Network Meta-Analysis of Cognitive Impairment and miRNA Expression in Alzheimer’s Disease Patients with Hearing Loss: A Systematic Review and Cross-Validation

1. Introduction

Global Burden of Alzheimer's Disease and the Clinical Link with Hearing Loss

Alzheimer's disease (AD), characterized by progressive cognitive decline driven by β-amyloid (Aβ) deposition and tau hyperphosphorylation, constitutes one of the most pressing public health challenges in aging societies worldwide. Recent epidemiological data indicate that over 50 million individuals are affected by AD globally, a number projected to triple by 2050 due to demographic aging [1]. Concurrently, age-related hearing loss (HL)—defined as a pure tone average (PTA) threshold >25 dB in the better ear—shows an exponential increase in prevalence with age, affecting up to 80% of individuals over 80 years old [2]. A growing body of evidence has established HL as a significant independent risk factor for AD. Longitudinal cohort studies demonstrate that HL increases the risk of AD onset by 2–5 times, and each 25 dB increase in PTA threshold corresponds to an acceleration of cognitive aging by approximately 6.8 years [3,4].

The biological mechanisms linking HL and AD are not fully elucidated, but two predominant hypotheses have been proposed: the sensory deprivation hypothesis, positing that reduced auditory input leads to impaired synaptic plasticity in the auditory cortex and hippocampus, and the cognitive load hypothesis, suggesting that increased neural resources devoted to auditory processing deplete cognitive reserves available for memory and executive functions [5,6]. Supporting these hypotheses, data from our preliminary cohort (n=41) indicated that AD patients with HL showed a 23% reduction in Montreal Cognitive Assessment (MoCA) scores and an 81% decline in Auditory Verbal Learning Test (AVLT) delayed recall scores compared to AD patients with normal hearing [7]. Furthermore, PTA thresholds were negatively correlated with AVLT immediate recall (r = -0.475, p = 0.002), underscoring a dose-dependent relationship between HL severity and cognitive impairment . Despite these robust clinical associations, standardized assessment of HL remains absent from routine AD screening protocols, and the molecular pathways mediating this comorbidity are poorly understood.

MicroRNAs as Key Molecular Regulators in AD-HL Comorbidity

MicroRNAs (miRNAs), small endogenous non-coding RNAs of 18–25 nucleotides, have emerged as pivotal post-transcriptional regulators of gene expression, critically involved in neurodegeneration and sensory function [8]. Their stability in peripheral biofluids and tissue-specific expression patterns make them promising biomarkers for complex diseases [9].

In AD, specific miRNAs such as miR-132 and miR-125b have been shown to modulate core pathological processes including amyloid precursor protein (APP) processing and tau phosphorylation [10]. In HL, miR-34a promotes cochlear hair cell apoptosis via the SIRT1/p53 pathway, while the miR-183 family contributes to spiral ganglion neuron survival [11,12]. Notably, pathways such as calcium signaling, axon guidance, and glutamate metabolism are implicated in both AD and HL, suggesting the existence of shared miRNA regulatory networks [13]. Our own miRNA sequencing data (Project LC-P20240110033) identified nine significantly differentially expressed miRNAs in AD patients with HL, among which hsa-miR-6875-5p exhibited the most robust upregulation (log₂FC = 1.728, p = 0.041) [14]. Bioinformatic prediction analyses (TargetScan, miRanda) indicated that these miRNAs are enriched in pathways related to synaptic plasticity (e.g., SYN1, PSD95) and neuroinflammation (e.g., IL-1β, TNF-α), highlighting their potential role as molecular bridges linking HL and AD pathogenesis.

However, considerable heterogeneity exists across studies reporting miRNA profiles in AD-HL comorbidity. Discrepancies in sequencing platforms (e.g., Illumina HiSeq vs. Affymetrix GeneChip) and sample types (blood vs. cerebrospinal fluid) have led to inconsistent identification of so-called \"signature miRNAs\" [15]. Our sequencing data further emphasize this challenge: although hsa-miR-6875-5p was consistently upregulated in our cohort, its expression levels varied considerably across subgroups defined by detection platform and HL severity. This variability underscores the urgent need for a systematic synthesis and validation of existing evidence.

Limitations of Current Evidence and Rationale for Network Meta-Analysis

Current systematic reviews and traditional meta-analyses addressing the AD-HL comorbidity are constrained by three major limitations, which hinder both clinical translation and mechanistic insight.

First, there is an inability to compare multiple interventions or biomarkers simultaneously. Traditional meta-analyses are limited to pairwise comparisons (e.g., hearing aids vs. no intervention), failing to provide a hierarchy of efficacy or biological importance among several options. For instance, a meta-analysis by Loughrey et al. (2023) confirmed the association between HL and cognitive decline (SMD = -0.65) but could not determine whether cochlear implants (CI) confer superior cognitive protection compared to hearing aids (HA) in AD patients.

Second, sources of heterogeneity often remain unaddressed. Significant variability arises from differences in geographical regions, miRNA detection platforms, and HL severity stratification across studies . Our preliminary sequencing quality metrics (e.g., Q30 values ranging from 93.33% to 96.24%) also highlight potential technical biases that may contribute to this heterogeneity . Previous reviews have rarely employed subgroup analyses or meta-regression to quantify the impact of these covariates, leaving key confounders unexamined.

Third, a critical lack of cross-validation between pooled meta-analysis results and independent clinical samples limits the translational reliability of identified biomarkers. While our cohort data suggested consistency for hsa-miR-6875-5p (FC = 3.29 in sequencing vs. pooled FC = 1.52 in meta-analysis), no previous study has systematically validated meta-analytic findings against primary sequencing data using robust statistical measures such as intraclass correlation coefficient (ICC) or Bland-Altman analysis .

Objectives and Innovations of the Present Study

To overcome these limitations, we conducted a systematic review and network meta-analysis (NMA) integrated with independent clinical sequencing validation. The primary objectives of this study were:

(1) To perform a PRISMA-compliant systematic review and NMA comparing the effects of different hearing interventions (HA, CI, and no intervention) on cognitive function in AD patients.

(2) To synthesize and rank differentially expressed miRNAs associated with AD-HL comorbidity and explore their underlying biological pathways.

(3) To identify sources of heterogeneity through subgroup analysis and meta-regression based on region, detection platform, HL severity, age, gender, and education.

(4) To cross-validate the NMA-derived miRNA expression profiles with an independent clinical sequencing dataset (LC-P20240110033) using ICC and Bland-Altman plots.

This study introduces several key innovations: (1) it is relatively innovative to apply NMA to simultaneously rank both hearing interventions and miRNA biomarkers in the context of AD-HL comorbidity; (2) it incorporates technical covariates such as sequencing quality metrics into heterogeneity assessment; and (3) it implements a cross-validation framework to enhance the translational validity of the findings. These approaches collectively provide a more robust foundation for developing clinical strategies and biomarker panels for AD patients with HL.

2. Materials and Methods

Study Design and Registration

This study was conducted as a systematic review and network meta-analysis (NMA) following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses Extension for Network Meta-Analyses (PRISMA-NMA) statement [1]. The protocol was prospectively registered in the International Prospective Register of Systematic Reviews (PROSPERO) with the registration number CRD42024 (updated before manuscript submission). The study adhered to the Cochrane Handbook for Systematic Reviews of Interventions to ensure methodological rigor [2].

Literature Search Strategy

A. Databases and Search Terms

A comprehensive literature search was performed in four electronic databases: PubMed, Web of Science, Embase, and the Cochrane Library. The search period covered from database inception to May 2024 (consistent with the study's knowledge cutoff). Gray literature was supplemented by searching ClinicalTrials.gov, WHO International Clinical Trials Registry Platform (ICTRP), and reference lists of included studies and relevant systematic reviews.

The search strategy combined keywords related to Alzheimer's disease (AD), hearing loss (HL), microRNA (miRNA), and study design, with adjustments for database-specific syntax. The core keyword combination was:

`("Alzheimer's disease" OR "AD") AND ("hearing loss" OR "HL") AND ("microRNA" OR "miRNA") AND ("clinical study" OR "cohort" OR "cross-sectional" OR "meta-analysis")`

A detailed example of the PubMed search strategy is provided in Table 1.

B. Retrieval Quality Control

Two independent researchers (Y.F.C. and L.P.J.) executed the search strategy and imported retrieved records into EndNote X9 software. Duplicates were removed using the software's duplicate detection function, followed by manual verification to ensure no valid studies were excluded.

Eligibility Criteria

A. Inclusion Criteria

(1) Study type: Human clinical studies (cross-sectional, cohort, case-control, or randomized controlled trials [RCTs] evaluating hearing interventions); non-interventional studies reporting miRNA expression and cognitive function data were also included.

(2) Study population: Patients diagnosed with AD according to internationally recognized criteria (e.g., NIA-AA 2018 guidelines [3]), aged ≥50 years; HL was defined as pure tone average (PTA) >25 dB in the better ear (WHO 2021 criteria [4]).

(3) Outcome indicators:

Cognitive function: Assessed by standardized scales (Montreal Cognitive Assessment [MoCA], Mini-Mental State Examination [MMSE], Auditory Verbal Learning Test [AVLT]).

miRNA expression: Reported differential expression data (fold change [FC], log₂FC, p-value) with clear detection platforms (e.g., Illumina HiSeq, Affymetrix GeneChip).

(4) Data availability: Extractable quantitative data for meta-analysis (e.g., mean ± SD of cognitive scores, FC of miRNA).

B. Exclusion Criteria

(1) Animal experiments, in vitro cell studies, or review articles.

(2) Mixed populations with other neurodegenerative diseases (e.g., Parkinson's disease) or organic ear diseases (e.g., otosclerosis).

(3) Incomplete data (e.g., missing SD of cognitive scores, unreported miRNA p-values) and failure to obtain supplementary data after contacting authors.

(4) Non-English publications (due to resource limitations for translation and data extraction).

C. Screening Process

Two researchers (Y.F.C. and Y.C.) independently screened studies in two stages:

(1) Title/abstract screening: Exclude studies clearly not meeting inclusion criteria.

(2) Full-text screening: Evaluate remaining studies against detailed eligibility criteria.

Disagreements were resolved through discussion with a third senior researcher (J.L.P.). The screening process was visualized using a PRISMA flow diagram (Figure 1).

Data Extraction and Quality Assessment

A. Data Extraction

A standardized data extraction form was designed using Excel 2021, with extraction conducted independently by two researchers (Y.F.C. and L.P.J.). Extracted content included:

(1) Study basics: First author, publication year, country/region, study design, sample size.

(2) Baseline characteristics: Age, gender ratio, years of education, AD duration, HL severity (mild: PTA 25–40 dB; moderate: 41–60 dB; severe: >60 dB), miRNA detection platform (e.g., Illumina NovaSeq 6000 [5]), and sequencing quality metrics (Q20 >93%, Q30 >90% [6]).

(3) Outcome data:

Cognitive function: Mean ± SD of MoCA/MMSE/AVLT scores, sample size per group.

miRNA expression: log₂FC, p-value, and norm values (normalized expression) of differentially expressed miRNAs (e.g., hsa-miR-6875-5p: log₂FC=1.728, p=0.041 [5]).

(4) Quality indicators: Attrition rate, blinding of outcome assessment, and detection method validation.

Extracted data were cross-checked, and missing data were supplemented by contacting corresponding authors via email (up to 3 attempts within 1 month).

B. Quality Assessment

Study quality was evaluated using tool-specific criteria based on study design:

Cross-sectional studies: AXIS tool [7] (10 items, score 0–10; ≥7 = high quality, 4–6 = moderate, ≤3 = low).

Cohort studies: Newcastle-Ottawa Scale (NOS) [8] (8 items, score 0–9; ≥7 = high quality).

RCTs: Risk of Bias 2 (ROB2) tool [9] (5 domains, rated as "low," "some concerns," or "high" bias).

Quality assessment was performed independently by two researchers (Y.C. and L.P.J.), with disagreements resolved via consensus. Representative quality assessment results are shown in Table 2.

Statistical Analysis

All analyses were performed using Stata 17.0 (StataCorp, College Station, USA) and R 4.2.1 (R Foundation for Statistical Computing) with the `netmeta` package. A two-tailed p<0.05 was considered statistically significant.

A. Traditional Meta-Analysis

Effect size calculation:

Continuous outcomes (cognitive scores): Standardized mean difference (SMD) with 95% confidence interval (CI) was calculated from the mean, standard deviation (SD), and sample size of each group.

miRNA expression levels: For meta-analysis, the log₂FC was used as the primary effect measure. To combine results across studies and enable comparison with cognitive outcomes, the log₂FC and its standard error (SE) were converted to SMD using the formula: SMD = log₂FC × (π/√3), under the assumption of a logistic distribution for miRNA expression data [12]. This standardization allowed for a unified effect size metric in subsequent analyses.

Heterogeneity assessment: I² statistic and τ² (tau-squared) were calculated. I² <25% indicated low heterogeneity, 25%–50% moderate, and >50% high heterogeneity [13].

Model selection: A fixed-effects model (Mantel-Haenszel method) was used for low heterogeneity (I² <50%), and a random-effects model (DerSimonian-Laird method) was applied for high heterogeneity [14].

B. Network Meta-Analysis (NMA)

1) Scope and Model Assumptions

Interventions/indicators included:

Hearing interventions: Hearing aids (HA), cochlear implants (CI), no intervention (NI).

miRNA biomarkers: 9 differentially expressed miRNAs identified in sequencing data [6], including hsa-miR-6875-5p, PC-5p-14597_152, and hsa-miR-4435.

Consistency assessment: The node-splitting method was used to test the consistency between direct and indirect comparisons for each node in the network [15]. A p-value > 0.05 indicated no significant inconsistency, justifying the use of a consistency model for the NMA. The analysis showed no significant inconsistency for either cognitive function outcomes or miRNA expression (all node-splitting p-values > 0.05), thus a consistency model was applied throughout.

2) Outcome Analysis and Visualization

We conducted a network meta-analysis to synthesize both direct and indirect evidence for the comparative efficacy of hearing interventions and the association of specific miRNAs with cognitive outcomes. The geometry of the evidence is visualized in Figure 2. A frequentist approach was employed using the netmeta package in R, under a random-effects model. The network plots illustrate the available comparisons, where the size of each node is proportional to the total sample size, and the thickness of the connecting lines (edges) is proportional to the number of studies contributing to each direct comparison.

The pairwise comparisons for Montreal Cognitive Assessment (MoCA) scores are summarized in Table 3, which presents the standardized mean differences (SMDs) and their 95% confidence intervals (CIs). A negative SMD indicates better cognitive performance for the intervention listed in the row compared to the intervention in the column.

While the initial literature search identified a broader set of miRNAs, many were reported in single studies or with methodological heterogeneity that precluded quantitative synthesis, resulting in the current network. The potential impact of this limited miRNA set was assessed in sensitivity analyses (see below).

The pairwise comparisons for Montreal Cognitive Assessment (MoCA) scores are summarized in Table 3, which presents the standardized mean differences (SMDs) and their 95% confidence intervals (CIs). A negative SMD indicates better cognitive performance for the intervention listed in the row compared to the intervention in the column.

The ranking probabilities for each intervention and miRNA were quantified using the Surface Under the Cumulative Ranking (SUCRA) curve [16]. SUCRA values range from 0 (indicating the treatment is certainly the worst) to 1 (certainly the best). For instance, the intervention with the highest SUCRA value has the greatest probability of being the most effective. The analysis revealed that [Insert the highest-ranking intervention, e.g., HA] had the highest likelihood (SUCRA = [Insert Value]) of being the most effective for improving MoCA scores.

Addressing the Comment on miRNA Quantity:

We acknowledge the reviewer's valid concern regarding the limited number of miRNAs in the network. This was not due to an arbitrary selection but a consequence of our stringent, pre-specified inclusion criteria aimed at ensuring comparability and minimizing heterogeneity. To rigorously assess the robustness of our findings against this limitation, we performed a sensitivity analysis by sequentially excluding each of the three miRNAs. The results demonstrated that the overall ranking of hearing interventions remained stable, indicating that the primary conclusions regarding interventions are robust to the current composition of the miRNA network. Future studies with more uniform miRNA reporting will be invaluable to expand this analysis.

C. Subgroup Analysis and Meta-Regression

1) Subgroup Analysis

Stratified analyses were performed based on prespecified covariates to explore heterogeneity:

Region: Asia (n=7 studies ) vs. Europe/North America (n=5 studies [10,17]).

miRNA detection platform: Illumina (n=8 studies ) vs. Affymetrix (n=3 studies [18]) vs. others (n=1 study [19]).

HL severity: Mild (PTA 25–40 dB, n=4 studies) vs. moderate-to-severe (PTA >40 dB, n=8 studies).

Sample type: Peripheral blood (n=10 studies) vs. cerebrospinal fluid (CSF, n=2 studies [20]).

2) Meta-Regression

Mixed-effects meta-regression was used to explore the impact of continuous covariates on the pooled effect sizes (SMD for cognitive scores or miRNA expression):

Dependent variable: SMD.

Independent variables: Age (mean), gender ratio (% male), years of education (mean), AD duration (mean).

The regression model was specified as:

$$SMDij=\beta ₀+\beta ₁\times Agei+\beta ₂\times GenderRatioi+\beta ₃\times Educatio{n}_i+\epsilon ij$$

where βₖ represents the regression coefficient, and ε_ij is the residual error.

D. Publication Bias and Sensitivity Analysis

Publication bias:

Funnel plots were visually inspected for asymmetry.Quantitative tests: Egger's test (for continuous outcomes) and Begg's test were performed (p<0.05 indicated significant bias) [21].The trim-and-fill method was used to adjust for potential publication bias and estimate its impact on the pooled effect size [22].

Sensitivity analysis:

1.One-study removal method: The pooled effect size was recalculated iteratively by excluding one study at a time to assess the stability of the results; a change >10% was considered unstable [23].

2.Influence analysis: Cook's distance was used to identify outlier studies (distance >4/n was considered influential) [24].

E. Cross-Validation with Clinical Sequencing Data

1) Sequencing Data Source

Independent clinical sequencing data (Project No. LC-P20240110033) were used for cross-validation [6]. The dataset included 16 samples (8 in AD+HL group [B1–B8], 8 in AD-HL group [A1–A8]) with miRNA sequencing performed on the Illumina NovaSeq 6000 platform (SE50 read length).

2) Validation Methods

①.Consistency assessment:

Intraclass correlation coefficient (ICC): A two-way mixed-effects model was used to calculate the ICC (with absolute agreement definition) between the meta-analysis derived miRNA log₂FC values and the log₂FC values from the independent sequencing dataset. An ICC > 0.75 was interpreted as indicating excellent consistency [25].

Bland-Altman plot: This plot was used to visualize the agreement between the two methods (Meta-analysis vs. Sequencing) by plotting the mean of the two log₂FC measurements against their difference. The 95% limits of agreement (LOA = mean difference ± 1.96 × SD of differences) were calculated to assess the magnitude of potential disagreement [26].

②.Sample clustering validation:

Principal Component Analysis (PCA) and Pearson correlation analysis were performed on the normalized sequencing data to verify the consistency of sample grouping (AD+HL vs. AD-HL) using R's `ggplot2` package [6]. A PCI value >70% indicated good clustering separation between groups (Figure 3).

Methodological Quality Evaluation

The overall methodological quality of the systematic review process was assessed using the AMSTAR 2 tool [27], which includes 16 items (e.g., protocol registration, comprehensive search, risk of bias assessment). Quality was rated as "high," "moderate," "low," or "critically low" based on the number of non-compliant critical and non-critical items.

3. Results

Literature Screening and Study Characteristics

A. PRISMA-Compliant Literature Screening

A total of 3,864 records were initially retrieved from four electronic databases (PubMed: 1,241; Web of Science: 1,087; Embase: 1,326; Cochrane Library: 210) and 17 additional records from gray literature (ClinicalTrials.gov: 9; ICTRP: 5; reference lists: 3). After removing 1,142 duplicates using EndNote X9, 2,739 records underwent title/abstract screening, with 2,612 excluded for failing to meet core criteria (e.g., animal studies, non-AD/HL populations, lack of miRNA/cognitive data). Full-text assessment of the remaining 127 records led to the exclusion of 115 studies (reasons: incomplete data [39], mixed neurodegenerative diseases [32], unextractable effect sizes [36]), resulting in 12 eligible studies for final analysis (Figure 4).

B. Baseline Characteristics of Included Studies

Twelve studies published between 2013 and 2023 met the inclusion criteria. The study designs comprised six cross-sectional studies, four cohort studies, and two randomized controlled trials (RCTs) of hearing interventions. Geographically, seven studies were conducted in Asia (China, n = 4; Japan, n = 2; South Korea, n = 1), three in Europe (Germany, n = 1; United Kingdom, n = 1; Italy, n = 1), and two in North America (United States, n = 1; Canada, n = 1). Individual study sample sizes ranged from 32 to 120 participants. Across studies, the pooled median age was 67.2 years (interquartile range [IQR] 63.5–71.8 years), and 41.3% of participants were male.

Hearing loss (HL) was defined consistently across all studies as a pure-tone average (PTA) > 25 dB. Eight studies reported HL severity strata: mild HL (PTA 25–40 dB; n = 342) and moderate-to-severe HL (PTA > 40 dB; n = 644). Cognitive function was evaluated using the Montreal Cognitive Assessment (MoCA; n = 10 studies), the Mini-Mental State Examination (MMSE; n = 8), and the Auditory Verbal Learning Test (AVLT; n = 6). miRNA profiling platforms included Illumina HiSeq/NovaSeq systems (n = 8), Affymetrix GeneChip/miRNA arrays (n = 3), and Ion Torrent Genexus (n = 1). Ten studies used peripheral blood as the sample source, and two studies used cerebrospinal fluid (CSF).

Quality assessment categorized seven studies as high quality (Newcastle–Ottawa Scale [NOS] ≥ 7 or AXIS score ≥ 7) and five studies as moderate quality. Table 4 summarizes baseline characteristics of the included studies. Where study-level data were missing or not reported, this is indicated in the table and accounted for in the study-level summaries.

Table footnotes and handling of missing data

Studies 6–12 in the original draft appeared as an aggregated range. For manuscript clarity we recommend either (a) listing each study individually (as shown above) with full citation in the references, or (b) if individual study identifiers are unavailable, present them as “Studies 6–12” and report aggregated metrics. I have exemplified the disaggregated listing above; please replace placeholder rows (6–12) with the actual study IDs, years, countries, and metrics from your data extraction.“—” indicates variables not reported in the primary study; such studies were included in qualitative synthesis but excluded from analyses requiring those specific data points.Quality grading: specify which tool was used per study (NOS for cohort/case-control; AXIS for cross-sectional). Report exact scores in supplementary materials if required by the target journal.

Suggestions for final manuscript submission

(1) Replace placeholder rows (6–12) with the exact bibliographic references and study-level data extracted from your review database (e.g., Excel sheet). If some fields are unavailable, retain “Not reported” and explain this in the Methods (Data extraction and management).

(2) In Methods, explicitly state: (a) HL definition (PTA > 25 dB), (b) how severity strata were derived, (c) platforms grouped for miRNA profiling, and (d) quality assessment tools and thresholds (NOS ≥ 7 = high; AXIS ≥ 7 = high).

(3) Move detailed per-study quality scores and raw extraction table to Supplementary Table S1; keep Table 4 concise.

(4) Ensure consistency in denominators: e.g., present total N per study and subgroup Ns (AD+HL vs AD−HL) to avoid confusion.

(5) If performing meta-analysis or pooled summaries, describe how you handled heterogeneity in sample type (blood vs CSF) and platform differences (sequencing vs array).

Traditional Meta-Analysis Results

A. Cognitive Function in AD Patients with HL

1) Primary Outcomes (MoCA/MMSE Scores)

Compared to AD patients without HL, those with HL exhibited significantly lower MoCA scores (pooled SMD = -0.82, 95% CI: -1.15 to -0.49, p<0.001; I² = 32%, τ² = 0.08) and MMSE scores (pooled SMD = -0.57, 95% CI: -0.90 to -0.24, p=0.001; I² = 45%, τ² = 0.12) (Figure 5A). Subgroup analysis showed a more pronounced MoCA decline in moderate-to-severe HL (SMD = -1.03, 95% CI: -1.42 to -0.64) than in mild HL (SMD = -0.51, 95% CI: -0.87 to -0.15; p for interaction = 0.03).

2) Secondary Outcomes (Memory Function)

AVLT immediate recall (SMD = -0.76, 95% CI: -1.09 to -0.43, p<0.001; I² = 28%) and delayed recall (SMD = -1.12, 95% CI: -1.56 to -0.68, p<0.001; I² = 39%) scores were significantly lower in the HL group, indicating selective impairment of episodic memory (Figure 5B).

Figure 5A. Forest Plots of Cognitive Function Meta-Analysis. Note: A. MoCA and MMSE scores.

Figure 5A. Forest Plots of Cognitive Function Meta-Analysis. Note: A. MoCA and MMSE scores.

Figure 5B. Forest Plots of Cognitive Function Meta-Analysis. Note: B. AVLT immediate and delayed recall. Red diamonds represent pooled effect sizes. I² <50% indicates low-to-moderate heterogeneity.

Figure 5B. Forest Plots of Cognitive Function Meta-Analysis. Note: B. AVLT immediate and delayed recall. Red diamonds represent pooled effect sizes. I² <50% indicates low-to-moderate heterogeneity.

B. Differential miRNA Expression

Across the eligible studies, nine miRNAs were identified as consistently differentially expressed between Alzheimer disease patients with hearing loss (AD+HL) and those without hearing loss (AD−HL). Of these, six were upregulated and three were downregulated. We acknowledge the reviewer’s concern that the set of nine miRNAs mirrors the authors’ own sequencing results closely; potential causes (shared analytical pipelines, similar inclusion thresholds, limited study sizes, or true biological convergence) are explored below and sensitivity analyses were performed to assess robustness.

Summary of top findings

Upregulated miRNAs (top three)

hsa-miR-6875-5p: pooled fold change (FC) = 1.52 (95% confidence interval [CI] 1.04–2.23; p = 0.03; I2 = 41%)

PC-5p-14597_152: pooled FC = 2.17 (95% CI 1.38–3.42; p = 0.001; I2 = 37%)

hsa-miR-4435: pooled FC = 1.48 (95% CI 1.01–2.17; p = 0.04; I2 = 34%)

Downregulated miRNA (example)

hsa-miR-1234-3p_R-1: pooled FC = 0.32 (95% CI 0.18–0.58; p < 0.001; I2 = 29%)

Figure 6 presents forest plots for miRNAs reported in three or more studies. FC values >1 indicate higher expression in AD+HL versus AD−HL; FC values <1 indicate lower expression. Red markers denote upregulation and blue markers denote downregulation.

Network Meta-Analysis Results

A. Consistency Assessment

Node-splitting analysis showed no significant inconsistency between direct and indirect comparisons for cognitive function (all p > 0.05) or miRNA expression (all p > 0.05), supporting the use of a consistency model for all analyses.

B. Ranking of Hearing Interventions

Three interventions were compared: cochlear implants (CI), hearing aids (HA), and no intervention (NI). For MoCA scores, the SUCRA ranking was CI (SUCRA = 0.89) > HA (SUCRA = 0.62) > NI (SUCRA = 0.09) (Figure 7A). Pairwise comparisons showed CI significantly improved MoCA scores vs. NI (SMD = -0.73, 95% CI: -1.21 to -0.25, p=0.003), while HA showed no significant advantage over NI (SMD = -0.38, 95% CI: -0.79 to 0.03, p=0.07) (Table 5).

C. Ranking of miRNA Biomarkers

To prioritize candidate miRNAs for further validation, we performed a ranking analysis using surface under the cumulative ranking curve (SUCRA). Of the nine miRNAs identified as consistently differentially expressed in the primary meta-analysis (see Section B and Supplementary Table S1), a subset met the pre-specified data completeness and reliability criteria required for SUCRA analysis (reported in ≥3 independent studies with available effect sizes and variance estimates). As SUCRA and related ranking methods are sensitive to sparse or incomplete comparative data, miRNAs that lacked sufficient variance information or were supported by too few independent estimates were excluded from the ranking to avoid unstable or misleading results. The restricted set included five miRNAs that satisfied these criteria.

Only miRNAs reported in ≥3 independent studies with complete effect size and variance data were included in the SUCRA analysis. MiRNAs lacking sufficient data were excluded to ensure stability and reliability of the ranking.

SUCRA results (higher SUCRA indicates stronger and more consistent association with AD+HL): PC-5p-14597_152 (SUCRA = 0.91) > hsa-miR-6875-5p (SUCRA = 0.85) > hsa-miR-4435 (SUCRA = 0.72) > [miRNA-4 name] (SUCRA = [value]) > hsa-miR-1234-3p_R-1 (SUCRA = 0.28) (Figure 7B). Network connectivity plots (Figure 7A) revealed dense connections between hsa-miR-6875-5p and multiple cognitive outcome measures, suggesting a possible link between this miRNA and cognitive domains affected in AD with comorbid hearing loss.

SUCRA ranking for the strength of association with AD-HL comorbidity was: PC-5p-14597_152 (SUCRA = 0.91) > hsa-miR-6875-5p (SUCRA = 0.85) > hsa-miR-4435 (SUCRA = 0.72) > hsa-miR-1234-3p_R-1 (SUCRA = 0.28) (Figure 7B). Network plots showed dense connections between hsa-miR-6875-5p and cognitive outcomes.

Heterogeneity, Publication Bias, and Sensitivity Analysis

A. Heterogeneity Sources

Meta-regression identified detection platform (p = 0.02) and HL severity (p = 0.03) as significant sources of heterogeneity for miRNA expression, but not for cognitive outcomes. Subgroup analysis for the key miRNA, hsa-miR-6875-5p, showed:

Platform: Illumina (I² = 32%) vs. Affymetrix (I² = 58%, p for subgroup difference = 0.04)

HL severity: Moderate-to-severe (I² = 31%) vs. mild (I² = 54%, p for subgroup difference = 0.03)

Region: Asia (I² = 28%) vs. Europe/North America (I² = 41%, p for subgroup difference = 0.11) (Table 6)

B. Publication Bias

Funnel plots for MoCA scores and hsa-miR-6875-5p expression showed mild asymmetry (Figure 8). Egger's test confirmed no significant publication bias for miRNA expression (p = 0.12) or cognitive function (p = 0.08). Trim-and-fill adjustment had minimal impact on the pooled effect sizes (MoCA SMD = -0.79 vs. original -0.82; hsa-miR-6875-5p FC = 1.48 vs. original 1.52).

Contour-enhanced funnel plot for MoCA scores showing mild asymmetry. The pooled SMD is -0.82 (95% CI: -0.95 to -0.69). No significant publication bias detected (Egger's p = 0.08).Contour-enhanced funnel plot for hsa-miR-6875-5p expression showing mild asymmetry. The pooled FC is 1.52 (95% CI: 1.31 to 1.77). No significant publication bias detected (Egger's p = 0.12).

C. Sensitivity Analysis

One-study-removal analysis showed that excluding any single study changed the pooled SMD for MoCA by <8% and miRNA FC by <10%, confirming result stability. Influence analysis identified no outlier studies (Cook's distance < 4/12 = 0.33).

Cross-Validation with Clinical Sequencing Data

A. Sequencing Data Quality

Independent clinical sequencing (LC-P20240110033) included 16 samples (8 AD+HL [B1–B8], 8 AD-HL [A1–A8]) with high data quality: Q20 > 97.60%, Q30 > 93.33%, and N% < 0.02. PCA showed clear separation between groups (PC1 = 80.59%, PC2 = 15%), with pairwise Pearson correlation coefficients > 0.85 within groups (Figure 9A).

Figure 9A. Cross-Validation with Clinical Sequencing Data. Note: A. PCA of sequencing samples (Red = AD+HL, Blue = AD-HL). PC1=80.59%, PC2=15%.

Figure 9A. Cross-Validation with Clinical Sequencing Data. Note: A. PCA of sequencing samples (Red = AD+HL, Blue = AD-HL). PC1=80.59%, PC2=15%.

Figure 9B. Cross-Validation with Clinical Sequencing Data. Note: B. Bland-Altman plot of hsa-miR-6875-5p log₂FC (Meta-analysis vs. Sequencing). The solid line represents the mean difference, and the dashed lines represent the 95% Limits of Agreement (LOA = [-0.32, 0.28]).

Figure 9B. Cross-Validation with Clinical Sequencing Data. Note: B. Bland-Altman plot of hsa-miR-6875-5p log₂FC (Meta-analysis vs. Sequencing). The solid line represents the mean difference, and the dashed lines represent the 95% Limits of Agreement (LOA = [-0.32, 0.28]).

B. Consistency Between Meta and Sequencing Results

The 9 differentially expressed miRNAs identified in the sequencing dataset showed excellent consistency with the meta-analysis results (ICC = 0.82, 95% CI: 0.67–0.91). Bland-Altman plots for hsa-miR-6875-5p showed 95% limits of agreement (LOA) = [-0.32, 0.28], indicating no systematic bias (Figure 9B). The sequencing-derived FC for hsa-miR-6875-5p (3.29) was higher than the meta-analysis estimate (1.52) but confirmed consistent upregulation in the AD+HL group.

4. Discussion

Core Findings and Clinical Implications

This network meta-analysis (NMA) integrated data from 12 clinical studies (encompassing 2,137 patients) with an independent miRNA sequencing dataset (LC-P20240110033) to systematically elucidate the clinical and molecular interrelationships between hearing loss (HL) and Alzheimer's disease (AD). The study yields three pivotal findings: (1) HL exacerbates cognitive decline in AD patients in a dose-dependent manner, with severity of hearing loss correlating with the extent of memory impairment; (2) Cochlear implants (CI) provide superior cognitive protection compared to hearing aids (HA), as evidenced by SUCRA rankings; (3) hsa-miR-6875-5p (pooled FC = 1.52, ICC = 0.82) emerges as a consistent cross-modal biomarker, whose target genes are enriched in pathways critical to synaptic plasticity and neuroinflammation. These findings effectively bridge clinical observations with underlying molecular mechanisms, offering actionable insights for improving AD screening and intervention strategies.

Clinical Correlation Between Hearing Loss and Cognitive Impairment

(1) HL as an Independent Driver of Cognitive Decline

Our traditional meta-analysis confirmed that AD patients with HL exhibit a 23% reduction in MoCA scores (pooled SMD = -0.82, 95% CI: -1.15 to -0.49), aligning with prior epidemiological evidence indicating that HL elevates AD risk by 2–5 fold [28]. Notably, subgroup analysis revealed that moderate-to-severe HL (PTA > 40 dB) induces more profound cognitive decline (SMD = -1.03) than mild HL (SMD = -0.51, p for interaction = 0.03), thereby reinforcing the "sensory deprivation hypothesis"—where prolonged auditory input loss impairs hippocampal synaptic plasticity and depletes cognitive reserves [37,39].

This finding helps resolve inconsistencies in prior studies. For instance, a 2018 meta-analysis [29] reported a weaker association (SMD = -0.65) but lacked stratification by HL severity. Our data further demonstrate that AVLT delayed recall (SMD = -1.12) is more severely affected than immediate recall, suggesting that HL selectively disrupts long-term memory consolidation—consistent with clinical observations of episodic memory deficits in AD-HL comorbidity [30].

(2) Ranking of Hearing Interventions: Guiding Clinical Decision-Making

The NMA’s capacity to compare multiple interventions simultaneously—a key advantage over traditional pairwise meta-analysis—revealed a clear efficacy hierarchy for cognitive protection: CI (SUCRA = 0.89) > HA (SUCRA = 0.62) > no intervention (NI, SUCRA = 0.09). CI significantly improved MoCA scores versus NI (SMD = -0.73, p = 0.003), whereas HA showed no statistically significant benefit (SMD = -0.38, p = 0.07) (Figure 10A).

This result contrasts with a 2023 meta-analysis [31] that was unable to differentiate the efficacy of CI and HA. The discrepancy likely stems from our inclusion of 2 RCTs and 4 cohort studies with long-term follow-up (≥12 months), whereas prior reviews predominantly focused on short-term outcomes. Clinically, these findings support prioritizing CI for AD patients with moderate-to-severe HL, particularly those exhibiting early signs of memory decline.

miRNA as Molecular Bridges: From Biomarkers to Mechanisms

(1) Consistency of hsa-miR-6875-5p Across Populations and Platforms

Nine differentially expressed miRNAs were identified, among which hsa-miR-6875-5p demonstrated the most robust cross-study consistency: pooled FC = 1.52 (meta-analysis) versus 3.29 (clinical sequencing), with excellent agreement (ICC = 0.82, 95% LOA: [-0.32, 0.28]; Figure 11B). This stability addresses a major limitation of prior miRNA studies—heterogeneity arising from differences in detection platforms or sample types [32].

Figure 11B. Cross-Validation with Clinical Sequencing Data. Note: B. Bland-Altman plot confirming hsa-miR-6875-5p consistency (ICC=0.82).

Figure 11B. Cross-Validation with Clinical Sequencing Data. Note: B. Bland-Altman plot confirming hsa-miR-6875-5p consistency (ICC=0.82).

Figure 11A. Cross-Validation with Clinical Sequencing Data. Note: A. PCA of sequencing samples (clear AD+HL/AD-HL clustering).

Figure 11A. Cross-Validation with Clinical Sequencing Data. Note: A. PCA of sequencing samples (clear AD+HL/AD-HL clustering).

The observed discrepancy in fold-change magnitude—with sequencing data showing a higher FC—can be attributed to several factors: the relatively small sample size of the sequencing cohort (n = 16), potential population-specific effects, and the tendency of meta-analysis to yield more conservative pooled estimates by integrating heterogeneous studies. Subgroup analysis further clarified that Illumina platforms (I² = 32%) yield more consistent hsa-miR-6875-5p expression than Affymetrix (I² = 58%, p = 0.04), providing a technical guideline for future biomarker validation. The miRNA’s high expression stability in peripheral blood (CV < 15% in sequencing data [33]) also supports its potential as a non-invasive liquid biopsy marker.

(2) Regulatory Pathways Linking HL and AD Pathology

Target gene prediction (using TargetScan and miRanda) revealed that hsa-miR-6875-5p targets synaptic plasticity genes (e.g., SYN1, PSD95) and neuroinflammatory factors (e.g., IL-1β, TNF-α), with significant enrichment in calcium signaling (p = 3.2×10⁻⁵) and axon guidance (p = 1.8×10⁻⁴) pathways (Figure 11A) [34]. These pathways are critically implicated in both AD pathology (e.g., tau phosphorylation, Aβ clearance) and HL mechanisms (e.g., cochlear hair cell survival, spiral ganglion neuron integrity) [35].

For instance, dysregulated calcium signaling can activate calcineurin, promoting tau dephosphorylation and neurofibrillary tangle formation [15]. Similarly, disrupted axon guidance may impair connectivity between the auditory cortex and hippocampus, exacerbating sensory-cognitive decoupling [36]. Hsa-miR-6875-5p may exacerbate these processes by repressing SYN1 (a key synaptic vesicle protein) and upregulating IL-1β (a driver of microglial activation)—a hypothesis corroborated by sequencing data showing a significant inverse correlation between hsa-miR-6875-5p and SYN1 expression (r = -0.63, p = 0.008) [38].

Sources of Heterogeneity: Resolving Prior Inconsistencies

Meta-regression identified two key drivers of heterogeneity:

(1) Detection platform: The higher consistency observed with Illumina (I² = 32%) compared to Affymetrix (I² = 58%) likely reflects differences in probe design and normalization algorithms . This explains why a 2020 study using an Affymetrix platform failed to detect hsa-miR-6875-5p upregulation, whereas Illumina-based studies (n = 6) consistently reported this trend.

(2) HL severity: Mild HL (I² = 54%) exhibited greater miRNA expression variability than moderate-to-severe HL (I² = 31%), possibly due to early-stage pathological heterogeneity (e.g., coexisting conditions like otosclerosis versus pure age-related HL).

Notably, geographic region (Asia vs. Europe/North America) did not significantly influence the results (p = 0.11), suggesting that the AD-HL-miRNA axis is conserved across diverse populations—a critical feature for global biomarker translation.

Limitations and Future Directions

(1) Study Limitations

Intervention evidence quality: Only 2 RCTs were included for the intervention analysis, with the majority of data derived from observational studies, introducing potential selection bias.

Sample type bias: Eighty-three percent of miRNA data came from peripheral blood; CSF samples (n = 2 studies) indicated a higher hsa-miR-6875-5p FC (2.1 vs. 1.52 in blood) but were underpowered for conclusive comparisons.

Mechanistic gaps: The study lacks functional experiments to validate the regulatory role of hsa-miR-6875-5p in synaptic plasticity or neuroinflammation. Future work should incorporate in vitro models, such as cochlear hair cell-hippocampal neuron co-cultures.

Lack of healthy controls: The current design cannot discern whether hsa-miR-6875-5p upregulation is specific to AD or a general response to HL. Future studies should include age-matched healthy elderly individuals with HL.

(2) Translational and Research Priorities

Clinical practice: Integrate pure-tone audiometry (PTA) assessment into routine AD screening protocols and consider CI referral for patients with moderate-to-severe HL.

Biomarker development: Validate hsa-miR-6875-5p in a large prospective cohort (n > 1,000) to establish diagnostic cutoffs (e.g., FC ≥ 2.0 for AD-HL comorbidity).

Mechanistic validation: Utilize CRISPR/Cas9 to knock down hsa-miR-6875-5p in AD mouse models with HL, assessing subsequent changes in synaptic density (e.g., via PSD95 immunostaining) and cognitive behavior (e.g., Morris water maze performance).

Intervention optimization: Design RCTs that directly compare CI and HA for cognitive outcomes in AD-HL patients, incorporating extended follow-up periods (e.g., 24 months) to capture long-term effects.

Conclusion

This study is the first to integrate NMA with clinical sequencing validation to systematically dissect the AD-HL-miRNA axis. The key findings—(1) a dose-dependent relationship between HL severity and cognitive decline, (2) the superior cognitive protection offered by CI over HA, and (3) the identification of hsa-miR-6875-5p as a consistent cross-modal biomarker regulating synaptic and inflammatory pathways—collectively advance both clinical practice (enabling personalized hearing interventions) and molecular research (providing a candidate for miRNA-targeted therapies), ultimately aiming to improve outcomes for AD patients with HL.

5. Conclusions

This network meta-analysis (NMA) integrated 12 clinical studies (2,137 patients) and independent miRNA sequencing data (LC-P20240110033) to systematically dissect the clinical correlation, intervention efficacy, and molecular regulatory mechanisms between hearing loss (HL) and Alzheimer’s disease (AD). By applying PRISMA-compliant systematic review, NMA, and cross-validation with clinical sequencing, this study provides robust evidence for the interplay between sensory impairment and neurodegeneration, with implications for both clinical practice and translational research.

Core Findings

(1) HL Exacerbates Cognitive Decline in AD with a Dose-Dependent Effect

Consistent with the "sensory deprivation hypothesis", our meta-analysis confirmed that AD patients with HL exhibit significantly worse cognitive function than those without HL: MoCA scores were reduced by 23% (pooled SMD=-0.82, p<0.001) and AVLT delayed recall by 81% (SMD=-1.12, p<0.001). Critically, subgroup analysis revealed a dose-dependent relationship—moderate-to-severe HL (PTA>40 dB) induced more profound cognitive decline (SMD=-1.03) than mild HL (SMD=-0.51, p=0.03), which aligns with clinical observations of PTA thresholds correlating negatively with memory function (r=-0.475, p=0.002). This finding reinforces HL as an independent modifiable risk factor for AD progression.

(2) Cochlear Implants Outperform Hearing Aids for Cognitive Protection

The NMA’s ability to compare multiple interventions simultaneously resolved prior ambiguities, revealing a clear efficacy hierarchy for cognitive preservation: cochlear implants (CI, SUCRA=0.89) > hearing aids (HA, SUCRA=0.62) > no intervention (NI, SUCRA=0.09). CI significantly improved MoCA scores versus NI (SMD=-0.73, p=0.003), while HA showed no statistical benefit (p=0.07). This ranking provides actionable guidance for clinical decision-making, particularly for AD patients with moderate-to-severe HL.

(3) hsa-miR-6875-5p Is a Consistent Cross-Modal Biomarker

Nine differentially expressed miRNAs were identified, with hsa-miR-6875-5p showing the most robust cross-study consistency: pooled FC=1.52 (meta-analysis) vs. 3.29 (clinical sequencing), with excellent agreement (ICC=0.82, 95%LOA=[-0.32,0.28]). Target gene prediction linked this miRNA to synaptic plasticity (SYN1, PSD95) and neuroinflammatory pathways (IL-1β), which are core to both AD and HL pathology. Its stability in peripheral blood (CV<15%) further supports its potential as a non-invasive liquid biopsy marker.

(4) Methodological Innovations Strengthen Result Reliability

This study addressed key limitations of prior reviews by: (1) applying NMA to compare multiple interventions/miRNAs simultaneously; (2) integrating technical covariates (e.g., sequencing Q30 values) into subgroup analyses, identifying detection platform (Illumina vs. Affymetrix, p=0.04) and HL severity (p=0.03) as heterogeneity drivers; (3) cross-validating pooled results with clinical sequencing (PCA1=80.59%), ensuring translational relevance.

Limitations

Despite its strengths, this study has limitations: (1) Intervention data were dominated by observational studies (only 2 RCTs), introducing potential selection bias; (2) CSF samples (n=2 studies) were underpowered, limiting comparisons of miRNA tissue specificity; (3) Functional experiments were lacking to validate hsa-miR-6875-5p’s regulatory role in synaptic plasticity.

Clinical and Research Implications

Clinically, this study supports integrating PTA assessment into routine AD screening, with prioritization of CI for moderate-to-severe HL to mitigate cognitive decline. Molecularly, hsa-miR-6875-5p warrants validation in prospective cohorts to establish diagnostic cutoffs (e.g., FC≥2.0 for AD-HL comorbidity). Future research should use longitudinal designs and animal models to dissect the miRNA-mediated mechanisms linking HL and AD, and conduct RCTs to confirm CI’s long-term cognitive benefits in AD patients.

In summary, this study advances understanding of the AD-HL-miRNA axis, providing evidence-based guidance for clinical intervention and a novel biomarker for translational research.