Longitudinal changes in the brain‐age gap in mild cognitive impairment and their relationships with neuropsychological functions and Alzheimer's disease biomarkers

Abstract

INTRODUCTION

The discrepancy between biological and modeled brain ages—the brain‐age gap (BAG)—could indicate potential neuropsychological changes. This study verified if and how longitudinal BAG changes were associated with neuropsychological functions and Alzheimer's disease–related biomarkers in individuals with mild cognitive impairment (MCI).

METHODS

One hundred thirty‐eight individuals with MCI and 103 healthy controls (HCs) with three rounds of magnetic resonance imaging scanning were selected from the Alzheimer's Disease Neuroimaging Initiative. We applied support vector regression on functional connectivity for modeling the brain age and further calculated the BAG.

RESULTS

Longitudinal BAG changes were higher in participants with MCI compared to HCs. Larger BAG fluctuations were correlated with poorer cognitive performance and more severe depressive symptoms in patients with MCI. Neurofilament light chain and phosphorylated tau levels were associated with the longitudinal BAG changes.

DISCUSSION

Present findings demonstrated the necessity of incorporating longitudinal BAG in monitoring the neuropsychological status among cognitively vulnerable populations.

Highlights

• Brain‐age gap (BAG) changes are sensitive indicators of cognitive vulnerability in aging.

• BAG changes were larger in patients with mild cognitive impairment than in the controls.

• Longitudinal BAG changes were associated with worse cognitive‐affective states.

• The plasma neurofilament light chain and cerebrospinal fluid phosphorylated tau levels were associated with the BAG changes.

1. BACKGROUND

The aging population is a global phenomenon. While some older people experience age‐appropriate neurocognitive changes, many present pathological age‐related neurocognitive decline and mild cognitive impairment (MCI). ¹ The annual conversion rate from MCI to Alzheimer's disease (AD) can reach ≈ 15% among clinic samples. ² Considering the phenomenal medical and psychosocial costs associated with loss of productivity and caregiving, acquiring early insights into neuropsychological changes over time is highly beneficial for disease management. ³ Researchers have estimated the pathological development of MCI using various biopsychosocial protective and risk factors. ⁴ , ⁵ These factors offer indirect estimations of brain status and cognitive health, leading to inconsistency in the robustness of associations between these factors and MCI pathological progression. ⁵ Furthermore, it is challenging to quantify these factors on a unified scale.

“Brain age” estimates predicted brain age using an index generated by computational algorithms based on neuroimaging data. Depending on pathological changes in the brain, brain age may differ from chronological age.

The gap between brain age and chronological age—the brain‐age gap (BAG) ⁶ —offers an objective measure useful for estimating brain pathology and MCI progression. Cross‐sectional studies have revealed that, relative to healthy controls, the BAG is larger among patients with MCI and AD. ⁷ The greater the BAG, the higher the risk of developing age‐inappropriate neurocognitive decline and emotional disturbances. ⁸ However, recent aging research on BAG has been primarily based on cross‐sectional studies. There has been strong advocacy for examining longitudinal fluctuations in cognitive performance, brain characteristics, and biomarker levels in estimating the progression of MCI and gaining insight into the rate and mechanisms of conversion to AD. ⁹ , ¹⁰ , ¹¹

AD‐related biological pathways, specifically amyloid beta (Aβ) and tau proteins, ¹² , ¹³ and neurofilament light chain (NfL), ¹⁴ can be associated with cognitive decline and BAG. Indeed, cross‐sectional studies found that BAG estimated by multimodal neuroimaging was positively associated with amyloid pathology detected on positron emission tomography, ¹⁵ global tau protein accumulation, ¹⁶ plasma NfL, ¹⁷ and neurodegenerative changes. ¹⁸ Thus, these biological factors also play a role in longitudinal BAG changes.

This study investigated the relationships among BAG, biological correlates, neurocognitive status, and affective states over time. We investigated the difference in longitudinal changes in BAG between people with MCI and their healthy counterparts. Furthermore, we assessed the relationship between BAG changes and neurocognitive and affective performances. Finally, we evaluated the association between baseline AD pathological proteins and longitudinal BAG changes. We hypothesized that people with MCI would have a significantly larger BAG at each time point compared to the healthy controls. Longitudinally, BAG changes would be greater in patients with MCI than in the controls. Furthermore, larger BAG changes would be associated with worse cognitive performance and more affective disturbances. Finally, AD pathological proteins would be associated with longitudinal changes in BAG.

2. METHODS

2.1. Participants

Participants aged between 55 and 90 were assessed longitudinally at multiple sites in North America through the Alzheimer's Disease Neuroimaging Initiative (ADNI). All participants provided written consent. We selected the participants with stable diagnoses throughout data collection and excluded those with conversion and/or reversion to avoid heterogeneity. We only include participants with at least three rounds of magnetic resonance imaging (MRI) scanning, each > 90 days apart. The average time interval between the first and the third scan was 2.99 years (1091 days) with a standard deviation (SD) of 1.31 years (478 days). To maximize the time interval between baseline biosample collection and subsequent assessments, we selected the functional MRI (fMRI) data from the most recent three time points by the time of inclusion (February 2024). The average time interval between biomarker assessment and the first MRI scan in our study ranges from 2.33 years to 4.71 years for different biomarkers. Due to the limited sample size of AD, only healthy controls (HCs) and patients with MCI were finally included. For detailed inclusion and exclusion criteria, see the ADNI Study Cohort Information (https://adni.loni.usc.edu/data‐samples/adni‐data/study‐cohort‐information/).

2.2. Measures

2.2.1. Neuropsychological assessment

General cognitive status was assessed by the Alzheimer's Disease Assessment Scale‐11‐item cognitive subscale (ADAS‐Cog 11). ¹⁹ Domain‐specific cognitive functions in memory, ²⁰ executive function, ²¹ language, and visual‐spatial ability ²² were evaluated using composite scores from the ADNI neuropsychological battery (Tables S1–S4 in Supporting Information). Neuro‐affective evaluations included the Neuropsychiatric Inventory (NPI) ²³ and the Geriatric Depression Scale (GDS). ²⁴ For more details, see Section S1 in supporting information and ADNI documentation (https://adni.loni.usc.edu/help‐faqs/adni‐documentation/). Neuropsychological outcomes were defined as (1) the longitudinal changes in performance (annual change rate or SD; see Formulas 1 and 2 in section 2.3.3), and (2) the performance at the final fMRI scan.

2.2.2. Biomarker measurement

Cerebrospinal fluid (CSF) and plasma samples were collected at a standardized time after a minimum 6 hour fast. Biomarkers were analyzed only if sample size was > 75 for statistical power. These included plasma levels of tau phosphorylated at threonine‐181 (p‐tau181), Aβ 42/40 ratio (Aβ 42/40), oligomer Aβ (OAB), NfL, and CSF levels of p‐tau and Aβ 42/40. Measurements were provided by the ADNI Biomarker Core (full list: Fluid Biomarker, https://adni.loni.usc.edu/data‐samples/adni‐data/biofluid‐biomarker/).

2.3. BAG calculation

2.3.1. Imaging processing

MRI data were collected from 3T scanners (https://adni.loni.usc.edu/data‐samples/adni‐data/neuroimaging/mri/ for details). The resting‐state fMRI (rs‐fMRI) was preprocessed using DPARSF 5.4 under DAPBI ²⁵ in MATLAB. The first five volumes of scans were removed for signal equilibrium and participants’ adaptation. We then applied slice timing correction, realignment, co‐registration, segmentation, normalization, head motion correction, detrending, nuisance covariate regression (Friston's 24 head motion parameters, global signal, white matter, and CSF signals), and filtering (0.01∼0.1 Hz). During normalization, poorly aligned images were manually reoriented. Participants with head motion > 3.0 mm or 3.0° were excluded (N = 15 for training dataset, and N = 12 for T1, N = 15 for T2, N = 11 for T3). After the preprocessing, resting‐state functional connectivity (rs‐FC) matrices (with Fisher Z transformation) at three time points were extracted using Power‐264 3mm functional brain atlas ²⁶ in GRETNA.

RESEARCH IN CONTEXT

1. Systematic review: We searched PubMed using terms: ((Brain‐Age Gap[tiab] or Brain‐Age Delta[tiab] or BrainAGE[tiab]) AND longitudinal). Most studies focused on longitudinal changes in brain‐age gap (BAG) in healthy and pathological aging or neurological and neuropsychiatric disorders, examining associated bio‐psycho‐social factors. Four studies had similar objectives, using structural or positron emission tomography imaging features for model construction and identifying differences in BAG changes across disease stages.

2. Interpretation: Our study bridges a gap by revealing that while cross‐sectional BAG, derived from brain functional characteristics, were similar between healthy controls and patients with mild cognitive impairment (MCI) with stable diagnoses, longitudinal BAG changes were larger in the MCI group and linked to worse neuropsychological outcomes. Baseline Alzheimer's disease pathology levels were associated with future BAG changes.

3. Future directions: Future studies investigating BAG changes using advanced deep learning with longitudinal multimodal neuroimaging and multi‐level bio‐psycho‐social data to predict disease conversion or reversion would offer significant clinical implications.

2.3.2. Model construction and parameter optimization

The BAG model was trained using support vector regression (SVR) with 10‐fold cross‐validation in MATLAB, using rs‐FC from a separate cohort of cognitive HCs from ADNI (N = 201, mean age = 71.31, SD = 8.20, range 56–95). In ADNI cohorts, only HCs with matching image acquisition parameters and not in the testing set were included. The age range of the training and testing datasets was comparable. Least absolute shrinkage and selection operator (LASSO) regression was used for feature selection, ²⁷ followed by hyperparameter optimization. The final model was selected based on the lowest mean squared error (MSE) score.

2.3.3. BAG calculation and longitudinal changes quantification

The optimized model was used to estimate predicted brain age and calculate the BAG at three time points by subtracting the predicted brain age from the chronological age. To correct age‐related bias in BAG ²⁸ and account for demographic and multi‐site imaging heterogeneity, ²⁹ we used the linear mixed‐effects model to control for chronological age, sex, and education as the fixed effects and site as a random effect on the BAG. The residual of the model was extracted as the final BAG. We quantified longitudinal BAG characteristics using (1) average annual change rate and (2) SD among the three time points.

$$\begin{array}{ccc}& & \mathrm{A}verageannualchangerate\hfill \\ & & {\displaystyle =\frac{{\displaystyle \frac{\left(\mathrm{B}AGT3-T2\right)}{\mathrm{Y}earbetweenT2T3}+\frac{\left(BAGT2-T1\right)}{YearbetweenT1T2}}}{2}}\hfill \end{array}$$ (1)

$$\begin{array}{ccc}& & \mathrm{S}D\hfill \\ & & =\frac{\sqrt{{\left(BAGT1-BAGmean\right)}^2+{\left(BAGT2-BAGmean\right)}^2+{\left(BAGT3-BAGmean\right)}^2}}{\sqrt{2}}\hfill \end{array}$$ (2)

2.4. Hypothesis testing

Group differences in BAG were tested via two‐sample t tests at all time points and for longitudinal BAG changes between HCs and MCI. Next, we performed Pearson correlation analysis between longitudinal BAG changes and changes in cognitive status and affective states, as well as their levels at the final time point. Pearson correlation analyses were conducted to assess the relationship between biomarker levels at baseline enrollment and longitudinal BAG changes. Significant results were validated with 10,000 times bootstrapping at a 95% confidence interval (CI) to ensure the robustness of the results (note that all CIs listed below are 95% CI). Statistical analyses were conducted in MATLAB. Results visualization was achieved with BrainNet in MATLAB, ³⁰ ggplots2, GGally, and ggpubr in RStudio.

3. RESULTS

3.1. Demographic characteristics of the participants

There were 103 and 138 individuals in the HC and MCI groups, respectively (see Table 1 for details). The chronological ages for each group at the three time points did not show any significant difference, t (239) = 0.27, p = 0.78 at T1; t (239) = 0.68, p = 0.50 at T2; t (239) = 1.08, p = 0.28 at T3. No group difference was observed in sex (p = 0.37) and education (p = 0.14).

TABLE 1.

Information on the demographic characteristics, brain‐age gap, cognitive status, affective state, and baseline plasma and cerebrospinal fluid biomarker levels in two groups.

	Time point 1		Time point 2		Time point T3
	HC	MCI	HC	MCI	HC	MCI
Demographics	Sample size		N = 103	N = 138
Age	74.25 (7.10)	73.99 (7.63)	75.81 (7.07)	75.16 (7.67)	77.67 (7.05)	76.63 (7.60)
Age range	(57.50, 90.30)	(58.50, 91.50)	(60.00, 93.40)	(56.70, 91.60)	(57.00, 92.60)	(57.00, 93.60)
Sex	F = 52/M = 51			F = 61/M = 78
Education	16.65 (2.35)			16.17(2.64)
Brain‐age gap
Brain‐age gap	0.14 (5.17)	2.4e‐14 (6.32)	−0.018 (4.75)	4e‐14 (5.61)	0.11 (5.15)	3.7e‐14 (5.62)
Annual change rate	−0.14 (2.71)			−0.05 (4.78)
Standard deviation	3.56 (1.85)			4.33 (2.57)
Neurocognitive status
ADAS	5.91 (3.11)	10.03 (6.46)	5.78 (3.10)	10.62 (6.79)	5.92 (3.64)	12.54 (9.81)
Memory	1.30 (0.70)	0.47 (0.96)	1.28 (0.76)	0.44 (1.05)	1.27 (0.81)	0.22 (1.22)
Executive function	1.02 (0.84)	0.50 (0.98)	0.91 (0.81)	0.42 (0.94)	0.77 (0.78)	0.27 (0.13)
Language	1.01 (0.68)	0.49 (0.83)	0.88 (0.73)	0.34 (0.91)	0.89 (0.78)	0.21 (1.08)
Visual‐spatial ability	0.15 (0.72)	0.01 (0.80)	0.24 (0.56)	0.04 (0.81)	0.15 (0.70)	0.05 (0.78)
Affective state
Neuropsychiatric Inventory	0.79 (1.87)	4.57 (6.17)	1.49 (2.89)	4.65 (6.55)	1.92 (4.31)	5.80 (9.70)
Geriatric Depression Scale	0.71 (1.17)	2.05 (2.21)	1.10 (1.77)	2.05 (2.15)	1.54 (2.17)	2.42 (2.39)
Biomarkers (Baseline)
Plasma NfL	34.43 (12.13)			33.11 (12.21)
Plasma p‐tau	13.49 (5.92)			14.47 (7.62)
Plasma Aβ 42/40	0.13 (0.01)			0.12 (0.01)
Plasma OAB	205.96 (57.68)			198.50 (49.97)
CSF p‐tau	23.67 (9.38)			32.44 (17.00)
CSF Aβ 42/40	0.10 (0.01)			0.07 (0.03)

Note: We reported the mean value for continuous variables and the count for categorical variables. The standard deviation was reported in brackets.

Abbreviations: Aβ, amyloid beta; ADAS, Alzheimer's Disease Assessment Scale; CSF, cerebrospinal fluid; HC, healthy control; MCI, mild cognitive impairment; NfL, neurofilament light chain; OAB, oligomer amyloid beta; p‐tau, phosphorylated tau.

3.2. Model performance

The best SVR model achieved excellent performance with mean absolute error (MAE) = 4.58, root mean squared error (RMSE) = 5.35, MSE = 28.67, with hyperparameters of box constraint = 0.001, linear kernel with kernel scale = 0.012, epsilon = 4.291. The best model included 21 pairs of rs‐FC (Figure 1).

FIGURE 1.

Visualization of 21 pairs of functional connectivity as features for the final model. Among the 34,716 pairs of functional connectivity, 21 pairs were selected as features for the final best support vector regression model. The functional connectivities were located around the temporal and occipital lobes, regions within the default mode network (i.e., medial prefrontal cortex, posterior parietal cortex), and the cerebellum.

3.3. Difference in cross‐sectional BAG

There was no difference in the cross‐sectional BAG between HCs and individuals with MCI at all three time points (Figure 2), t (239) = 0.19, p = 0.85 at T1; t (239) = −0.026, p = 0.98 at T2; t (239) = 0.16, p = 0.87 at T3.

FIGURE 2.

Brain‐age gap at three time points. HC, healthy control; MCI, mild cognitive impairment.

3.4. Difference in longitudinal changes in BAG

The SD of the BAG among the three time points was significantly higher in individuals with MCI than in HC (Figure 3A), t (239) = −2.57, p = 0.011. There was no difference in the annual change rate between the two groups (Figure 3B), t (239) = −0.16, p = 0.87.

FIGURE 3.

Standard deviation and annual change rate of the brain‐age gap from the two groups. HC, healthy control; MCI, mild cognitive impairment.

3.5. Longitudinal relationship between brain‐age gaps and cognitive status

A higher annual change rate of ADAS‐Cog 11 was associated with a larger annual change rate of BAG in MCI (Figure 4A), r  = 0.32, p < 0.001, CI = [0.0091, 0.6361], but not in HC, r = 0.11, p = 0.29. The SD of ADAS‐Cog 11 was not associated with the SD of BAG in either group, r = –0.13, p = 0.19 in HC and r = 0.024, p = 0.78 in MCI. The larger annual change rate of BAG appeared to be associated with higher ADAS‐Cog 11 (i.e., worse performance) at T3 in MCI, r = 0.18, p = 0.041, but became insignificant after bootstrapping CI = [−0.0075, 0.4862]. The annual change rate of BAG was not associated with ADAS‐Cog 11 at T3 in HC, r = −0.045, p = 0.66. The SD of BAG was not associated with ADAS‐Cog 11 at T3 in either group, r =−0.13, p = 0.21 in HC and r = 0.023, p = 0.79 in MCI.

FIGURE 4.

Correlation between brain‐age gap change rate and Alzheimer's Disease Assessment Scale change rate, memory, and visual‐spatial performance at T3. ADAS: Alzheimer's Disease Assessment Scale‐Cognitive 11; HC, healthy control; MCI, mild cognitive impairment.

Interestingly, the annual change rate of memory performance composite score was negatively associated with the annual change rate of BAG in MCI, r = −0.38, p < 0.001, CI = [−0.6296, −0.1040], but not in HC, r = 0.006, p = 0.95. The SD of the memory performance composite score was not associated with the SD of the BAG in either group, r = 0.054, p = 0.59 in HC and r = 0.036, p = 0.68 in MCI. A higher annual change rate of BAG was associated with poorer memory at T3 in the MCI group (Figure 4B), r  = −0.19, p = 0.04, CI = [−0.3837, −0.0023], but not in HC, r = 0.051, p = 0.62. The SD of the BAG was not associated with memory performance at T3 in either group, r = 0.20, p = 0.05 in HC and r = 0.004, p = 0.97 in MCI.

The annual change rate of the visual‐spatial domain composite score was not associated with the annual change rate of BAG in either group, r = −0.13, p = 0.21 in HC, r = −0.056, p = 0.55 in MCI. The SD of the visual‐spatial score was not associated with the SD of BAG in either group, r = −0.018, p = 0.86 in HC, r = 0.086, p = 0.32 in MCI. A higher annual change rate of BAG was associated with lower visual‐spatial score at T3 in the MCI group (Figure 4C), r  = −0.25, p = 0.006, CI = [−0.4440, −0.0864], but not in HC, r = −0.092, p = 0.38. The SD of BAG was not associated with visual‐spatial score at T3 in either group, r = 0.12, p = 0.23 in HC and r = −0.085, p = 0.36 in MCI.

Longitudinal changes in BAG were not related to changes in executive functions or their performance at the final assessment time. The annual change rate of language domain composite score was negatively associated with the annual change rate of BAG in MCI, r = −0.26, p = 0.005, CI = [−0.5296, −0.0284], but not in HC, r = 0.039, p = 0.71. Detailed domain‐specific results are listed in Supporting Information.

3.6. Longitudinal relationship between BAG and affective states

A higher annual change rate of NPI total score was associated with a larger annual change rate of BAG in HC (Figure 5A), r  = 0.22, p = 0.046, CI = [0.0634, 0.4133], but not in MCI, r = −0.019, p = 0.85. The SD of NPI was not associated with the SD of BAG in either group, r = 0.0074, p = 0.94 in HC and r = 0.012, p = 0.90 in MCI. A higher annual change rate of BAG was associated with a higher NPI total score at T3 in the HC group (Figure 5B), r  = 0.21, p = 0.035, CI = [0.0845, 0.3673], but not in MCI, r = 0.03, p = 0.74. The SD of BAG was not associated with NPI at T3 in either group, r = −0.052, p = 0.61 in HC and r = 0.084, p = 0.35 in MCI. Subdomain results are listed in Supporting Information.

FIGURE 5.

Correlation between the brain‐age gap change rate and the change rate of the Neuropsychiatric Inventory total score, the total score of the Neuropsychiatric Inventory, and the Geriatric Depression Scale at T3. HC, healthy control; MCI, mild cognitive impairment.

Notably, a higher annual change rate of the GDS was associated with a larger annual change rate of BAG in MCI (Figure 5C), r  = 0.31, p = 0.0005, CI = [0.0062, 0.6052], but not in HC, r = −0.16, p = 0.11. The SD of GDS was not associated with the SD of BAG in either group, r = −0.12, p = 0.22 in HC and r = 0.028, p = 0.75 in MCI. The annual change rate of BAG was not associated with GDS at T3 in either group, r = 0.30, p = 0.77 in HC, r = 0.11, p = 0.21 in MCI. The SD of BAG was not associated with GDS at T3 in either group, r = −0.14, p = 0.18 in HC and r = 0.17, p = 0.05 in MCI.

3.7. Relationship between baseline biomarkers and longitudinal changes in BAG

Plasma NfL at baseline enrollment was negatively associated with the SD of BAG among all the subjects (Figure S1A in supporting information), r = −0.19, p = 0.026, CI = [−0.3243, −0.0416]. Further analysis revealed that a higher baseline plasma NfL was also associated with larger BAG at T1 among HC (Figure S1B), r = −0.43, p = 0.002, CI = [0.1801, 0.6059]. The plasma level of OAβ was significantly related to the SD of BAG in HC, but became non‐significant after the bootstrapping, r = −0.25, p = 0.049, CI = [−0.4531, 0.0100]. However, the plasma Aβ 42/40 ratio and p‐tau was not associated with longitudinal changes in BAG.

In healthy controls, CSF p‐tau at baseline enrollment was negatively associated with the SD of BAG (Figure S1C), r = −0.23, p = 0.042, CI = [−0.4080, −0.0091]. The BAG changes were not related to the baseline CSF Aβ 42/40 ratio.

4. DISCUSSION

Our findings confirmed that longitudinal relative to cross‐sectional changes in the BAG were sensitive indicators of neurocognitive vulnerability. AD‐related proteins (NfL and p‐tau) were significantly associated with BAG changes, and BAG variability was higher in the MCI cohort than HCs. In patients with MCI, BAG fluctuations correlated with cognitive decline and affective disturbances, and were negatively associated with memory and visual‐spatial ability performance at the last measurement.

4.1. Longitudinal changes in BAG and MCI pathology

Although cross‐sectional BAG has been linked to future cognitive decline, ⁸ , ³¹ few studies have investigated its longitudinal pattern and the role in neuropsychological functioning. Our results verified greater BAG changes in patients with MCI than HCs, consistent with the observation of accelerated brain aging among those with progressive MCI. ³² Different longitudinal versus cross‐sectional estimations emphasize the advantage of incorporating longitudinal observations of BAG when predicting MCI pathology and neurocognitive decline in cognitively vulnerable populations.

Although some studies reported significant cross‐sectional BAG differences among patients with AD, MCI, and HCs, ⁷ , ¹⁵ our findings did not reveal such differences between those with MCI and HCs, aligning with previous evidence. ³³ Brain or biological age prediction based on epigenetic or structural brain features ³¹ , ³⁴ also reported associations between cognitive decline and accelerated brain aging. Discrepancy in cross‐sectional comparisons might stem from different data modalities for modeling, as multi‐modal or epigenetic‐based prediction showed larger effects. Additionally, considering the BAG exhibited more rapid increases in individuals with progressive MCI or AD compared to HCs or those with stable MCI, ³² our MCI participants’ impairment may not have reached the severity threshold for detectable cross‐sectional BAG differences. Overall, our brain‐age performed comparably or better than existing models across diverse age and cognitive impairment levels (details in Table S5 in supporting information).

4.2. Longitudinal changes in BAG and neuropsychological functions

We observed that larger BAG variations were associated with worse future general cognitive status and steeper decline or poor future status in memory and visual‐spatial abilities, especially in those with MCI. This highlights the effectiveness of leveraging the longitudinal BAG changes in managing neurocognitive vulnerability and progression. This aligns with the understanding of age‐related decline in visuospatial performance ³⁵ and declarative memory. ³⁶ BAG changes showed weaker associations with executive and language functions, suggesting different paces of neurocognitive changes. For example, language function is relatively more resistant to the negative impact of aging. ³⁷ , ³⁸ Nonetheless, longer longitudinal monitoring may clarify the relationship between changes in BAG and executive ³⁹ and language functions. ⁴⁰

Neuropsychiatric symptoms, particularly depression and apathy, are common in those with dementia and MCI. ⁴¹ , ⁴² Our findings indicate that longitudinal changes in the BAG could estimate the levels and progression of affective disturbances in patients with MCI—larger BAG changes corresponded to more severe depressive and apathy symptoms. This extends previous research on BAG in depression. ⁴³ We noticed that stronger associations between annual NPI total score changes and BAG changes in HCs than those with MCI may reflect general neurobiological changes driven by aging in the HCs, while MCI cases may show predominant affective symptoms as pathology progresses. This also suggests that neuropsychiatric symptoms in our MCI participants remained negatively stable over time.

Overall, despite MCI participants being in preclinical or prodromal stages of AD, their reported neuropsychiatric symptoms already indicated brain malfunctioning that hindered effective affective regulation and cognitive functioning, ⁴⁴ , ⁴⁵ which can be sensitively detected by closely monitoring the longitudinal BAG changes.

4.3. Biological correlates of longitudinal changes in BAG

Our research identified AD‐related biological pathways, particularly plasma NfL, as potential risk indicators of longitudinal BAG changes, even in the absence of clinical symptoms.

We observed baseline NfL negatively associated with BAG changes only in HCs, despite NfL's generally detrimental effect on cognition. ⁴⁶ , ⁴⁷ Given its link to poorer cognitive functions and white matter hyperintensity at a single time point but not their progression, ⁴⁷ we speculate that NfL indicates abnormal brain aging in axonal loss at an earlier stage rather than later accelerated changes. The Aβ 42/40 ratios in plasma and CSF were not associated with longitudinal BAG changes, which might be attributed to fluctuations in biomarker levels over time or the silence of Aβ pathways underlying BAG changes. We did not observe the prognostic role of p‐tau on cognitive decline because its baseline level was linked to smaller BAG changes. ⁴⁸ We speculated that individuals with a higher p‐tau might perceive cognitive symptoms earlier and then proactively adopt preventive strategies to mitigate this negative impact over time, considering participants’ pathological processes remained stable over 10 to 15 years between baseline assessment and BAG estimation. Alternatively, early tau presence in cognitively healthy individuals might signify a protective mechanism counteracting accelerated brain aging, unlike its detrimental impact in AD. ⁴⁹

4.4. Limitations

First, our BAG prediction model used rs‐fMRI and a straightforward machine‐learning approach for clinical practicality, with moderate (7 or 10 minutes) to balance motion risk and scan length. Future studies using advanced deep learning tools with longer (> 30 minutes), ⁵⁰ harmonized, multi‐site, multimodal data would improve reliability and predictive performance. Second, to maximize sample size and longitudinal coverage, we set the minimum interval between scans at 90 days, and some participants had relatively short total intervals (256 days), warranting caution in interpretation. Third, we studied subjects with stable cognitive status to reduce heterogeneity. Future research on longitudinal BAG changes and disease conversion or reversion would help identify high‐risk preclinical populations. Fourth, due to limited biomarkers and sample size, we focused on key AD‐related pathological biomarkers at baseline. Future research should explore genetic, biological, and neuropsychological pathways to better understand longitudinal BAG changes and inform prevention strategies.

CONFLICT OF INTEREST STATEMENT

The authors have no conflicts of interest to declare.

CONSENT STATEMENT

All human subjects provided informed consent. Samples of the consent form can be found in the ADNI Documentation section on the ADNI official website (https://adni.loni.usc.edu/help‐faqs/adni‐documentation/).

Supporting information

Supporting Information

Supporting Information

Jin RR, Gu Y, Lee TMC. Longitudinal changes in the brain‐age gap in mild cognitive impairment and their relationships with neuropsychological functions and Alzheimer's disease biomarkers. Alzheimer's Dement. 2025;17:e70180. 10.1002/dad2.70180