Neurobiological markers across joint profiles of subjective cognitive decline and objective cognitive function in older adults

Abstract

INTRODUCTION

Subjective cognitive concerns frequently diverge from objective cognitive performance in cognitively unimpaired (CU) older adults, yet the neurobiological basis of this mismatch remains unclear.

METHODS

In 648 participants from the Investigating Gains in Neurocognition in an Intervention Trial of Exercise (IGNITE), we defined four profiles by integrating subjective and objective cognitive status. We examined associations with plasma neurofilament light chain (NfL), phosphorylated tau 217 (p‐tau217), glial fibrillary acidic protein (GFAP), a magnetic resonance imaging–based volumetric Alzheimer's disease (AD) signature reflecting atrophy, and brain‐predicted age difference (brain‐PAD).

RESULTS

Joint profiles were differentially associated with NfL (P = 0.0427) and brain‐PAD (P = 0.0296). Follow‐up contrasts further indicated higher NfL and lower volumetric AD signature in the concordant lower functioning profile, and higher brain‐PAD in discordant profiles. p‐tau217 and GFAP did not differ across profiles.

DISCUSSION

Joint subjective–objective cognitive profiles may capture biologically meaningful heterogeneity relevant to neurodegeneration and brain aging in older adults.

TRIAL REGISTRATION

ClinicalTrials.gov: NCT02875301

Highlights

• A data‐driven approach identified four joint subjective–objective cognitive profiles.

• Joint cognitive profiles mapped onto distinct neurobiological markers.

• Concordant lower functioning was linked to higher neurofilament light chain.

• Discordant profiles showed higher brain‐predicted age difference.

• Phosphorylated tau 217 and glial fibrillary acidic protein did not differ across subjective–objective profiles.

1. BACKGROUND

Subjective cognitive complaints and performance on objective neuropsychological tests capture complementary perspectives on cognitive aging but often diverge in cognitively unimpaired (CU) older adults. This divergence can reflect measurement context, including differences between performance under structured testing and real‐world cognitive demands, practice effects on repeated assessment, and ceiling or floor constraints of standardized neuropsychological tests that limit sensitivity to subtle changes that might be more detectable in everyday contexts. ¹ , ² , ³ Socio‐emotional factors such as mood and anxiety, personality and health beliefs, and individual differences in education attainment or cognitive reserve can differentially influence subjective ratings and objective test performance. ⁴ , ⁵ , ⁶ Early biological changes that arise before clinically detectable cognitive impairment, including amyloid and tau accumulation and neurodegeneration or atrophy, may further contribute to disagreement between self‐report and neuropsychological test performance. ⁷ , ⁸

Prior studies have frequently examined subjective complaints and objective cognition separately or treated one as a predictor or covariate of the other, leaving the joint phenotype largely uncharacterized and contributing to mixed conclusions about clinical significance. ⁹ , ¹⁰ Studies that evaluate both self‐reported cognitive decline and objective test performance demonstrate heterogeneous patterns, including subjective cognitive decline (SCD) despite preserved objective performance, reduced insight/awareness characterized by impaired performance with limited complaint, and concordant normal or concordant impairment patterns. ¹¹ , ¹² , ¹³ , ¹⁴ These profiles differ in everyday functioning and in risk for subsequent cognitive decline and clinical outcomes, although the direction of risk appears to vary across cohorts. ¹⁵ , ¹⁶ However, the biological features underlying those profiles in CU older adults remain incompletely characterized, as many studies focus on a single biomarker, clinic‐based populations, or restrict to the SCD profile alone without parallel evaluation of alternative subjective–objective configurations.

Integrating neurobiological markers with subjective–objective profiles may improve biological characterization of cognitive status in CU older adults. Biomarkers that index neurodegeneration and global brain aging, including plasma neurofilament light chain (NfL), structural magnetic resonance imaging (MRI) composites of atrophy in Alzheimer's disease (AD) vulnerable regions, ¹⁷ and brain‐predicted age difference (brain‐PAD), provide information on neuronal injury and systems‐level aging. In contrast, plasma phosphorylated tau 217 (p‐tau217) indexes AD‐related amyloid and tau pathology, whereas glial fibrillary acidic protein (GFAP) indexes astroglial activation associated with AD‐related inflammatory processes. Evaluating these biomarkers across profile types may clarify whether discordance between subjective reports and objective performance, or differences across profiles, is accompanied by measurable neurobiological differences. Accordingly, prespecified and transparent definitions of subjective and objective status are essential for valid inference of potential biomarker differences.

Prior studies commonly used single instruments or simple threshold rules to classify subjective or objective status. ¹⁸ , ¹⁹ Subjective cognitive status is often defined by one self‐report questionnaire or even a single item, which may not robustly capture the breadth and heterogeneity of everyday cognitive concerns. Objective cognitive status is frequently determined by a single test score from a brief screening measure or a fixed z score cutoff in a single domain, which overlooks multidimensional performance and increases the risk of misclassifying an individual's cognitive status. Data‐driven approaches that aggregate multiple self‐report indicators into a latent subjective construct and consider objective performance across several cognitive domains can better capture individual differences and yield more accurate cognitive profiles. Whether biomarker patterns differ across profiles defined with modern, data‐driven approaches remains unclear in CU older adults.

In this study, we examined whether subjective–objective profiles derived from data‐driven approaches were associated with differences in biomarker patterns in a community sample of CU older adults. Given that NfL and structural MRI atrophy primarily reflect downstream neurodegeneration, whereas p‐tau217 and GFAP capture more proximal AD‐related tau and glial processes, we hypothesized that markers of neurodegeneration and global brain aging, including NfL, the structural MRI AD signature, and brain‐PAD, would differ across subjective–objective profiles. In contrast, we predicted that markers of AD‐related pathology and astroglial activation, including p‐tau217 and GFAP, would show limited separation across profiles.

RESEARCH IN CONTEXT

1. Systematic review: We reviewed the literature using PubMed and reference lists of relevant articles to identify studies on subjective cognitive decline, objective cognition, and neurobiological markers in cognitively unimpaired older adults. Prior work has largely examined subjective and objective cognition separately; fewer studies have evaluated their joint profiles across plasma and magnetic resonance imaging–derived biomarkers.

2. Interpretation: We found that joint profiles of subjective cognitive decline and objective cognitive function were differentially associated with neurobiological markers. The strongest associations were observed for neurofilament light chain and brain‐predicted age difference, whereas phosphorylated tau 217 and glial fibrillary acidic protein were not associated with profile status. These findings suggest that joint subjective–objective cognitive profiles are associated with differences in markers of neurodegeneration and brain aging in cognitively unimpaired older adults.

3. Future directions: Future longitudinal studies are needed to determine whether these joint cognitive profiles predict subsequent neurodegeneration, brain aging, cognitive decline, or dementia risk, and assess their stability over time.

2. METHODS

2.1. Participants

Participants (N = 648) were enrolled in the Investigating Gains in Neurocognition in an Intervention Trial of Exercise (IGNITE) study, a multicenter randomized clinical trial (ClinicalTrials.gov: NCT02875301) assessing the effect of aerobic exercise on cognition and brain health. IGNITE was conducted across three sites in the United States: Boston, Pittsburgh, and Kansas City. The present analyses were restricted to baseline data collected prior to randomization. Participants were community‐dwelling older adults, recruited via advertisements and community outreach, aged 65 to 80 years at baseline, free of neurological disease diagnoses (e.g., multiple sclerosis, Parkinson's disease, dementia, or stroke), and considered inactive (engaged in < 20 minutes of moderate‐intensity physical activity three times per week). Additional demographic and clinical characteristics of the analytic sample, including sex, race, and selected medical comorbidities, are summarized in Table 1. Consensus adjudication was reached after a comprehensive cognitive assessment to exclude individuals with probable mild cognitive impairment (MCI) or dementia, ²⁰ but given the known limitations of neuropsychological testing and the variable definitions of MCI, it is possible that some included participants were near the MCI range. Full exclusion criteria for the IGNITE study have been published previously. ²⁰ The study was approved by the institutional review board at each site, and all participants provided written informed consent before data collection.

TABLE 1.

Descriptive characteristics of participants at Investigating Gains in Neurocognition in an Intervention Trial of Exercise (IGNITE) baseline.

Characteristic	Overall, N = 648; Mean (SD) or n (%)
Age, years	69.88 (3.75)
Sex
Male	187 (29%)
Female	461 (71%)
Site
Pitt	219 (34%)
KU	214 (33%)
NEU	215 (33%)
Race
Black	123 (19%)
White	491 (75.8%)
Asian	10 (1.5%)
Multi‐race	16 (2.5%)
Native Hawaiian	2 (0.3%)
Another race	4 (0.6%)
Refused to answer	2 (0.3%)
Ethnicity
Hispanic or Latinx	20 (3%)
Not Hispanic or Latinx	628 (97%)
Education, years	16.32 (2.21)
BMI, kg/m2	29.74 (5.75)
Hypertension	377 (58%)
Type 2 diabetes	103 (16%)
Hyperlipidemia	301 (46%)
Cardiovascular disease	132 (20%)
APOE ε4
Non‐carrier	466 (73%)
Carrier	174 (27%)
Missing	8
Neurobiological markers
NfL, pg/mL	16.81 (6.82)
Missing	43
AD signature, mm3	64,907.46 (5439.43)
Missing	7
Brain‐PAD, years	−4.01 (6.66)
Missing	7
pTau217, pg/mL	0.43 (0.28)
Missing	23
GFAP, pg/mL	178.75 (96.77)
Missing	42

Abbreviations: AD, Alzheimer's disease; APOE, apolipoprotein E; BMI, body mass index; brain‐PAD, brain‐predicted age difference; GFAP, glial fibrillary acidic protein; KU, University of Kansas; NEU, Northeastern University; NfL, neurofilament light chain; Pitt, University of Pittsburgh; p‐tau217, phosphorylated tau 217; SD, standard deviation.

2.2. Measures

2.2.1. Neuropsychological assessment

Objective cognition was assessed using a comprehensive neuropsychological battery of verbal, paper–pencil, and computerized assessments. Testing was administered by certified psychometricians and completed across 2 days. Raw scores were normalized and combined using a confirmatory factor analysis to produce five latent factors ²¹ for cognitive domains, including executive function (EF)/attentional control (Flanker Task; Stroop Task [incongruent trial]; Dimensional Change Card Sort task; Trail Making Test, Part B), episodic memory (Brief Visuospatial Memory Test, Picture Sequencing Test, Hopkins Verbal Learning Test, Logical Memory Task, Verbal Paired Associates), processing speed (Letter Comparison Test; Digit Symbol Substitution Test; Trail Making Test, Part A), working memory (N‐Back Working Memory Task, Spatial Working Memory Task, List Sorting Working Memory Task), and visuospatial function (Matrix Reasoning, Spatial Relations, Clock Draw).

2.2.2. Objective status classes

Objective cognitive status (higher/lower performance) was derived from the five latent factors using a data‐driven clustering approach. We applied k‐means clustering with k = 2 to the five‐dimensional factor space, using multiple random starts to reduce sensitivity to initialization. The number of clusters was prespecified as k = 2 to align with the 2 × 2 subjective–objective framework and clinical interpretability. This choice was also supported by the highest average silhouette width across candidate k, ranging from 2 to 8 (Figure S1; details in Supplementary Methods in supporting information). Clusters were labeled by their multivariate centroids: the lower centroid cluster was labeled lower objective testing performance, and the other was labeled higher objective testing performance. Cluster quality was summarized by average silhouette width and by domain‐specific standardized mean difference (Hedges g) comparing higher and lower performance. The binary objective status indicator was carried forward into the factorial models.

2.2.3. Subjective cognition composite

Four subjective cognition questionnaires were administered: the Everyday Cognition Questionnaire (ECOG), ²² Patient‐Reported Outcomes Measurement Information System (PROMIS) Applied Cognition: Abilities, PROMIS Applied Cognition: General Concerns, ²³ and the Cognitive Function Index (CFI). ²⁴ Details of each questionnaire are available in Supplementary Methods.

A confirmatory factor analysis (CFA) was performed to derive the latent architecture of the questionnaires related to subjective complaints. Prior to conducting the CFA, all questionnaire outcomes were examined to identify missing data and out‐of‐range values. Two participants had missing data in the PROMIS questionnaires, and 16 participants had out‐of‐range values (scores ± 4 standard deviation [SD] from the mean) in at least one questionnaire outcome. We imputed the missing values in the questionnaire outcomes with the median value for each outcome and retained all other out‐of‐range values. There was no significant difference in goodness‐of‐fit indices when excluding these out‐of‐range values. Z standardization was used to ensure that outcomes with different response scales were on a standardized and comparable metric.

The CFA was conducted using the lavaan package (version 0.6‐19) in R (version 4.2.1) with maximum likelihood estimation. We tested the fit of several CFA models that differed in the following ways: (1) a single‐factor model incorporating all individual questionnaire outcomes using the ECOG total score, (2) a single‐factor model including individual ECOG subdomain scores as indicators, (3) a single‐factor model excluding the ECOG visuospatial/perceptual subdomain from the latent factor and allowing correlated residuals between two PROMIS outcomes, and (4) a hierarchical model in which five ECOG subdomain scores loaded on a first‐order ECOG factor that, in turn, loaded onto a higher‐order subjective cognitive complaints factor (Figure S2 in supporting information). Established goodness‐of‐fit indices used to assess model fit included χ², the χ²/df ratio, the CFI, the Tucker–Lewis Index (TLI), the root mean square error of approximation (RMSEA), the standardized root mean square residual (SRMR), and the Akaike information criterion (AIC; Table S1 in supporting information). After the best‐fit model was identified, the factor scores were extracted for further analysis.

2.2.4. Subjective status classes

The subjective complaints factor scores were reversed so that higher values indicated better subjective cognition. The oriented factor scores were then entered into a two‐component Gaussian mixture model, yielding posterior probability P(non‐elevated). Primary classification was based on the maximum posterior probability. The classification strategies were tailored to the measurement structure of each construct: objective cognition was defined within a five‐dimensional neuropsychological performance space, whereas subjective cognition was modeled as a unidimensional latent factor derived from self‐report measures.

We further defined “confident” classifications as P(non‐elevated) ≥ 0.80 or P(non‐elevated) ≤ 0.20, with intermediate values (0.20–0.80) considered a gray zone. Model selection and robustness analyses are reported in Supplementary Methods. Statistical models were repeated in the confident subset as a sensitivity analysis.

2.2.5. Neurobiological markers

Approximately 47 cc of blood was collected during fasting conditions in the morning between 8:00 am and 11:00 am. Isolated plasma was immediately stored at −80°C. Apolipoprotein E (APOE) genotype was determined using TaqMan assays, and participants with at least one APOE ε4 allele were classified as carriers. Plasma p‐tau217, NfL, and GFAP were measured using the SIMOA platform (Quanterix). Specifically, NfL and GFAP were analyzed using the N2 PB assay (No. 103520) on a SIMOA‐HD X at the University of Kansas Medical Center. P‐tau217 was measured using the ALZpath assay kit (No. 104371) at the Department of Psychiatry, University of Pittsburgh. Quality control samples were assessed at the beginning and the end of each run to evaluate reproducibility. The average within‐run coefficients of variation (CVs) were p‐tau217 = 11.0%, NfL = 6.3%, and GFAP = 9.7%. The mean between‐run CVs were p‐tau217 = 11.4%, NfL = 8%, and GFAP = 14.8%. Details on blood biomarker collection and processing have been described previously. ²⁵

MRI acquisition procedures have been described previously. ²⁰ Briefly, MRI scans were completed on a Siemens Prisma 3T scanner with a 64‐channel head coil at the University of Pittsburgh and Northeastern University and on a Siemens Skyra 3T scanner with a 32‐channel head coil at the University of Kansas Medical Center. High‐resolution T1‐weighted 3D MPRAGE (magnetization‐prepared rapid gradient echo imaging) images were collected on all participants (0.8 × 0.8 × 0.8 mm voxels, 224 slices, 0.8 mm slice thickness, repetition time =  2400.0 ms, echo time =  2.31 ms, inversion time = 1060 ms, flip angle  =  8 degrees). Because MRI data were acquired across multiple study sites, study site was included as a covariate in all statistical models to account for potential site or scanner‐related variability.

MPRAGE images were submitted to the Computational Anatomy Toolbox (CAT12) implemented in the Statistical Parametric Mapping 12 tool (SPM12) to derive a quantitative image quality rating (IQR) indicating the quality of each T1‐weighted image. ²⁶ This image quality metric reflects an aggregate composite of noise, bias, and resolution on a percentage scale, with higher values (100% maximum) indicating better image quality.

T1‐weighted images were processed using FreeSurfer 6.0 (https://surfer.nmr.mgh.harvard.edu) via the “recon‐all” pipeline, with the ‐3T flag (recon‐all ‐3T). AD signatures were computed using a previously published method. ¹⁷ The calculation process included an adjustment for estimated intracranial volume based on the Desikan–Killiany 2006 atlas. This adjustment was applied to five volumetric regions of interest (ROIs): hippocampus, precuneus, parahippocampus, entorhinal region, and inferior parietal cortex. The final AD signature was derived as a sum of the volume of these adjusted ROIs; therefore, higher values indicate preserved brain structure in these regions.

Brain age estimation was performed on T1‐weighted images using the brainageR analysis pipeline (github.com/james‐cole/brainageR). ²⁷ The pipeline first initiates voxel‐level preprocessing using the SPM12 toolbox, including segmentation of raw T1‐weighted images into gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF), followed by non‐linear spatial normalization to a template image using SPM12's DARTEL (Diffeomorphic Anatomical Registration using Exponentiated Lie Algebra) toolbox. Images were then resampled to 1.5 mm voxels and smoothed with a 4 mm full‐width at half‐maximum (FWHM) Gaussian kernel. Probabilistic tissue maps were visually inspected to assess segmentation quality. Brain predicted age was then estimated using the pretrained brainageR model, which was originally developed using data from 3377 healthy adults aged 18 to 92 years by combining normalized tissue maps, reducing them to 435 principal components, and training a Gaussian process regression model. This pretrained model was applied to vectorized and masked images in the current study to generate a brain‐predicted age score for each participant.

Brain‐PAD was then calculated as the difference between predicted brain age and chronological age, using the whole‐year scale for chronological age. Therefore, a positive brain‐PAD (i.e., older brain age relative to chronological age) suggests advanced brain aging, while a negative brain‐PAD (i.e., younger brain age relative to chronological age) represents preserved brain aging. To account for a potential age bias in brain age estimates, ²⁸ we include chronological age as a covariate in the statistical models.

2.3. Statistical analysis

We assessed the normality of continuous variables through visual inspection and skewness and kurtosis analysis. Plasma biomarkers with right skewness (NfL, p‐tau217, GFAP) were log‐transformed. Outliers in plasma biomarkers were pre‐identified with the following threshold: ²⁵ NFL > 76 pg/mL (n = 2); p‐tau217 > 2.16 pg/mL (n = 2); GFAP > 1010 pg/mL (n = 2).

For each biomarker, we fit a covariate‐adjusted 2 × 2 factorial linear model with binary subjective status (elevated vs. non‐elevated complaints), binary objective status (higher vs. lower performance), and their interaction. Covariates were specified a priori and included age, sex, study site, education, body mass index, APOE ε4 status for plasma biomarkers, and age, sex, study site, education, APOE ε4 status, and IQR for MRI AD signature and brain‐PAD. The primary test of association was the subjective x objective interaction. To localize where differences occurred, we decomposed each model into four prespecified simple‐effect contrasts using marginal means: (1) objective status (higher vs. lower performance) within non‐elevated complaints stratum, (2) objective status (higher vs. lower performance) within elevated complaints stratum, (3) subjective status (elevated vs. non‐elevated complaints) within higher cognitive performance stratum, and (4) subjective status (elevated vs. non‐elevated complaints) within lower cognitive performance stratum. Model‐based marginal means and the four prespecified contrasts were estimated with emmeans in R and are reported as point estimates with 95% confidence intervals (CIs) and P values. Estimated marginal mean difference (EMD) for contrasts was calculated as low−high performance or elevated−non‐elevated complaints within the indicated stratum; positive values indicate higher outcomes in the first group listed. Standardized effect sizes were summarized as Hedges g computed from the model residual variance. For outcomes analyzed on the log scale, estimates and intervals are reported on the log‐transformed scale without back‐transformation.

For completeness, profile‐level means and all pairwise comparisons across the four subjective–objective profiles were summarized as descriptive secondary analysis, with 95% CIs and P values. Inferential conclusions are derived from the prespecified factorial models and the four planned simple‐effect contrasts.

For sensitivity analyses, models were additionally adjusted for self‐reported race (White vs. non‐White), re‐estimated after excluding 54 participants whose cognitive performance was near but did not meet criteria for MCI, and repeated in the subset of participants with high posterior probabilities for subjective classification (P[non‐elevated] ≥ 0.80 or ≤ 0.20), excluding those in the gray zone (0.20–0.80).

All analyses were conducted in R (version 4.2.1), using emmeans for marginal means and contrasts, and effectsize for standardized effects; k‐means clustering used base stats, and mixed modeling used mclust.

3. RESULTS

3.1. Sample characteristics

A total of 648 CU older adults were included. Demographic and clinical characteristics of the sample are shown in Table 1.

3.2. Subjective cognition composite

The best‐fit model for subjective cognitive function was a second‐order model with a general subjective cognitive complaints composite (χ² = 72.006, df = 18, χ²/df = 4.00, P < 0.001, CFI = 0.973, TLI = 0.958, RMSEA = 0.068, SRMR = 0.03). Figure 1 shows the latent factor construct of the optimal model with standardized factor loadings on the paths. All loadings were statistically significant (P < 0.001), and all measures had loadings > 0.55. Factor scores were used for subsequent classification of subjective status.

FIGURE 1.

The best‐fit subjective cognitive complaints factor structure derived from the CFA. Circles depict latent factors, while rectangles reflect observed variables representing questionnaire outcome scores. Single‐headed arrows represent paths of factor loadings, and the double‐headed arrow represents residual correlations. Standardized factor loadings are shown on the single‐headed arrows. Error items were not included in the figure. CFA, confirmatory factor analysis; CFI, Cognitive Function Index; ECOG, Everyday Cognition Questionnaire; PROMIS, Patient‐Reported Outcomes Measurement Information System.

3.3. Subjective–objective profile construction and counts

A two‐component Gaussian mixture model (GMM) applied to the subjective factor scores identified elevated and non‐elevated subjective complaints status, yielding n = 443 in the non‐elevated class and n = 205 in the elevated class. The distribution of posterior probability is shown in Figure S3B in supporting information. Bootstrap evaluation of the posterior probability mapping indicated generally stable separation, with reduced precision confined to the upper P = 0.80 boundary (Figure S3D).

Objective cognitive performance across five domains was classified into two objective classes with k‐means (Figure S1). The two objective clusters differed consistently across all five cognitive domains, with the largest separations in working memory and EF/attentional control, with smaller differences in episodic memory (Figure S4 in supporting information). Class sizes were n = 370 for higher objective performance and n = 278 for lower objective performance.

Crossing subjective and objective status produced four profiles: concordant higher functioning (non‐elevated complaints, higher performance; n = 277), SCD‐like (elevated complaints, higher performance; n = 93), reduced insight (non‐elevated complaints, lower performance; n = 166), and concordant lower functioning (elevated complaints, lower performance; n = 112; Table 2). Key demographics and biomarkers of interest by profile are summarized in Table 2.

TABLE 2.

Key demographics and biomarkers of interest by subjective–objective profile. Profiles were defined by crossing subjective complaint status (non‐elevated vs, elevated) and objective performance (higher vs. lower): concordant higher functioning (non‐elevated complaints, higher performance), SCD‐like (elevated complaints, higher performance), reduced insight (non‐elevated complaints, lower performance), and concordant lower functioning (elevated complaints, lower performance).

Characteristic	Concordant higher functioning	SCD‐like	Reduced insight	Concordant lower functioning	P value*
Number	277	93	166	112
Age, years	69.10 (3.41)	69.14 (3.37)	70.95 (3.97)	70.81 (3.94)	<0.001
Sex					0.423
Male	75 (27%)	23 (25%)	55 (33%)	34 (30%)
Female	202 (73%)	70 (75%)	111 (67%)	78 (70%)
Site					0.29
Pitt	89 (32%)	35 (38%)	61 (37%)	34 (30%)
KU	101 (36%)	33 (35%)	47 (28%)	33 (29%)
NEU	87 (31%)	25 (27%)	58 (35%)	45 (40%)
Education, years	16.80 (2.03)	16.68 (1.92)	15.74 (2.35)	15.68 (2.33)	<0.001
BMI, kg/m2	29.70 (5.97)	28.93 (5.09)	30.21 (5.81)	29.82 (5.61)	0.355
APOE ε4					0.043
Non‐carrier	198 (71%)	78 (85%)	114 (71%)	76 (68%)
Carrier	79 (29%)	14 (15%)	46 (29%)	35 (32%)
Missing	0	1	6	1
Neurobiological markers
NfL, pg/mL	16.27 (5.78)	15.30 (5.27)	17.35 (7.60)	18.66 (8.62)	0.029
Missing	13	6	16	8
AD signature, mm3	65,530.19 (5620.66)	65,811.01 (4824.90)	64,101.31 (5376.61)	63,773.85 (5259.69)	0.002
Missing	0	1	5	1
Brain‐PAD, years	−4.98 (6.75)	−3.64 (6.30)	−2.78 (6.61)	−3.67 (6.50)	0.018
Missing	0	1	5	1
p‐tau217, pg/mL	0.42 (0.25)	0.41 (0.26)	0.45 (0.31)	0.47 (0.32)	0.443
Missing	6	3	10	4
GFAP, pg/mL	170.31 (87.76)	184.55 (112.68)	184.17 (103.63)	187.42 (93.72)	0.324
Missing	13	6	16	7

^*Differences between groups were determined using the Kruskal–Wallis rank sum test for continuous variables and Pearson chi‐squared test for categorical variables.

Abbreviations: AD, Alzheimer's disease; APOE, apolipoprotein E; BMI, body mass index; brain‐PAD, brain predicted age difference; GFAP, glial fibrillary acidic protein; KU, University of Kansas; NEU, Northeastern University; NfL, neurofilament light chain; Pitt, University of Pittsburgh; p‐tau217, phosphorylated tau 217; SCD, subjective cognitive decline.

3.4. Biomarker patterns across profiles

Subjective–objective interaction and standardized mean differences across the four prespecified contrasts are summarized in Figure 2. Effect sizes varied by biomarker family. Markers of neurodegeneration and global brain aging showed the largest separations: brain‐PAD was higher in discordant profiles than in concordant higher functioning, with no further increase in concordant lower functioning; differences in NfL were most prominent when objective performance was poorer in the presence of subjective complaints. The structural MRI AD signature demonstrated a similar directionality of results to NfL. In contrast, p‐tau217 and GFAP showed minimal separation across profiles, with estimates near zero and confidence intervals spanning zero.

FIGURE 2.

Subjective x objective interaction and adjusted simple effects across biomarkers. The forest plot shows standardized mean difference (Hedges g) for each biomarker across the prespecified simple‐effect contrasts. Points indicate effect size estimates, and horizontal lines indicate 95% confidence intervals. Simple effects are displayed using the standardized subjective–objective profile labels within the indicated stratum; positive Hedges g values indicate higher outcomes in the first group listed. For NfL, p‐tau217, GFAP, and brain‐PAD, higher values reflect greater pathological burden or older‐appearing brain, whereas for MRI volumetric AD signature, lower values reflect greater neurodegeneration. AD, Alzheimer's disease; brain‐PAD, brain predicted age difference; CI, confidence interval; GFAP, glial fibrillary acidic protein; MRI, magnetic resonance imaging; NfL, neurofilament light chain; p‐tau217, phosphorylated tau 217; SCD, subjective cognitive decline.

3.4.1. Biomarkers of neurodegeneration

For NfL, the subjective x objective interaction was statistically significant (P = 0.0427). Higher NfL was observed with lower objective performance only when subjective complaints were elevated (lower vs. higher performance within elevated complaints: EMD [log] = 0.130, 95% CI [0.027, 0.232], P = 0.0134; Figure 3A), whereas the corresponding simple effect within non‐elevated complaints was smaller and not significant (lower vs. higher performance within non‐elevated complaints: EMD [log] = 0.002, 95% CI [–0.077, 0.080], P = 0.970).

FIGURE 3.

Neurodegeneration biomarkers across subjective–objective cognitive profiles. A, NfL. B, MRI AD signature. For each biomarker, the top panel shows prespecified simple‐effect contrasts from the adjusted model, displayed using the standardized profile labels within the indicated stratum, with points indicating adjusted mean differences and horizontal bars indicating 95% confidence intervals; exact P values are shown for each contrast. The bottom panel shows the biomarker distribution by profile, with small colored points indicating individual participants, boxplots showing the median, interquartile range, and 1.5 times the interquartile range, and white circles with vertical error bars indicating adjusted estimated marginal means and 95% confidence intervals. Profiles were defined by crossing subjective complaint status (non‐elevated vs. elevated) and objective performance (higher vs. lower): concordant higher functioning (non‐elevated complaints, higher performance), SCD‐like (elevated complaints, higher performance), reduced insight (non‐elevated complaints, lower performance), and concordant lower functioning (elevated complaints, lower performance). AD, Alzheimer's disease; MRI, magnetic resonance imaging; NfL, neurofilament light chain; SCD, subjective cognitive decline.

For the MRI volumetric AD signature, in which higher values indicate greater preservation of AD‐vulnerable regional volumes, the subjective x objective interaction was not statistically significant (P = 0.39), indicating that the association between objective performance and AD signature did not differ significantly by subjective complaints status. Nonetheless, the simple effect estimates followed the same general pattern as NfL, but were attenuated (lower vs. higher performance within elevated complaints: EMD = –1842 mm³, 95% CI [–3231, –453.6], P = 0.0094; lower vs. higher performance within non‐elevated complaints: EMD = –1086 mm³, 95% CI [–2191, 20.1], P = 0.054; Figure 3B).

3.4.2. Global brain aging

For brain‐PAD, the subjective x objective interaction was significant (P = 0.0296). Within higher objective performance, worse subjective status was associated with higher brain‐PAD (elevated vs. non‐elevated complaints within higher performance: EMD = 1.48 years, 95% CI [0.06, 2.89], P = 0.0404). Within non‐elevated ,   complaints, lower versus higher objective performance was likewise associated with higher brain‐PAD (EMD = 2.07 years; 95% CI [0.75, 3.39], P = 0.0021). When one dimension was already in poorer status, the other dimension contributed little additional difference (within lower objective performance or elevated complaints, both P > 0.05). Adjusted profile mean differences are shown in Figure 4.

FIGURE 4.

Brain‐PAD difference across subjective–objective cognitive profiles. The top panel shows prespecified simple‐effect contrasts from the adjusted model, displayed using the standardized profile labels within the indicated stratum. Points show adjusted mean differences in years, and horizontal bars show 95% confidence intervals; exact P values are shown for each contrast. The bottom panel shows brain‐PAD by subjective–objective profile. Profiles were defined by crossing subjective complaint status (non‐elevated vs. elevated) and objective performance (higher vs. lower): concordant higher functioning (non‐elevated complaints, higher performance), SCD‐like (elevated complaints, higher performance), reduced insight (non‐elevated complaints, lower performance), and concordant lower functioning (elevated complaints, lower performance). Small colored points indicate individual participants. In each boxplot, the center line indicates the median, the box indicates the interquartile range, and the whiskers indicate 1.5 times the interquartile range. White circles indicate adjusted estimated marginal means, and vertical error bars indicate 95% confidence intervals. brain‐PAD, brain predicted age difference; SCD, subjective cognitive decline.

3.4.3. p‐tau217 and GFAP markers

Profiles showed minimal separation for p‐tau217 and GFAP (interaction P = 0.81 and 0.64, respectively). None of the simple‐effect contrasts were statistically significant. Adjusted profile mean differences are provided in Figure 5.

FIGURE 5.

AD‐related pathology and astroglial activation biomarkers across subjective‐objective cognitive profiles. A, p‐tau217. B, GFAP. For each biomarker, the top panel shows prespecified simple effect contrasts from the adjusted model, displayed using the standardized profile labels within the indicated stratum, with points indicating adjusted mean differences and horizontal bars indicating 95% confidence intervals; exact P values are shown for each contrast. The bottom panel shows the biomarker distribution by profile, with small colored points indicating individual participants, boxplots showing the median, interquartile range, and 1.5 times the interquartile range, and white circles with vertical error bars indicating adjusted estimated marginal means and 95% confidence intervals. Profiles were defined by crossing subjective complaint status (non‐elevated vs. elevated) and objective performance (higher vs. lower): concordant higher functioning (non‐elevated complaints, higher performance), SCD‐like (elevated complaints, higher performance), reduced insight (non‐elevated complaints, lower performance), and concordant lower functioning (elevated complaints, lower performance). AD, Alzheimer's disease; GFAP, glial fibrillary acidic protein; p‐tau217, phosphorylated tau 217; SCD, subjective cognitive decline.

3.5. Descriptive pairwise profile comparisons

All six pairwise profile comparisons for each biomarker are summarized in Table S2 in supporting information. Patterns were consistent with the simple‐effects results described above.

3.6. Sensitivity results

Sensitivity analyses showed consistent patterns across alternative model specifications (Table S3 in supporting information). When 54 participants whose cognitive performance near MCI thresholds were excluded, statistical support for the interaction terms was attenuated for NfL (P = 0.058) and brain‐PAD (P = 0.077), but the direction of the effects and the qualitative pattern of simple effects were unchanged (Figure S5 in supporting information).

Finally, a total of 189 participants had intermediate posterior probabilities for the non‐elevated subjective complaints class, with 0.20 < P(non‐elevated) < 0.80. Baseline demographic characteristics did not differ significantly between ambiguous and confident cases (Table S4 in supporting information). Analyses restricted to the confident‐only subset (P[non‐elevated] ≥ 0.80 or ≤ 0.20), which excluded those 189 individuals, yielded similar results. The key simple effects remained directionally consistent (Figure S6 in supporting information).

4. DISCUSSION

In this community‐dwelling cohort of older adults, we examined the association between neurobiological markers and jointly defined subjective–objective cognitive profiles. Consistent with our hypothesis that markers of neurodegeneration and global brain aging would differ across profiles, three observations emerged. First, plasma NfL levels were elevated in the concordant lower functioning profile, in which greater subjective complaints co‐occurred with lower objective performance. MRI volumetric AD signature showed a directionally similar reduction in this profile, although the overall interaction was not significant, consistent with a pattern linked to neurodegeneration. Second, brain‐PAD was elevated in both discordant profiles, namely the reduced insight and SCD‐like groups, compared to the concordant higher functioning group. Notably, the concordant lower functioning group did not demonstrate further elevation beyond the discordant groups, suggesting that deviation in either subjective or objective cognition alone was associated with an older‐appearing brain. Third, p‐tau217 and GFAP showed minimal separation across profiles.

Taken together, these findings suggest partially distinct biological mechanisms underlying concordant and discordant configurations of subjective and objective cognition. Elevation of NfL in the concordant lower functioning profile is consistent with processes linked to neuroaxonal injury, whereas the association of brain‐PAD with discordant profiles suggests sensitivity to broader aging‐related changes that are not restricted to overt performance deficits. The absence of differentiation in p‐tau217 and GFAP further indicates that AD‐related pathology and astroglial activation related to neuroinflammation are unlikely to account for these profile distinctions in this cohort. By distinguishing convergence from divergence between perceived and measured cognition, our findings provide a potential framework for understanding the long‐recognized inconsistency between subjective complaints and objective deficits. ²⁹ , ³⁰ Rather than reflecting measurement noise alone, discrepancy between subjective and objective measures may capture biologically meaningful variations across aging trajectories, with global brain aging measures characterizing discordant states and neurodegeneration markers becoming more prominent when both dimensions consistently reflect poorer cognition.

The profile in which subjective complaints and objective deficits converged warrants particular attention in this context. The neurodegeneration pattern observed here aligns with prior evidence that subjective complaints are most prognostically meaningful when accompanied by objective deficits and biological vulnerability. ³¹ , ³² Cohort studies have shown that subjective complaints are more strongly associated with subsequent decline when amyloid positivity or plasma markers of neurodegeneration are present, whereas in their absence, subjective complaints more often relate to mood, sleep, or other psychological or behavioral factors. ³³ , ³⁴ , ³⁵ , ³⁶ In the present study, NfL was elevated in the concordant lower functioning profile, consistent with neuroaxonal injury, and AD signature volume was directionally reduced in the same group. Although the structural marker did not reach statistical significance, the convergent pattern across plasma and MRI measures supports the interpretation that co‐occurring subjective and objective difficulty reflects processes linked to neurodegeneration in clinically unimpaired older adults.

A different pattern emerged for global brain aging. In the present study, brain‐PAD was elevated in both discordant cognitive profiles, including individuals with lower objective performance despite few complaints and those with greater complaints despite preserved performance. Brain‐PAD has been linked to cardiovascular risk, white matter hyperintensities, cardiorespiratory fitness, and mortality, and predicts cognitive change beyond chronological age. ³⁷ , ³⁸ , ³⁹ , ⁴⁰ Those associations support its role as an integrative index of global aging rather than a marker of a single disease process. The pattern observed here suggests that divergence between subjective and objective cognition may correspond to variation in broader aging‐related processes not fully captured by traditional AD‐specific biomarkers. Consistent with this interpretation, p‐tau217 and GFAP did not differentiate the profiles, indicating that AD pathology and neuroinflammation may not fully explain the observed distinctions in this cohort. It is also possible that some participants were on trajectories related to non‐AD mechanisms, which may help explain the limited association with AD‐specific biomarkers. Longitudinal follow‐up with clinical and etiologic diagnostic data will be important for clarifying whether these subjective–objective profiles differentially predict AD and non‐AD outcomes.

Our findings have potential clinical and research implications. First, joint consideration of subjective complaints and objective performance may provide greater specificity regarding underlying neural processes than either measure alone. Individuals with discordance between complaints and performance who also show an older‐appearing brain may represent a subgroup at increased risk for future decline and could be candidates for closer monitoring or preventive strategies targeting vascular and metabolic health. In contrast, concordant lower functioning accompanied by elevated NfL may reflect early neurodegenerative processes detectable in a clinically unimpaired population. This profile‐based framework may help refine risk stratification and inform the selection of biomarker endpoints that correspond to specific neural processes observed within each cognitive profile, although replication and longitudinal validation are needed. Future longitudinal and intervention‐focused analyses will also be important for determining whether exercise or related behavioral factors influence transitions between these subjective–objective profiles. Second, by explicitly modeling the joint subjective–objective profiles, our approach offers a potential explanation for the heterogeneity observed in prior studies, in which subjective and objective cognition were often analyzed separately or hierarchically. Framing subjective and objective cognition as interacting dimensions, rather than as partially overlapping but independently examined constructs, may clarify variability in aging trajectories among CU older adults.

This study has several strengths. First, the objective and subjective measurement coverage is broad. Objective cognition was summarized across five cognitive domains (episodic memory, processing speed, working memory, executive function, and visuospatial function), and subjective status was summarized from four validated self‐report questionnaires into a composite. This approach reduces dependence on any single instrument, improves reliability and construct coverage, mitigates ceiling and floor effects in objective testing, and dampens idiosyncratic response bias in self‐report. Second, profile assignment was data driven rather than using arbitrary cut‐points. Objective performance was classified via k‐means, and subjective status was modeled with a Gaussian mixture that yields posterior probabilities, enabling probability‐based thresholds and confidence‐band sensitivity analyses. Third, the biomarker panel is broad and mechanistically informative. Plasma NfL and structural MRI AD signature index neurodegeneration; brain‐PAD captures global brain aging; p‐tau217 and GFAP index amyloid‐ and tau‐related pathology and astroglial neuroinflammatory processes. Evaluating all five biomarkers together strengthens biological interpretation. Finally, the community‐based nature of the cohort enhances generalizability beyond memory clinic populations that are often enriched for more advanced or overtly symptomatic individuals.

Several limitations merit emphasis. First, the subjective categories were derived from a two‐component Gaussian mixture. Bootstrap analyses indicated limited precision at the upper “few complaints‐confident” boundary, reflecting residual overlap and fewer observations at high scores. Findings that depend on this boundary should therefore be interpreted with caution. Second, sample sizes for discordant profiles were smaller, limiting statistical power for some simple‐effect tests and likely contributing to the non‐significant interaction for the AD signature despite a consistent direction. Third, the cross‐sectional design precludes temporal inference, and therefore, whether discordant states precede concordant lower functioning remains to be tested. Finally, generalizability is still bounded by the age range and relatively high education attainment of the cohort, and the findings may not extend to younger adults or to samples with different demographic or clinical characteristics. Compared to memory clinic cohorts, community‐based samples may also exhibit subtler clinical and biomarker differences, which could attenuate some between‐group associations. In addition, non‐neurologic comorbidities and other unmeasured influences on neuropsychological performance, including factors such as cognitive reserve, testing context, anxiety, and fatigue, may have contributed to subjective–objective discordance. These influences are not unique to the present sample, but represent a broader source of variation in neuropsychological assessment across both clinical and research settings. Because profiles were derived within a screened CU cohort, the higher and lower classifications reflect relative variation rather than clinically defined impairment.

Overall, this study demonstrated that in CU older adults, the joint consideration of subjective and objective cognition is informative for understanding underlying neural processes. Neurodegeneration markers, including elevated NfL and lower volumetric AD signature, were most evident when complaint and test scores were concordant, whereas global brain‐PAD characterized discordant profiles in either direction. AD‐related neuropathology and astroglial markers were not associated with any profile. Together, these findings suggest that joint evaluation of subjective and objective cognition reflects distinct processes related to neurodegeneration and global brain aging in CU older adults.

AUTHOR CONTRIBUTIONS

Kirk I. Erickson, Lu Wan, and Chaeryon Kang participated in the design of the study, and contributed to data analysis and interpretation of the results; Patricio Solis‐Urra, Kelsey R. Sewell, George Grove, Haiqing Huang, Rae Harrison, Thomas K. Karikari, and Anna L. Marsland contributed to data acquisition, analysis, and interpretation of the data; Lu Wan and Chaeryon Kang contributed to methodology and statistical analysis. All authors contributed to the manuscript writing and approved the final version of the manuscript.

FUNDING INFORMATION

This study was funded by the National Institutes of Health (R01 AG053952) awarded to K.I.E., J.M.B., A.F.K., E.M., and to K.I.E. (R35 AG072307).

CONFLICTS OF INTEREST STATEMENT

The authors declare no conflicts of interest relevant to this article. K.I.E. consults for MedRhythms, Inc. and Neo Auvra, Inc. T.K.K. consults for Quanterix Corporation, SpearBio Inc., Neurogen Biomarking LLC, and Alzheon; has served on advisory boards for Siemens Healthineers and Neurogen Biomarking LLC; and has received honorarium from Cell Signaling Technology. T.K.K. has also received royalties from Bioventix for the transfer of specific antibodies and assays to third party organizations. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

All participants provided written informed consent prior to any study‐related procedures.

Supporting information

Supporting Information

Supporting Information