EOAD-Signature Atrophy Predicts Dementia in Early-Onset MCI due to Alzheimer Disease

Abstract

Background and Objectives

Early-onset Alzheimer disease (EOAD) is associated with substantial variability in clinical progression, and reliable biomarkers to predict the transition from mild cognitive impairment (MCI) to dementia remain limited. Structural MRI measures have demonstrated prognostic value in late-onset Alzheimer disease, but their utility for predicting progression in EOAD is less well understood. The goal was to examine whether baseline cortical atrophy predicts progression to dementia in patients with MCI because of EOAD.

Methods

This study included a well-characterized cohort of patients with EOAD enrolled in the large multisite natural history Longitudinal Early-Onset Alzheimer's Disease Study. Participants underwent standardized clinical assessments and structural MRI at baseline. Participants were aged between 40 and 64 years with biomarker-supported sporadic EOAD at the MCI stage. Cortical atrophy was measured within the EOAD-signature, a set of predominantly parieto-temporal regions showing greater atrophy in EOAD than in controls. Clinical severity was measured with the global Clinical Dementia Rating. Cox proportional hazards models estimated the association between baseline EOAD-signature atrophy burden and the hazard of progression to dementia over time. We evaluated whether EOAD-signature atrophy improved prognostic performance beyond baseline clinical severity using likelihood ratio tests, Akaike Information Criterion (AIC), and Harrell concordance index.

Results

A total of 130 patients with MCI due to EOAD (mean age 59.6 ± 4.1 years; 49% female) and 97 cognitively normal controls (mean age 56.9 ± 6.0 years; 64% female) were included. Greater baseline atrophy within the EOAD-signature predicted faster progression to dementia (hazard ratio [HR] = 1.24 per 1-SD increase in atrophy; 95% CI 1.13–1.37; p < 0.002). Adding EOAD-signature atrophy burden to a model including baseline clinical severity significantly improved model fit (ΔAIC = −4.5; likelihood ratio test p = 0.011).

Discussion

Baseline cortical atrophy within the EOAD-signature predicts progression from MCI to dementia in EOAD and provides prognostic information beyond baseline clinical severity. These findings support the potential value of EOAD-signature atrophy as an MRI-based biomarker for individualized prognostication and clinical trial stratification.

Introduction

Sporadic early-onset Alzheimer disease (EOAD) is defined by the presence of Alzheimer disease (AD) pathology with clinical onset younger than 65 years.¹ The clinical profile of EOAD is characterized by greater impairment in nonamnestic cognitive domains^2,3 and faster decline compared with late-onset AD (LOAD).⁴ Although this relatively faster rate of progression is true as an average, the trajectory of deterioration varies widely between patients,^5-7 and the field lacks sensitive tools to predict the timing of the loss of functional independence as the patient's cognitive functional status progresses from mild cognitive impairment (MCI) stage to mild dementia and beyond. This is an important unmet need because the capability to estimate disease progression in a personalized manner is critical to optimize patient selection for clinical trials, tailor management and support programs, and establish appropriate goals of care.

Prior work has indicated that baseline clinical and demographic characteristics predict faster disease progression in EOAD, including education,^8-11 gender,¹⁰ and, in particular, performance on neuropsychological measures.^12,13 However, neuropsychological measures are subject to biases, such as practice and floor effects, or individual variability often observed in longitudinal analyses.

Neuroimaging biomarkers could offer an alternative method for prognostication in patients with EOAD at the clinical stage of MCI. Previous research highlights the potential of using baseline imaging features to predict the trajectory of clinical decline in various neurodegenerative syndromes.^14-18 One such strategy is to study the prognostic significance of neurodegeneration within specific large-scale functional networks (e.g., the default mode, salience, frontoparietal control, limbic, dorsal attention, and language networks).¹⁷ Progressive neurodegeneration within these networks likely underlies the loss of cognitive function in multiple domains that leads to the loss of functional independence in EOAD. For instance, in a cohort of individuals with early-stage primary progressive aphasia (PPA), we found that the extent of cortical atrophy within the language, frontoparietal, and limbic networks was an important predictor of progression from MCI to dementia.¹⁷ In another study focusing on individuals with posterior cortical atrophy (PCA), atrophy burden within the dorsal attention network predicted the rate of clinical progression over a follow-up period spanning an average of ∼3 years.¹⁹

Another potential strategy for prognostication is to quantify the magnitude of atrophy in distributed regions that are particularly vulnerable in EOAD using data-driven, disease-specific regions of interest. This approach acknowledges that EOAD-related neurodegeneration often extends beyond the boundaries of several canonical large-scale functional networks, reflecting the spatial heterogeneity of the disease. In a sample of patients with sporadic EOAD at the MCI or mild dementia stage, we recently defined an MRI biomarker of neurodegeneration—the EOAD-signature—that is atrophic within the entire group and includes areas primarily in medial and lateral parietal cortex and posterior lateral temporal cortex, with lesser involvement of the lateral prefrontal cortex and fusiform gyrus.²⁰ This biomarker demonstrated clinical validity by showing a robust correlation with the magnitude of cognitive and functional impairment at baseline.²⁰ However, whether this biomarker can also predict future loss of functional independence in EOAD has not yet been examined.

Here, we applied a predictive modeling framework to examine the utility of baseline cortical atrophy within the EOAD-signature for predicting the progression to dementia in patients with EOAD at the MCI stage. We hypothesized that baseline atrophy within the EOAD-specific signature would provide incremental prognostic information for progression from MCI to dementia beyond baseline clinical severity. We also compared the prognostic utility of the EOAD-signature MRI measure with that of standard morphometric measures as well as functional network-specific measures. The development of imaging biomarkers for prognostication in EOAD could provide better information to clinicians as well as patients and their families about what to expect and enable personalized interventions to putatively improve clinical outcomes.

Methods

Participants

All diagnoses within Longitudinal Early-Onset Alzheimer's Disease Study (LEADS) are made through consensus criteria (see Apostolova et al., 2021).²¹ Data from 130 patients with EOAD at the MCI stage and 97 cognitively normal (CN) age-matched individuals were obtained from the LEADS.²¹ Participants were aged between 40 and 64 years at enrollment, fluent in English, had a knowledgeable informant, were in good general health, and were without a history of other neurologic or psychiatric disorders. Individuals with known familial genetic mutations were excluded.²² CN participants (M_age = 56.85 ± 5.97, 35M/62F) did not have cognitive deficits on neuropsychological testing. Full demographic and clinical characteristics are summarized in Table 1.

Table 1.

Demographic and Baseline Clinical Features of the MCI due to Patients With EOAD

	EOAD	Controls
N	130	97
Sex (M/F)	66/64	35/62
Education	15.58 ± 2.27	16.67 ± 2.36
Age at first scan	59.59 ± 4.07	56.85 ± 5.97
Race/ethnicity
White and non-Hispanic	155(88.5%)	65 (67%)
White and Hispanic	7(5.4%)	7 (7.2%)
Black and non-Hispanic	8(6.1%)	14 (14.4%)
Black and Hispanic	—	1 (1%)
Asian and non-Hispanic	—	6 (6.2%)
Mixed and non-Hispanic	—	2 (2.1%)
Mixed and Hispanic	—	2 (2.1%)
Diagnostic breakdown	—
Amnestic-predominant single or multidomain	104(80%)	—
Posterior cortical atrophy	10(7.7%)	—
Primary progressive aphasia	7(5.3%)	—
Nonamnestic single or multidomain	8(6.2%)	—
Unknown	1(0.7%)	—
CDR	0.5	0
CDR-SB	2.7 ± 0.96	0
MMSE	24.02 ± 3.59	29.17 ± 0.94
Amyloid PET centiloid values	97.17 ± 29.5	6.17 ± 14.21
Progression to dementia (yes/no)	84/46	—
Duration of follow-up after baseline (mo) (all)	21.73 ± 11.12	—
Duration of follow-up after baseline (mo) (individuals who progressed to dementia)	20.37 ± 9.97	—
Duration of follow-up after baseline (mo) (individuals who remained classified at MCI at follow-up)	24.26 ± 12.7	—

Abbreviations: CDR = Clinical Dementia Rating; CDR-SB = Clinical Dementia Rating scale Sum of Boxes; EOAD = early-onset Alzheimer disease; MCI = mild cognitive impairment; MMSE = Mini-Mental State Examination.

Values are presented as mean (SD).

All LEADS participants underwent neuroimaging analysis, including structural MRI and amyloid PET using 18F-Florbetaben. Amyloid PET positivity in patients with EOAD was determined using both visual reads and global standardized uptake value ratio quantification (eMethods).

Participants included in this analysis were required to also meet the following inclusion criteria: (1) mild cognitive impairment at baseline, defined by a global Clinical Dementia Rating (CDR) score of 0.5; (2) availability of baseline research MRI data; and (3) a minimum of 2 longitudinal research visits, including baseline and at least one annual follow-up. Participants were enrolled between 2018 and 2023, with follow-up through 2024. Progression to dementia was operationalized as the first follow-up visit at which a participant met criteria for a CDR global = 1. The mean time between the baseline clinical assessment and MRI scan was 17.74 days.

Neuroimaging Data Acquisition and Preprocessing

MRI data were acquired with a 3.0 T scanner (14 sites used Siemens scanners, 3 used GE scanners, and 1 used a Phillips scanner) using ADNI-3 harmonized protocols after certification of each site using phantom and human scans. Acquired sequences included a high-resolution 3D T1-weighted sequence (e.g., Siemens MPRAGE parameters are TR = 2,300 ms, TE = 2.98 ms, flip angle = 9°, slice thickness = 1 mm, field of view = 240 × 256 mm, 208 sagittal slices, and 2× acceleration).²⁰

Structural 3D T1-weighted MRI scans were processed using FreeSurfer (version 6.0) to derive cortical thickness estimates. All images underwent visual inspection to identify major artifacts and assess overall quality before processing, as previously described.²⁰ Standard FreeSurfer preprocessing included intensity normalization, skull stripping, automated white matter segmentation, and reconstruction of the gray–white and pial surfaces. Cortical surface reconstructions were visually inspected for accuracy in each participant. Cortical thickness was calculated as the closest distance from the gray/white boundary to the gray/CSF boundary at each vertex on the tessellated surface. Whole-brain cortical thickness maps were also registered to the fsaverage template and smoothed using a geodesic Gaussian kernel (15-mm FWHM).²⁰

Next, we investigated the magnitude of AD-related cortical atrophy in EOAD within cortical regions identified in the EOAD-signature regions of interest (ROI) label, as described in the study by Touroutoglou et al. (Figure 1A).²⁰ The EOAD-signature was originally defined in a discovery sample from Massachusetts General Hospital and replicated in an independent LEADS validation cohort that included participants from this study. Using spherical surface registration to fsaverage, 8 EOAD-signature cortical ROIs were projected onto each participant's native cortical surface: the (1) inferior parietal lobule, (2) precuneus, (3) caudal lateral temporal cortex, (4) posterior cingulate cortex, (5) superior parietal lobule, (6) midlateral temporal cortex, (7) middle frontal gyrus, and the (8) superior frontal gyrus. The mean cortical thickness was calculated for each ROI by averaging vertex-wise thickness values within the labeled region. A composite EOAD-signature measure was then derived by combining all ROIs and averaging cortical thickness across the resulting label.

Figure 1. EOAD-Signature.

The early-onset Alzheimer disease (EOAD) signature adapted from the study by Touroutoglou et al., 2023. Colored vertices on the cortical surface maps indicate areas where patients with EOAD showed greater cortical atrophy compared with low-amyloid cognitively normal (CN) participants within the Longitudinal Early-Onset Alzheimer's Disease Study (LEADS) cohort.

We also investigated the magnitude of cortical atrophy within a priori defined large-scale canonical functional networks, derived from a well-established parcellation of the cerebral cortex derived from 1,000 healthy young adults.²³ Based on the 17-network solution of this parcellation, bilateral networks were defined in accordance with our prior work: language, default mode, frontoparietal control, dorsal attention, salience, limbic, and somatomotor (eFigure 1).¹⁷ The somatomotor network was used as a control network, given its primary role in sensorimotor function rather than in cognitive-affective processes relevant for maintaining functional independence. These network labels, originally defined in the fsaverage template, were mapped to each participant's cortical surface to extract mean cortical thickness within each network. We further quantified standard morphometric measures, such as mean cortical thickness and average hippocampal volume, based on FreeSurfer's automatic segmentation algorithms and adjusted for the estimated total intracranial volume.

To account for individual variability in cortical thickness and volume measures, all values were converted to W-scores, which are analogous to Z-scores adjusted for covariates of no interest (age and sex in this study; see Touroutoglou et al.).²⁰ In brief, we first performed a multiple regression analysis using mean cortical thickness/volume data obtained from CN participants, which resulted in beta coefficient values for age and sex as well as individual values of residuals. Using these parameters, we then computed W-scores for all patients with EOAD with the following formula:

$$W_{ik}=\frac{T_{\text{ik}}-\widehat{T_{ik}}}{SDi}$$

Where $T_{ik}$ = the observed mean cortical thickness/volume of label i and patient k, $\widehat{T_{ik}}$ = the predicted mean cortical thickness/volume of label i and patient k given this patient's age and sex as well as the beta coefficients obtained from CN participants, and SD = the standard deviation of the individual residuals obtained from CN participants for label i.

To characterize the spatial distribution of cortical atrophy burden in our sample, we followed a similar approach as in our own prior work.²⁰ A whole-cortex vertex-wise 2-class general linear model implemented in FreeSurfer identified regions where individuals with MCI due to EOAD showed reduced cortical thickness relative to cognitively normal LEADS participants (p < 10⁻¹² with false discovery rate correction). The resulting vertex-wise map was binarized, and group-level cortical thickness differences (mm) were quantified within this mask. Maps were then thresholded at 0.2 mm, retaining regions showing ≥0.2-mm thinning in EOAD-MCI relative to controls.

Statistical Analysis

To test whether EOAD-signature atrophy predicted progression from MCI to dementia, we fit a univariable Cox regression model with EOAD-signature W-score as the predictor variable. The outcome variable was time from baseline to progression to dementia (event = 1) or last follow-up for censored participants (event = 0), measured in months. Additional univariable Cox models examined the association between progression risk and atrophy within each cognitive-affective network, as well as in the whole cortex and hippocampus. As a supplementary analysis, we assessed the predictive value of each EOAD-signature ROI using univariable Cox models. Model performance was compared using the Akaike Information Criterion (AIC), with lower values indicating better fit.

We further examined whether EOAD-signature atrophy burden added predictive information beyond baseline cognitive functional status. We built a base model, including age, sex, and baseline CDR-SB, and an extended model adding the EOAD-signature W-score to the base model. The incremental predictive value of the extended model was evaluated within a prognostic framework using complementary metrics. Predictive discrimination was assessed using Harrell concordance index (C-index), assessing the model's ability to correctly rank individuals by risk of progression. A likelihood ratio test was used to assess whether the extended model improved model performance relative to the base model, and relative model fit was compared using Akaike Information Criterion (AIC). These likelihood-based approaches have been advocated as statistically efficient measures of added predictive value in prognostic modeling.²⁴

The Cox proportional hazards models were performed using the survival package in R version 4.2.1. Before model fitting, all EOAD-signature and network-level predictor variables were inverted so that more positive W-scores would indicate thinner cortex (i.e., greater atrophy). We used the false discovery rate method to correct the p-values obtained from the univariable Cox regression models for multiple comparisons, assessing statistical significance at p < 0.05.

Sample size adequacy for the extended model was confirmed using the framework of Riley et al. for time-to-event prediction models, implemented using the pmsampsize R package.²⁵ Additional details are provided in eMethods. Internal validation of the extended model was performed using bootstrap resampling (1,000 iterations).²⁶ Optimism-corrected discrimination (Harrell C-index) and calibration (calibration slope, O/E ratio at 24 months, bootstrap-corrected plots) were derived from this procedure, along with the scaled Brier score for overall model performance.

All variables included in the final models—age, sex, baseline CDR-SB, and EOAD-signature/network W-score or MRI volumetric measures—were complete across all 130 participants. There were no missing data. This study was reported in accordance with the TRIPOD-AI guidelines for transparent reporting of multivariable prediction models.

Standard Protocol Approvals, Registrations, and Patient Consents

Institutional review board (IRB) approval was obtained through a central IRB overseen by Indiana University, and informed consent was obtained in written form from study participants or authorized representatives.

Data Availability

LEADS data can be requested from the LEADS Data Sharing Committee. Analysis code used for statistical modeling and internal validation is made publicly available in a GitHub repository (github.com/thiagoparanhos/EOAD-prognostication). This study is not preregistered; the parent LEADS cohort is described in the study by Apostolova et al. in 2021.²¹

Results

During the follow-up interval, 84 of 130 patients (∼65%) progressed from MCI to dementia; the remaining 46 patients had a cognitive functional status of MCI at the last follow-up (Table 1). Demographic and clinical characteristics of progressors and nonprogressors are presented on Table 2. MCI due to EOAD patients who progressed to dementia had greater cognitive impairment at baseline with higher CDR-SB scores (t = 5.87, p < 0.001) and lower MMSE scores (t = −4.65, p < 0.001). Progressors were also younger at baseline (t = −2.14, p = 0.036) and had a higher proportion of women (x² = 5.08, p = 0.024). The 2 groups did not differ in average years of education (t = −0.54, p = 0.51).

Table 2.

Demographic and Baseline Clinical Features of the MCI due to EOAD Patients Who Progressed or Not to Dementia

	Progressors (n = 84)	Nonprogressors (n = 46)	Group difference (p)
Sex (M/F)	43%	65%	0.024
Education	15.49 ± 2.33	15.76 ± 2.18	0.508
Age at scan	59 ± 3.75	60.65 ± 4.44	0.036
CDR-SB	3.03 ± 0.88	2.12 ± 0.82	<0.001
MMSE	23.05 ± 3.48	25.80 ± 3.09	<0.001

Abbreviations: CDR-SB = Clinical Dementia Rating scale Sum of Boxes; EOAD = early-onset Alzheimer disease; MCI = mild cognitive impairment; MMSE = Mini-Mental State Examination.

Values are presented as mean (SD).

Topography of Baseline Cortical Atrophy

The vertex-wise atrophy surface map revealed a pattern of atrophy affecting the medial and lateral parietal cortex and posterior lateral temporal cortex. Compared with the original EOAD-signature map, which included patients at MCI and mild and dementia stage, this MCI due to EOAD cohort showed relative sparing of the lateral prefrontal cortex, fusiform gyrus, and superior parietal lobule (Figure 2). For a direct comparison of this sample with the original EOAD-signature map, see eFigure 1 in this article and Figure 6 of the study by Touroutoglou et al.²⁰

Figure 2. Spatial Topography of Cortical Atrophy in MCI due to EOAD.

Colored vertices on the cortical surface maps indicate areas where MCI due to EOAD patients, on average, showed atrophy relative to CN participants from the LEADS cohort. The group difference (“gamma”) maps were thresholded at 0.2 mm and masked by areas showing statistically significant group differences (p < 10⁻¹², false discovery rate (FDR)–corrected). CN = cognitively normal; EOAD = early-onset Alzheimer disease; LEADS = Longitudinal Early-Onset Alzheimer's Disease Study; MCI = mild cognitive impairment.

Neuroanatomical Predictors of Progression to Dementia

The univariable Cox regression models indicated that age, but not sex or education, predicted the rate of progression from MCI to dementia (HR = 0.92, 95% CI 0.88–0.97, p < 0.0022). The Cox regression model using cortical atrophy in the EOAD-signature as predictor variable demonstrated that the magnitude of baseline cortical atrophy in the EOAD-signature was predictive of time to progression from MCI to dementia (x² = 20.17, HR = 1.24, 95% CI 1.13–1.37, p < 0.0022, AIC = 659.58). That is, for every 1 standard deviation greater magnitude of atrophy in the EOAD-signature, the patient's risk of progression from MCI to dementia was 1.24 times greater. Univariate survival curves for baseline atrophy within the EOAD-signature are shown in Figure 3.

Figure 3. Univariate Survival Curves Showing the Cumulative Proportion of Patients Who Have Progressed From MCI to Dementia as a Function of the Baseline Magnitude of Atrophy Within the EOAD-Signature.

MCI due to EOAD patients with relatively greater baseline atrophy in the EOAD-signature (i.e., cortical thickness that is more than 1 SD thinner than the group mean), represented as a dashed black line in the plot above, showed more rapid progression from MCI to dementia than those with an average magnitude of atrophy (i.e., cortical thickness within 1 SD of the mean) and 1 SD thicker than the group mean. MCI due to EOAD patients with relatively less baseline atrophy in the EOAD-signature (i.e., cortical thickness that is more than 1 SD thicker than the group mean), represented as a dashed gray line, showed a slower progression from MCI to dementia than those with an average magnitude of atrophy. EOAD = early-onset Alzheimer disease; MCI = mild cognitive impairment.

Baseline cortical thickness across the whole cortex (mean cortical thickness) and within each brain network, except the somatomotor network, predicted progression from MCI to dementia. By contrast, average hippocampal volume did not significantly predict progression (Table 3). Among all models, EOAD-signature atrophy yielded the lowest AIC value, indicating the best model fit (Table 3), followed by the frontoparietal control network. A supplementary analysis showed that the frontoparietal control network comprised approximately one-third of the EOAD-signature (eFigure 2). Regions within the frontoparietal control network that overlapped with the EOAD-signature provided a better model fit for predicting progression to dementia than nonoverlapping regions (AIC = 656.71 vs 664.64; eFigure 3). A univariate survival curve for baseline frontoparietal control network atrophy is shown in eFigure 4. All individual EOAD-signature ROIs predicted progression to dementia, with the right midlateral temporal cortex ROI showing the highest hazard ratio and lowest AIC (HR = 1.46, p < 0.001; AIC = 653.59) (eTable 1).

Table 3.

Univariable Cox Regression Models to Predict Time of Progression From MCI to Dementia in EOAD

Covariate	X2	HR	95% CI	p ≤	AIC
Age	10.54	0.92	0.88–0.97	<0.0026a	667.34
Sex (0 = F, 1 = M)	2.82	0.69	0.45–1.06	0.1076	—
Education	0.68	0.96	0.88–1.06	0.41	—
EOAD-signature	20.17	1.24	1.13–1.37	<0.0026a	659.78
Frontoparietal control	17.71	1.26	1.13–1.40	<0.0026a	661.82
Dorsal attention	12.64	1.22	1.09–1.36	<0.0026a	665.33
Default mode	13.21	1.29	1.12–1.48	<0.0026a	664.85
Language	7.73	1.28	1.07–1.52	0.0081a	670.08
Limbic	7.35	1.33	1.08–1.64	0.0101a	669.63
Salience	8.01	1.30	1.08–1.56	0.0081a	669.43
Somatomotor	2.05	1.15	0.95–1.4	0.1647	—
Whole brain thickness	10.66	1.31	1.11–1.54	<0.0026a	667.24
Hippocampal volume	3.64	1.21	0.99–1.48	0.0728	—

Abbreviations: AIC = Akaike Information Criterion; EOAD = early-onset Alzheimer Disease; FDR = false discovery rate; HR = hazard ratio; MCI = mild cognitive impairment.

The X², HR, 95% CI, and p values are displayed for the univariable cox regression models. The AIC values are provided for the significant models.

^aSignificant at FDR p < 0.05.

The multivariable extended model indicates that EOAD-signature atrophy significantly predicted progression to dementia beyond baseline cognitive functional status (HR = 1.14, p-value <0.009). In addition, although discriminability was similar between models (Base_C-index = 0.754; Extended_C-index = 0.757), the extended model demonstrated an improved fit according to the likelihood ratio test (χ²⁽¹⁾ = 6.5, p = 0.011) and a lower AIC (ΔAIC = −4.5), indicating superior predictive performance after accounting for model complexity.

Internal validation of the extended model demonstrated minimal overfitting. The optimism-corrected Harrell C-index for the extended model was 0.747. The apparent calibration slope was 1.00 (95% CI 0.73–1.27), and the optimism-corrected calibration slope was 0.86, indicating only mild overfitting. The observed-to-expected (O/E) ratio at 24 months was 1.18. The scaled Brier score was 31.7%, reflecting good overall predictive performance. Bootstrap-corrected calibration plots are presented in eFigure 5.

Discussion

In patients with EOAD, variability in symptom progression highlights the need for biologically valid biomarkers to prognosticate at the individual level, enabling personalized care management. The objective of this study was to evaluate the utility of our recently developed MRI biomarker—the EOAD-signature—for predicting clinical progression to dementia in individuals with EOAD at the MCI level of severity. Given the important roles of canonical large-scale functional networks in cognitive and affective functions supporting functional independence, we additionally compared the predictive value of atrophy within these networks. When patients with EOAD were scanned when their symptoms were at the MCI stage, atrophy in the EOAD-signature set of regions was the strongest predictor of progression to dementia, outperforming baseline atrophy in cortical functional networks. Importantly, EOAD-signature atrophy provided added predictive value beyond baseline cognitive functional status. Our analysis further showed the cortical atrophy in the frontoparietal control network (whose regions have substantial overlap with the EOAD-signature) to be the second best predictor of progression from MCI to dementia in EOAD (outperforming the predictive value of the other 5 networks). These findings suggest that the EOAD-signature MRI biomarker holds value as a tool for clinical prognostication, enabling more personalized approaches to management and care planning in EOAD.

Approximately 65% of participants in our sample progressed from MCI to dementia within the timeframe of follow-up (an average of 2 years), a progression rate that is notably higher than those reported in previous studies on early-onset MCI due to probable AD.^8,27 However, direct comparisons across studies are challenging because of differences in baseline cognitive impairment levels, as measured by the CDR-SB scores. For example, in the study by Tábuas-Pereira²⁷ et al. and Ye et al.,¹² baseline CDR-SB scores were 1.41 ± 0.9 and 1.4 ± 0.8, respectively, whereas our cohort exhibited a substantially higher baseline CDR-SB of 2.7 ± 0.96. This difference indicates greater baseline cognitive-functional impairment in our participants who likely contributed to the faster progression rate observed in our study. In addition, Tábuas-Pereira et al.²⁷ analyzed a relatively small cohort (n = 34) with confirmed AD biomarkers, while Ye et al.¹² did not incorporate biomarker evidence in their diagnostic criteria, increasing the risk of misclassification by including participants with non-AD conditions. Our findings also indicate that within the EOAD spectrum, older age was associated with a lower likelihood of progression to dementia. This finding aligns with prior literature suggesting that rates of cognitive decline decrease with increasing age of symptom onset in AD.^28,29 By contrast, sex and educational attainment were not significant predictors of progression in our cohort, in line with recent evidence.³⁰

At baseline, this sample of MCI due to EOAD patients exhibited prominent atrophy in medial and lateral parietal and posterior lateral temporal cortex (Figure 2). Using the same statistical threshold, we observed an overall milder severity of atrophy compared with the original EOAD-signature sample, which included both MCI and mild dementia cases.²⁰ The cortical regions relatively spared in the MCI due to EOAD sample were distributed in the lateral prefrontal cortex, fusiform gyrus, and superior parietal lobule (Figure 2 and eFigure 1). The EOAD-signature largely overlaps with temporoparietal regions that correlate with measures of global cognitive and functional status in early AD.³¹ Functional independence can be conceptualized in the ability to perform everyday activities across multiple domains without substantial external assistance.³² The complex nature of these activities is subserved by large-scale brain networks, which become dysfunctional as neurodegeneration progresses.³³ The EOAD-signature spans the boundaries of several of these networks,²⁰ including (1) the frontoparietal control network, associated with executive functions, encompassing regions in the caudal middle frontal gyrus, superior frontal gyrus, superior parietal cortex, and parts of the inferior parietal lobule, including the supramarginal gyrus³⁴; (2) the default mode network, which supports episodic and autobiographical memory and social cognition, among other functions, and includes key hubs such as the posterior cingulate cortex/precuneus, inferior parietal lobule, and lateral temporal regions³⁵; and (3) the language network in the left hemisphere, centered around the perisylvian cortex.³⁶ These networks support processes related to executive functions, episodic memory, and auditory-verbal working memory and comprehension, which are cognitive domains known to be predictive of clinical worsening in EOAD.⁸

Of these networks, baseline atrophy in the frontoparietal control network provided the best model fit to progression to dementia. The frontoparietal control network plays a vital role in supporting executive functions by enabling adaptive behavior in response to specific tasks and goals.^34,37,38 It also facilitates top-down regulation of activity across other brain regions to ensure coordinated processing. Impairments in executive functions are among the strongest predictors of functional independence and survival in AD.^39-41 Indeed, measures of baseline executive function have been shown to increase the accuracy of multimodal biomarker-based models that predict progression to AD dementia.⁴² Thus, although our EOAD cohort was clinically characterized as being mostly amnestic MCI at baseline (80%), baseline atrophy within the frontoparietal control network likely indicated the relatively imminent compromise of nonmnemonic functions, which better track with functional impairment. When the prediction models were compared, we found that baseline atrophy in the EOAD-signature showed maximal sensitivity (lowest AIC) in predicting progression to dementia in patients with EOAD at the MCI stage (Table 3), suggesting that the EOAD-signature may capture predictive variance beyond that afforded by executive functioning alone. In addition, the frontoparietal control network regions that overlap with the EOAD-signature provided a better model fit for dementia progression than frontoparietal control network regions without EOAD-signature overlap (eFigure 3), further supporting the specificity and usefulness of the EOAD-signature as a predictive biomarker in EOAD.

Our findings build on previous research indicating that atrophy primarily localized in the parietal and lateral temporal regions predicts cognitive decline in EOAD.^13,43 A key strength of our analysis is the use of an imaging biomarker specific to EOAD²⁰ that is validated against standard measures of clinical severity used in therapeutic trials. Moreover, we use an outcome measure (CDR) that is directly relevant to real-world clinical applications, including the assessment of functional independence, care escalation needs, and cost-related burden,⁴⁴ and that is the standard outcome measure in the field for assessing functional integrity in patients with neurodegenerative disorders.

Individual variability in clinical progression is an important factor to be considered in clinical trial design for patients with AD. The natural heterogeneity in clinical progression might create the false appearance of group differences (treatment vs placebo) when there are none.⁴⁵ This issue seems to be even more important for EOAD, in which cognitive decline is steeper and more variable,^6,45 making it harder to detect true potential treatment effects. Our work suggests that the baseline magnitude of atrophy in the EOAD-signature set of brain regions could be used in clinical trials to stratify the likelihood of progression to dementia across intervention (or placebo) groups, minimizing the problem of natural heterogeneity in disease progression in this population.

A limitation of this work is that the model was developed and tested within a single cohort, including predominantly non-Hispanic White participants, which precluded formal assessment of model fairness across racial and ethnic groups and limits conclusions regarding its generalizability to independent populations. In addition, our cohort mainly consisted of patients with amnestic single-domain or multidomain MCI. As such, it remains unclear to what extent baseline atrophy in the EOAD-signature would maintain predictive power over all EOAD phenotypic subtypes (e.g., those with predominantly behavioral/dysexecutive, visuo-spatial, or language phenotypes). Here, loss of functional independence is indexed by the CDR instrument. Future studies should also investigate whether EOAD-signature atrophy predicts convergent functional end points such as loss of advanced and instrumental activities of daily living, institutionalization, or mortality. In addition, incorporating neuropsychological assessment metrics as outcome variables would further allow assessment of whether EOAD-signature atrophy predicts decline in specific cognitive domains. Further translational research is needed to integrate computational approaches using structural brain MRI into clinical settings for prognostic purposes. For instance, patients with EOAD at the MCI stage who exhibit less atrophy burden within the EOAD-signature may have an extended period at the MCI stage, maintaining functional independence and potentially being able to benefit more from therapeutic interventions such as antiamyloid monoclonal antibody treatments. This study focused on MRI, but we also previously showed the importance of tau burden in early, atypical AD.⁴⁶ Future studies should consider using multimodal prediction models, including measures of cortical atrophy and plasma or PET biomarkers, to examine complementary and synergistic contributions to predict clinical decline in EOAD. Finally, although the clinical application of EOAD-signature atrophy quantification requires T1-weighted MRI processing and W-score estimation procedures which are not routinely implemented in clinical practice, automatic pipelines can be developed using open-source software such as FreeSurfer and institution-specific or open-source multisite normative data. In addition, the recent replication of the EOAD-signature in an Appalachian rural population further validates and supports its use in prognostication across distinct clinical cohorts of EOAD.⁴⁷

In conclusion, leveraging a relatively simple and cost-effective MRI-based measure, which is typically possible to obtain from scans collected in routine MRI protocols during diagnostic assessment, could facilitate the creation of individualized prognostic and monitoring strategies in clinical practice and in clinical trials. Future work should prioritize external validation in independent cohorts to establish generalizability and robustness across different populations and settings. After successful external validation, development of user-friendly implementation tools—such as web-based calculators or structured scoring sheets—should be considered to facilitate adoption in clinical practice and trial design.

Glossary

AIC

Akaike Information Criterion

CDR

Clinical Dementia Rating

CN

cognitively normal

EOAD

early-onset Alzheimer disease

LEADS

Longitudinal Early-Onset Alzheimer's Disease Study

MCI

mild cognitive impairment

ROI

regions of interest

Author Contributions

T. Paranhos: drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data. Y. Katsumi: analysis or interpretation of data. M.J. Brickhouse: analysis or interpretation of data. A. Eloyan: study concept or design; analysis or interpretation of data. R. Eckbo: analysis or interpretation of data. A. Zaitsev: analysis or interpretation of data. A Du: analysis or interpretation of data. R. La Joie: study concept or design. M. Thangarajah: major role in the acquisition of data. A. Taurone: major role in the acquisition of data. P. Vemuri: major role in the acquisition of data. C.R. Jack: major role in the acquisition of data. D.B. Hammers: major role in the acquisition of data. P.S. Aisen: major role in the acquisition of data. L.A. Beckett: major role in the acquisition of data. R. Koeppe: major role in the acquisition of data. W.A. Kukull: major role in the acquisition of data. A.W. Toga: major role in the acquisition of data. A. Atri: major role in the acquisition of data. D.G. Clark: major role in the acquisition of data. G.S. Day: major role in the acquisition of data. R. Duara: major role in the acquisition of data. N.R. Graff-Radford: major role in the acquisition of data. I.M. Grant: major role in the acquisition of data. L.S. Honig: major role in the acquisition of data. E. Johnson: major role in the acquisition of data D.T. Jones: major role in the acquisition of data. J.C. Masdeu: major role in the acquisition of data M.F. Mendez: major role in the acquisition of data. E.S. Musiek: major role in the acquisition of data. C.U. Onyike: major role in the acquisition of data. M. Riddle: major role in the acquisition of data. E. Rogalski: major role in the acquisition of data. S. Salloway: major role in the acquisition of data. S.J. Sha: major role in the acquisition of data. R.S. Turner: major role in the acquisition of data. T. Wingo: major role in the acquisition of data. D.A. Wolk: major role in the acquisition of data. K.B. Womack: major role in the acquisition of data. M.C. Carrillo:drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design. G.D. Rabinovici: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design. L.G. Apostolova: drafting/revision of the manuscript for content; major role in the acquisition of data; study concept or design. B.C. Dickerson: drafting/revision of the manuscript for content, including medical writing for content; study concept or design; major role in the acquisition of data, analysis or interpretation of data. M.C. Eldaief: drafting/revision of the manuscript for content, including medical writing for content; study concept or design; analysis or interpretation of data. A. Touroutoglou: drafting/revision of the manuscript for content, including medical writing for content; major role in the acquisition of data; study concept or design; analysis or interpretation of data.

Study Funding

This study is generously supported by R56 AG057195, U01AG6057195, U24AG021886, Alzheimer's Association LEADS GENETICS-19-639372, Alzheimer's Association LDRFP-21-818464, Alzheimer's Association LDRFP-21-824473, and Alzheimer's Association LDRFP-21-828356. NACC is funded by the National Institute on Aging (NIA; U24 AG072122). National Alzheimer's Coordinating Center (NACC) data are contributed by the following NIA-funded Alzheimer's Disease Research Centers (ADRCs): P30 AG010133, P30 AG062422, P30 AG066462, P30 AG066507, P30 AG062421, P30 AG066506, P30 AG072977, P30 AG066444, P30 AG066515, P30 AG062677, P30 AG072980, P30 AG072979, and P30 AG066511 as well as R01 DC014296, R01 AG085377, R21 AG080588, R21 DC019567, R21 AG073744, K23 DC016912, P50 AG005134, and P30 AG062421, and by the Tommy Rickles Chair in Primary Progressive Aphasia Research. This research was carried out in part at the Athinoula A. Martinos Center for Biomedical Imaging at the Massachusetts General Hospital (MGH), using resources provided by the Center for Functional Neuroimaging Technologies, P41EB015896, a P41 Biotechnology Resource Grant supported by the National Institute of Biomedical Imaging and Bioengineering (NIBIB), NIH. This work also involved the use of instrumentation supported by the NIH Shared Instrumentation Grant Program and/or High-End Instrumentation Grant Program; specifically, grant number(s) S10RR021110, S10RR023043, and S10RR023401.

Disclosure

The authors report no relevant disclosures. Go to Neurology.org/N for full disclosures.