Can auditory processing dysfunction indicate early cognitive decline?

Abstract

INTRODUCTION

Central auditory processing (CAP) is crucial for speech perception and is also fundamental for cognitive function. This study investigated whether gap detection threshold (GDT) could serve as an early marker for identifying individuals with cognitive impairment (CI) at high risk of dementia.

METHODS

Sixty‐four older adults underwent peripheral auditory, cognitive, and CAP assessments. Machine learning and resting state electroencephalography (EEG)/event‐related potential (ERP) analyses explored predictors and neural correlates of CI.

RESULTS

GDT was significantly higher in those with CI (mean ± standard deviation: 8.25 ± 6.14 versus 5.98 ± 3.44 ms, respectively, p = 0.034), and negatively correlated with cognitive test scores (e.g., Addenbrooke's Cognitive Examination III: r = –0.40, p = 0.001). GDT emerged as a key predictor. EEG showed altered auditory connectivity and ERP revealed reduced N1/N2 amplitudes in high‐GDT individuals (false discovery rate corrected p < 0.05).

DISCUSSION

GDT may reflect early neurophysiological changes in individuals with CI and has potential as a non‐invasive biomarker.

Highlights

• Central auditory processing (CAP) test scores were found to be significantly correlated with cognitive tests.

• By machine learning, the best variable gap detection threshold (GDT) for predicting cognitive impairment was screened out.

• GDT subgroup analysis was performed within the normal control (NC) group. Compared to the low GDT subgroup, the high GDT subgroup had lower amplitudes of the cognitive components of the event‐related potential and many differences in functional connectivity, indicating that GDT has predictive value for changes in cognitive function.

1. BACKGROUND

Age‐related hearing loss (ARHL), or presbycusis, is the foremost chronic sensory dysfunction among the elderly, affecting two thirds of individuals > 60 years of age. ¹ Hearing loss (HL) is independently linked to depression, cognitive impairment (CI), and dementia. ² , ³ According to the Lancet Commission, HL beginning in midlife contributes to a 9% increase in dementia incidence. However, the association of HL with cognition is not solely attributable to peripheral auditory decline. Central auditory processing disorder (CAPD), which involves the brain's interpretation of auditory signals, may play a more significant role.

CAPD can occur ahead of peripheral HL in the life span, ⁴ , ⁵ and may be more closely associated with CI in aging populations. ⁶ Many studies indicate that CAPD assessments may facilitate the early detection of CI before its full establishment. ⁷ , ⁸ , ⁹ For example, poor performance in dichotic tests is associated with increased risk of dementia in prospective studies. ¹⁰ However, existing evidence on the relationship between central auditory processing (CAP) and cognition remains inconsistent. A systematic review covering 25 studies concluded that most reported a positive association between speech‐in‐noise perception and cognition, though only a few showed significant differences between subjects with and without CI. ¹¹ Discrepancies across studies may be due to variability in both cognitive measures and CAP tests. Studies focusing on working memory typically report positive associations, ¹² while those on crystallized intelligence often do not. ¹³ Behavioral CAP assessments are diverse and lack standardization, complicating their use in predicting CI.

Because both CAP and cognitive function are commonly assessed with behavioral measurements prone to subjective variability, ¹⁴ objective measures could provide valuable complementary information. Electroencephalography (EEG), including resting‐state quantitative EEG (rs‐qEEG) and event‐related potential (ERP), has been widely used to study brain activity in relation to CI and dementia. ¹⁵ , ¹⁶ Abnormal EEG patterns have been reported in children with CAPD, ¹⁷ and ERP changes have been linked to cognitive deficits. ¹⁸ , ¹⁹ , ²⁰ However, studies simultaneously examining CAP and cognition using EEG/ERP, particularly in older adults, remain scarce.

The present study aimed to refine the connection between CAPD and CI to explore whether CAPD test(s) can serve as an early indicator for identifying individuals at high risk of future cognitive decline in the elderly. Comprehensive audiological evaluations were conducted, including peripheral and central auditory functions. A machine learning algorithm was applied to identify optimal CAPD predictors of CI from various auditory assessments. Both rs‐qEEG and ERP analyses were integrated with behavioral CAP and cognitive assessments to verify whether specific audiometric variables could serve as precursory indicators of dementia‐risk trajectories.

2. METHODS

2.1. Participants and study design

A total of 64 participants were recruited in the otolaryngology clinic of Shanghai Sixth People's Hospital Affiliated Jiao Tong University School of Medicine from August 2019 to December 2020. The inclusion criteria included (1) age of ≥ 55 years, (2) right‐handedness, (3) ability to comply with the instructions, and (4) understanding of the study purpose to provide informed consent. The exclusion criteria included (1) otologic conditions or (2) a history of psychiatric or neurological disorders, cardiovascular disease, diabetes, or alcohol dependency.

All participants were interviewed for medical history and demographic information and examined with audiometric tests, CAP tests, and neuropsychological evaluations. All data collection including the review of medical history, neurological examination, and neuropsychological assessments, and diagnosis of CI and Alzheimer's disease (AD) were carried out by two experienced geriatricians independently. Twenty‐nine were diagnosed with either subjective cognition decline (SCD, according to the SCD‐I framework [2014], ²¹ n = 12), mild cognitive impairment (MCI, according to the Jak/Bondi neuropsychological criteria [2014], n = 12), or AD (according to the National Institute on Aging–Alzheimer's Association clinical criteria [2018], n = 5) and were classified as the CI group. The other 35 elderly were categorized as the normal control (NC) group.

This study was approved by the ethics committee of the Sixth People's Hospital Affiliated to the Shanghai Jiao Tong University (Clinical Trial Number: ChiCTR‐RPC‐17012580).

2.2. Evaluations

All the evaluations requiring acoustic stimulation and/or recording of electrical responses were carried out in a sound‐proof room meeting professional standards.

2.2.1. Peripheral hearing assessment

Pure‐tone audiometry (0.25–8 kHz) and auditory brainstem response (ABR; 80 dB nHL, 13.1 Hz) assessed peripheral hearing. ABR waves I/V amplitudes and latencies were analyzed after 1024‐trial averaging (200–2000 Hz filtering).

2.2.2. CAP tests

Four CAP tests were conducted to cover three major domains in CAP: hearing in noise by speech‐in‐noise test (SINT) and competing sentence test (CST), binaural integration by dichotic listening (DL) test, temporal processing by gap‐detection threshold (GDT) tests.

RESEARCH IN CONTEXT

1. Systematic review: We conducted a literature review using databases such as PubMed to identify studies examining the relationship between central auditory processing (CAP) and cognitive decline. Prior research has primarily relied on behavioral tests, with inconsistent findings regarding which CAP measures best predict cognitive impairment (CI). Few studies incorporated objective methods such as electroencephalography (EEG) or machine learning to explore neurophysiological correlates or predictive potential.

2. Interpretation: Our findings demonstrate that gap detection threshold (GDT) is the only CAP measure that consistently distinguishes CI from normal aging and correlates with cognitive performance. Machine learning confirmed GDT as a key predictive feature. EEG and event‐related potential analyses revealed that GDT is associated with altered functional brain connectivity and reduced auditory attention responses, supporting its role as a sensitive marker of early cognitive decline.

3. Future directions: Future research should focus on longitudinal validation of GDT as a biomarker for cognitive decline, its neural mechanisms across diverse populations, and integration into routine audiological and cognitive screening protocols.

The Mandarin version of the hearing‐in‐noise test (MHINT) materials developed with MATLAB (R2012a) software by the House Ear Institute was used in both SINT and CST; each comprises 240 Chinese sentences lists (20 sentences in each list). Each sentence consists of 10 words, recorded from a male speaker. The sentences were balanced for naturalness, length, and intelligibility. In both tests, the sentences were presented bilaterally. In SINT, the sentence was presented under the masker by white noise with –5 dB signal‐to‐noise ratio (SNR). In CST, on the other hand, the competing sentences were presented by competing speech at an SNR of –3 dB. In each test, each subject was tested with two sentence lists, containing 20 sentences with 200 words in total. The performance was scored as the percentage of correctly repeated words out of the total.

The DL test used a digital version comprising 20 pairs of double‐digit numbers presented in random order. One number in each pair was presented to one ear, and the two numbers in each pair were aligned at onset and offset. Subjects were instructed to repeat both numbers. A response was judged correct if the two numbers were repeated correctly regardless of the order. The performance was scored as the percentage of correct responses out of the total 20 trials.

The GDT test is designed to determine the subject's ability to detect a silent gap between two noise bursts. In this study, it was measured using a three‐interval forced‐choice procedure implemented in MATLAB. Each interval consists of a 1000 ms noise burst. A silent gap, with its duration ranging from 1 to 20 ms, was inserted in the middle of the noise burst, which was randomly selected from one of the three intervals. Subjects were instructed to identify the interval in which the noise burst contained a gap (perceived as a discontinuity). They could do this by pressing a button corresponding to the targeted interval or clicking on the targeted interval displayed on the monitor screen. If the subject's response was correct, the gap size was decreased by 2 ms; if incorrect, it was increased by 1 ms. This procedure was repeated until 12 reversals were achieved. The GDT was calculated as the average gap duration for at least the last eight reversals.

2.2.3. Assessment of cognitive function

Cognitive status was assessed using the Montreal Cognitive Assessment Basic (MoCA‐B) and Addenbrooke's Cognitive Examination III (ACE‐III). Additional tests included the animal fluency test (AFT), Symbol Digit Modalities Test (SDMT), and the Boston Naming Test (BNT).

2.2.4. EEG data collection

EEG was recorded with a 256 channel system at 1000 Hz. A 5 minute eyes‐closed rsEEG was collected. ERP was obtained via an auditory oddball task with 85% 1 kHz standard tones and 15% 2 kHz target tones across 1000 trials (ISI: 850–1450 ms). Data were pre‐processed in EEGLAB with bandpass filtering and independent components analysis for artifact removal.

2.3. Predicting candidate selection by machine learning

Random forest recursive feature elimination (RF‐RFE) was used for selecting CAPD candidates as cognitive decline indicators, with 20 iterations of nested cross‐validation. Data were split into training and test sets to prevent leakage. A 5‐fold cross‐validation was performed on the training set, followed by feature importance evaluation using the test set. Extreme Gradient Boosting (XGBoost) ²² and random forest (RF) ²³ were applied to validate the important features selected. For interpretation of black‐box classifiers, the shapviz package was used to compute Shapley values. ²⁴

2.4. EEG data analysis

Spectrum analysis was performed using MATLAB2022a (Welch method), yielding power spectral density (µV²/Hz) across 150 rs‐EEG epochs. Source localization used EEGLAB, FieldTrip, and region of interest (ROI)–based analysis (see Table S1 in supporting information). Connectivity between ROIs and the dorsal attention network was calculated via phase‐locking value and stored in adjacency matrices. Power was averaged across theta, alpha1/2, beta1/2/3, and gamma bands. ERP analysis was also conducted using MATLAB2022a scripts.

2.5. Statistical analysis

Normality was assessed using the Shapiro–Wilk test. Continuous variables were expressed as mean ± standard deviation. Demographics were compared using t tests or Wilcoxon tests. Spearman partial correlations examined links between auditory and cognitive measures. EEG data were corrected using Bonferroni and false discovery rate (FDR) methods; post hoc tests used 1000 times bootstrap. Analyses were performed in R v4.2.2 and MATLAB2022a, with p < 0.05 (two sided) considered significant.

3. RESULTS

3.1. Participants characteristics

A total of 64 participants with mean age of 67.09 ± 6.29 years were enrolled in this study; 23 (36%) were female. Demographic characteristics and test outcomes stratified by cognition function are shown in Table 1. No significant differences were found between NC and CI groups in age, education, sex, or peripheral hearing (all p > 0.05). The CI group exhibited a significantly larger GDT score (p = 0.034), while other CAP tests showed no differences. All the cognitive tests between the two groups had a significant difference (all p < 0.005).

TABLE 1.

Characteristics in each group.

	Total	Normal control	Cognitive impairment
Characteristic	(n = 64)	(n = 35)	(n = 29)	p
Age, mean ± SD, years	67.09 ± 6.29	65.77 ± 5.97	68.69 ± 6.39	0.064
Education, mean ± SD, years	11.13 ± 3.72	11.24 ± 3.79	11 ± 3.7	0.989
Sex, number (M / F)	41/23	24/11	17/12	0.573
PTA, mean ± SD, dB HL	23.22 ± 12.49	21.88 ± 12.51	24.83 ± 12.48	0.338
ABR, mean ± SD
LV latency	5.65 ± 0.28	5.66 ± 0.23	5.63 ± 0.33	0.603
LV amplitude	0.4 ± 0.16	0.37 ± 0.12	0.44 ± 0.19	0.225
GDT	7.01 ± 4.94	5.98 ± 3.44	8.25 ± 6.14	0.034*
DL left	0.57 ± 0.2	0.59 ± 0.18	0.55 ± 0.22	0.522
DL right	0.63 ± 0.19	0.67 ± 0.18	0.59 ± 0.2	0.149
SIQT	0.96 ± 0.09	0.97 ± 0.09	0.95 ± 0.1	0.378
SINT	0.39 ± 0.21	0.4 ± 0.21	0.37 ± 0.22	0.512
CST	0.23 ± 0.17	0.25 ± 0.18	0.22 ± 0.15	0.566
MoCA	23.94 ± 3.96	25.4 ± 2.91	22.17 ± 4.37	<0.001***
ACE‐III	76.83 ± 11.26	80.83 ± 9.5	72 ± 11.47	0.003**
AFT	15.63 ± 4.62	17.29 ± 4.18	13.62 ± 4.37	0.001**
BNT	23.2 ± 4.11	24.54 ± 3.51	21.59 ± 4.25	0.004**
SDMT	34.47 ± 13.8	39.06 ± 14.9	28.93 ± 10.02	0.002**

Abbreviations: ABR, auditory brainstem response; ACE‐III, Addenbrooke's Cognitive Examination III; AFT, animal fluency test; BNT, Boston Naming Test; CST, competing sentence test; dB, decibel; DL, dichotic listening; GDT, gap detection threshold; HL, hearing level; LV, Left V wave; MoCA, Montreal Cognitive Assessment; PTA, pure tone average; SD, standard deviation; SDMT, Symbol Digit Modalities Test; SINT, speech‐in‐noise test; SIQT, speech‐in‐quiet text.

^*** p < 0.001,

^** p < 0.01.

^* p < 0.05.

3.2. Correlation analysis and machine learning

As only GDT differed significantly between CI and NC groups, correlation analysis focused on GDT. GDT was negatively correlated with all five cognitive test scores: MoCA (r = –0.34, p = 0.006), ACE‐III (r = –0.40, p = 0.001), AFT (r = –0.29, p = 0.02), BNT (r = –0.31, p = 0.01), and SDMT (r = –0.19, p = 0.12; Figure 1A).

FIGURE 1.

Correlation analysis and summary of machine learning performance. (A) Heatmap of correlations between auditory/vestibular tests and cognitive evaluations (**p < 0.01, *p < 0.05). (B) The AUC values of the training and validation sets for RF and XGBoost models (with the bars for mean values and error bars for SD). (C) Prediction performance of evaluated variables (including GDT) represented by SHAP values. The size of the bubble represents the relative importance of each variable within the predictive model to individual cognitive evaluations; the color intensity of the bubbles denotes the SHAP values. Variables without bubbles were not included in the model. ABR, auditory brainstem response; ACE‐III, Addenbrooke's Cognitive Examination III; AFT, Animals Fluency Test; AUC, area under the curve; BNT, Boston Naming Test; CAPT, Central auditory processing test; CST, competing sentence test; DL, dichotic listening; GDT, gap detection threshold; LV,Left V wave; MoCA, Montreal Cognitive Assessment; PTA, pure tone average; RF, random forest; SDMT, Symbol Digit Modalities Test; SHAP, SHapley Additive exPlanations; SINT, speech‐in‐noise test; SIQT, speech‐in‐quiet text; XGBoost, Extreme Gradient Boosting.

Nested cross‐validation was performed using RF‐RFE, and variables with variable importance > 1 and occurrence times > 200 were selected for the next step. GDT, DL Left, SINT, and CST were all initially considered the functional significant indicators of CI. RF outperformed XGBoost with higher areas under the curve across cognitive tests (Figure 1B).

SHapley Additive exPlanations (SHAP) values showed GDT had strong predictive power for all cognitive scores (Figure 1C). Further analysis focused on NC group to assess whether GDT could predict early cognitive decline within the normal range.

3.3. Characteristics of subjects in NC group

All the subjects in the NC group showed normal cognitive scores in the tests evaluated in this study; they showed a significant amount of variation in their performance in gap detection test. We therefore distinguished all the subjects in the NC group by their GDT: 15 with highest GDT (H‐GDT) and 15 with lowest (L‐GDT). The GDT values were 3.78 ± 0.45 for the L‐GDT subgroup and 6.98 ± 0.78 for the H‐GDT subgroup with a significant difference (p = 0.004), although there was no significant difference in all the cognitive test scores between the two subgroups of NC subjects (all p > 0.05; Table 2).

TABLE 2.

Characteristics in each GDT subgroup.

Characteristic	Lowest GDT	Highest GDT	p
	(n = 15)	(n = 15)
Age, mean ± SD, years	60.27 ± 6.76	65.40 ± 6.02	0.037 *
Education, mean ± SD, years	12.21 ± 3.68	11.00 ± 3.55	0.401
Sex, number (M/F)	4/11	3/12	0.679
PTA, mean ± SD, dB HL	13.25 ± 3.57	13.17 ± 5.25	0.96
ABR, mean ± SD
LV latency	5.55 ± 0.21	5.58 ± 0.19	0.696
LV amplitude	0.45 ± 0.15	0.45 ± 0.20	0.934
GDT	3.78 ± 0.45	6.98 ± 0.78	0.004*
DL left	0.69 ± 0.15	0.65 ± 0.15	0.395
DL right	0.75 ± 0.17	0.66 ± 0.18	0.164
SIQT	0.96 ± 0.13	0.97 ± 0.05	0.791
SINT	0.52 ± 0.13	0.50 ± 0.19	0.395
CST	0.36 ± 0.21	0.30 ± 0.13	0.377
MoCA	25.87 ± 2.50	24.50 ± 3.04	0.201
ACE‐III	81.80 ± 6.57	77.87 ± 10.07	0.216
AFT	18.64 ± 5.42	16.13 ± 4.85	0.228
BNT	24.27 ± 2.09	22.50 ± 4.72	0.204
SDMT	42.93 ± 12.78	32.53 ± 16.54	0.064

Abbreviations: ABR, auditory brainstem response; ACE‐III, Addenbrooke's Cognitive Examination III; AFT, Animals Fluency Test; BNT, Boston Naming Test; CST, competing sentence test; dB, decibel; DL, dichotic listening; GDT, gap detection threshold; HL, hearing level; LV, Left V wave; MoCA, Montreal Cognitive Assessment; PTA, pure tone average; SD, standard deviation; SDMT, Symbol Digit Modalities Test; SINT, speech‐in‐noise test; SIQT, speech‐in‐quiet text.

^* p < 0.05.

A significant correlation was found between GDT values and the scores of ACE‐III (r = –0.4075, p = 0.0349), BNT (r = –0.4027, p = 0.0373), and SDMT (r = –0.4669, p = 0.0141), but not for MoCA and AFT (Figure 2).

FIGURE 2.

Correlations between GDT and cognitive tests. ACE‐III, Addenbrooke's Cognitive Examination III; AFT, Animals Fluency Test; BNT, Boston Naming Test; GDT, gap detection threshold; MoCA, Montreal Cognitive Assessment; SDMT, Symbol Digit Modalities Test.

3.4. Resting‐state EEG and functional connectivity

The rs‐EEG showed no significant difference at any frequency band (all p > 0.05; Figure 3A) between CI and NC groups. However, interesting differences in functional connectivity were found between the GDT groups. First, the lagged phase synchronization analysis indicated a stronger functional connectivity between the angular gyrus (ANG) and the triangular part of inferior frontal gyrus (IFGtriang) in the delta (FDR p = 0.030), theta (FDR p = 0.017), beta1 (FDR p = 0.030), and beta2 (FDR p = 0.048) bands in H‐GDT subgroup (Figure 3B,C,E, and F). Second, within the theta band, the H‐GDT subgroup showed a significant augmentation of functional connectivity between the right IFGtriang and the adjacent cortical regions of the same side, namely, the middle temporal gyrus (MTG; FDR p = 0.050) and the superior frontal gyrus dorsolateral (SFGdor) (FDR p = 0.046), and also encompassing the superior frontal gyrus medial orbital (ORBsupmed; FDR p = 0.046; Figure 3C). Third, increased functional connectivity in the H‐GDT subgroup was also seen between the right ANG and the right middle frontal gyrus (MFG) in both delta (FDR p = 0.026) and theta (FDR p = 0.019) bands (Figure 3B, C), and between the MFG and the SFGdor in the alpha2 band (FDR p = 0.014; Figure 3D). Fourth, however, a notable reduction in functional connectivity was found in the H‐GDT subgroup in the gamma band between superior frontal gyrus medial (SFGmed.L) and the inferior frontal gyrus orbital part (ORBinf; FDR p = 0.006), and between superior frontal gyrus orbital part (ORBsup) and middle frontal gyrus orbital part (ORBmid; FDR p = 0.020), as well as ORBinf (FDR p = 0.002; Figure 3G).

FIGURE 3.

Resting‐state electroencephalography analysis. (A) The power spectrum densities (PSDs) between two subgroups. The functional connectivity comparisons are shown in Delta (B), Theta (C), Alpha2 (D), Beta1 (E), Beta2 (F), Gamma (G) frequency band, with red lines indicating a stronger connectivity in the H‐GDT subgroup and blue lines a weaker one as compared with the L‐GDT subgroup. SFGmed: Superior frontal gyrus, medial; MFG: Middle frontal gyrus; IFGtriang: Inferior frontal gyrus, triangular part; ANG: Angular gyrus; ORBinf: Inferior frontal gyrus, orbital part; ORBsup: Superior frontal gyrus, orbital part; MTG: Middle temporal gyrus; SFGdor: Superior frontal gyrus, dorsolateral; ORBsupmed: Superior frontal gyrus, medial orbital; ORBmid: middle frontal gyrus, orbital part.

3.5. ERP

Subjects in the H‐GDT subgroup showed a significantly lower amplitude in N1 (latency 140 ms), and N3 (latency 550 ms) recorded from the Fz electrode, and N1, N2 (latency 250–300 ms) from the Cz electrode than those in the L‐GDT subgroup in the response to deviant stimuli in the odd‐ball paradigm (Figure 4A, B; all FDR p < 0.05). In addition, their response to the standard stimuli also showed a lower N1 amplitude from Fz (Figure 4C, FDR p < 0.05), but not from Cz (Figure 4D) compared to those in L‐GDT subgroup. Consequently, the H‐GDT subgroup (Figure 4E, F) showed significantly smaller target‐standard differences at N2 and N3 than the L‐GDT subgroup as recorded from Fz and at N2 from Cz (all FDR p < 0.05). However, no difference was seen between the two subgroups in any of the recordings from the Pz electrode (all p > 0.05) (data not shown).

FIGURE 4.

Event‐related potential comparisons between GDT groups. A, C and E show the target, standard responses and target‐standard difference respectively recorded from Fz electrode; B, D and F show the corresponding responses from Cz. Scalp maps depict the voltage distribution in the specified time windows, revealing clear midline parietal maxima in all groups. All ERP components with significant differences between the two subgroups were marked by the gray bars. GDT, gap detection threshold; H‐GDT, highest gap detection threshold; L‐GDT, lowest gap detection threshold.

3.6. Correlation analysis between ERP and cognitive tests

Across all the subjects in the NC group, correlation analyses were carried out between ERP amplitude and all the cognitive test scores. The ERP responses to deviant stimuli were found to be correlated only with the SDMT score (r = 0.3970, p = 0.0494 for N1 from Fz and r = 0.4598, p = 0.0158 for N2 from Cz). The correlation analyses were also carried out between the target‐standard difference and the cognitive scores but no significant correlation was found (Figure S1 in supporting information).

4. DISCUSSION

This study explored the predictive value of auditory and non‐auditory variables for cognitive decline (CD), with a focus on CAP abilities. We found that: (1) while the scores of every CAP test were significantly correlated with some of the cognitive test(s), only GDT distinguished the CI and NC groups in our sample; (2) GDT was also picked up as the top predictor by machine learning models (RF and XGBoost), with the highest SHAP values across all cognitive scores; (3) even among NC individuals with normal behavioral cognitive test scores, GDT variation was substantial, correlated with cognitive scores, and reflected in EEG/ERP patterns—suggesting that GDT is sensitive to early cognitive changes not yet behaviorally evident.

We focused on identifying predictors of CI rather than AD, primarily because our interest lies in hearing impairment (HI)–related CI. Compared to AD, which has a relatively well‐defined neuropathological basis, CI is more prevalent, and its mechanisms remain less clearly understood—especially among individuals with HI, for whom CI often represents a prodromal stage of AD. More importantly, only a very small portion of CI subjects eventually develop AD. Therefore, a strong predictor for CI may not be useful to predict AD.

The present study is distinguished from previous reports investigating the relationship between CAPD and CI by its focus on the changes of CAP and cognitive functions in subjects with normal behavioral scores on cognitive tests. Differences in CAP have been found between subjects with and without CI in many previous studies. ⁶ , ⁷ , ²⁵ , ²⁶ The present study provided a strong support to the predictive value of GDT for CI or cognitive decline before occurrence of CI by the evidence summarized above.

Unlike prior studies that typically emphasized either CAP–behavioral cognitive correlations ⁶ , ²⁷ , ²⁸ , ²⁹ or subjective–objective cognitive assessments, ³⁰ , ³¹ , ³² we integrated CAP tests, cognitive scores, ERP, and EEG connectivity. This comprehensive approach provided clearer links among CAPD, cognition, and brain physiology. Notably, we focused on how CAP variation—especially in GDT—correlates with cognitive differences within a normal population.

It is worth noticing that GDT is the only CAP variable that shows a consistent correlation with the score of cognitive tests used in the present study. This is interesting because across different aspects of CAP, temporal process (especially as evaluated by GDT) was considered not strongly connected with CI by several studies. ³³ , ³⁴ However, there is growing evidence supporting its role: children with CD show poorer gap detection, ³⁵ and GDT has been linked to cortical atrophy and progression from MCI to AD. ³⁶ In our study, the cognitive tests adapted in the present study (MoCA, ACE‐III, AFT, BNT, and SDMT) nicely covered broad aspects of cognitive function, ³⁷ , ³⁸ , ³⁹ , ⁴⁰ strengthening the reliability of our findings.

Machine learning models (RF‐RFE with nested cross‐validation) further highlighted GDT's predictive power. GDT consistently ranked as the most informative variable, outperforming others in predicting cognitive scores. These algorithms help minimize human selection bias and enhance generalizability, ⁴¹ offering a novel methodological strength rarely applied in this research field.

There is currently no universal GDT cutoff due to methodological and demographic variations. In our sample, the L‐GDT subgroup showed GDT values within reported normal ranges, ⁴² , ⁴³ whereas the H‐GDT subgroup exhibited significantly elevated values. While both groups were within the NC population, their neural and cognitive profiles differed. The causality between CAPD and CD/CI remains debated. Some propose that cognitive dysfunction leads to top‐down CAP impairment, ⁴⁴ , ⁴⁵ while others argue CAPD contributes to cognitive deterioration. ²⁹ , ⁴⁴ Our results, showing early CAP dysfunction preceding measurable cognitive decline, lend support to the latter view.

EEG results revealed altered functional connectivity in the H‐GDT group. Specifically, we observed increased low‐to‐mid frequency connectivity between temporo‐parietal regions and the ANG, areas associated with semantic processing, attention, and executive control. ⁴⁶ , ⁴⁷ This may reflect compensatory neural recruitment, aligning with theories that individuals with emerging cognitive decline engage broader brain networks to maintain function. ⁴⁸ , ⁴⁹ Conversely, reduced gamma connectivity between inferior and superior frontal gyri in H‐GDT participants may indicate early auditory processing deficits. ⁵⁰

In ERP analyses, no differences were observed in early sensory components (< 250 ms) or responses to standard stimuli between H‐GDT and L‐GDT groups. However, cognitive components—those typically associated with attention and novelty detection—differed significantly. This suggests that changes in cognitive Auditory Evoked Potentials (AEP) components may emerge earlier than those in sensory processing, and that GDT is tightly linked to these early shifts.

Unexpectedly, we did not observe a typical P300 peak in ERP responses to deviant tones. Instead, negative deflections were observed, potentially due to the simplicity of the auditory stimuli used. As the P300 is usually evoked by more cognitively demanding stimuli, this may explain its absence in our data.

Our study has several limitations. First, the cross‐sectional design prevents causal inference about the relationship between CAPD and CD. We are addressing this limitation through ongoing longitudinal studies. Second, due to sample size constraints, we categorized participants broadly into NC and CI groups, rather than stratifying by specific stages of cognitive decline. This limits the granularity of our analysis regarding CAP severity and CI severity.

In conclusion, GDT emerged as a promising, non‐invasive biomarker for early cognitive changes, even in individuals without clinical impairment. Its association with both cognitive scores and brain network alterations highlights its potential for identifying individuals at risk for CD. The combination of behavioral assessments, EEG measures, and machine learning offers a comprehensive framework for future research and clinical screening efforts.

AUTHOR CONTRIBUTIONS

Wang Hui conceived and designed the study. Wang Jian and Yin Shankai revised the manuscript and provided a review. Ma Xinrong performed the analyses and wrote the manuscript. Li Jiayu and Wang Ying helped analyze the data. Li Shiyuan, Guo Junjie, Yu Xiao, Shen Wenxin, Dong Hongyu, Huang Shujian, and Li Linpeng helped collect the data. All authors read and approved the final manuscript.

CONFLICT OF INTEREST STATEMENT

All authors declare that there are no conflicts of interest related to this work. Author disclosures are available in the supporting information.

Supporting information

Supporting Information

Supporting Information

Ma X, Li J, Wang Y, et al. Can auditory processing dysfunction indicate early cognitive decline? Alzheimer's Dement. 2025;17:e70189. 10.1002/dad2.70189

DATA AVAILABILITY STATEMENT

The data of this study are not publicly available due to the issue of intellectual property, but are available from the corresponding author on reasonable request. All human subjects provided informed consent.