysis plan, which was registered at ClinicalTrials.gov before the unmasking of trial data, and re-ran the analysis, stratifying by recruitment source.

We conducted 5 additional sensitivity analyses. Under a missing at random assumption, missing outcome and covariate information was imputed using multiple imputations by chained equations for participants who were alive at the time of the study visit. We re-ran models, including available Year 1 outcome data. We also modeled the original SPPB score, to allow comparisons with published literature. We conducted a per-protocol analysis excluding study participants who did not follow their assigned intervention. Finally, we estimated the complier average causal effect (CACE) of hearing intervention among participants assigned to hearing intervention. Methodological details for sensitivity analyses are provided in supplemental materials (S3 File).

Because the trial was not powered to detect intervention differences for this pre-specified exploratory outcome, we focused on patterns of associations for hypothesis-generation instead of hypothesis-testing for statistical significance. All analyses were conducted using Stata 18.0 (StataCorp, College Station, TX).

Limitations

The trial was not powered to detect intervention differences on physical function decline and so our results should be considered hypothesis-generating on their own, although they may contribute to future systematic reviews and meta-analyses. Although all participants completed the baseline visit, the COVID-19 shutdown resulted in missing outcome data, particularly in Years 1 and 2. Results were unchanged in sensitivity analysis using multiple imputation. Given that missingness was due to study visit disruptions rather than participant characteristics, the mechanism for missingness could plausibly be considered missing at random (MAR) conditional on observed covariates. However, we cannot rule out the possibility that missingness was related to unmeasured factors. We may have been lacking strong auxiliary variables related to physical function and to missingness, reducing the potential efficiency gains typically observed under MAR assumptions. Participants are from four US study sites and primarily self-identified White or Black race, and so generalizability of study findings may be limited to other geographic locations and race/ethnicity groups. Participants were not masked to intervention assignment, but physical function was measured objectively in our study, reducing the possibility of bias in outcome reporting by treatment group.