Integration of neuroimaging signatures of aging and Alzheimer's with fluid biomarkers

Abstract

INTRODUCTION

Machine learning algorithms applied to neuroimaging may aid diagnosis and disease monitoring. This study aimed to derive neuroimaging signatures of brain aging (Spatial Pattern of Atrophy for Recognition–Brain Age [SPARE‐BA]) and Alzheimer's disease (SPARE‐AD) in a diverse sample and to examine their associations with plasma biomarkers, diagnosis, and disease progression.

METHODS

SPARE‐BA and SPARE‐AD were calculated using T1‐images from 1297 participants without dementia. Associations between neuroimaging signatures and plasma biomarkers were assessed with partial correlations. Logistic regression and cox proportional hazard models evaluated diagnostic accuracy and clinical progression.

RESULTS

SPARE‐AD displayed the strongest association with plasma phosphorylated tau 217, whereas SPARE‐BA correlated most strongly with plasma neurofilament light. The combination of neuroimaging signatures with plasma biomarkers achieved the highest diagnostic discriminability and strongest associations with clinical progression.

DISCUSSION

Neuroimaging signatures of aging and Alzheimer's disease display distinct associations with plasma biomarkers. Neuroimaging signatures combined with plasma biomarkers may improve the detection and prediction of cognitive impairment.

Highlights

• SPARE‐AD had the strongest partial correlation with plasma phosphorylated tau 217.

• SPARE‐BA demonstrated a higher partial correlation with plasma NfL.

• SPARE‐AD combined with plasma biomarkers had the highest diagnostic accuracy.

• SPARE‐BA combined with plasma biomarkers best predicted cognitive impairment.

1. BACKGROUND

The expanding landscape of treatments for neurodegenerative diseases places an impetus on identifying early vulnerability for cognitive decline and risk for future progression. ¹ Structural brain magnetic resonance imaging (MRI) is widely deployed in clinical settings to aid diagnosis of Alzheimer's disease and related dementias (ADRD); however, detection of subtle neurodegenerative changes remains a challenge, and predictions of individual clinical trajectories are limited. ² Cerebral atrophy is driven by the culmination of multiple neuropathological and age‐related processes, ³ making it highly variable and non‐specific across individuals. Machine‐learning algorithms can assist with decomposing this heterogeneity by capturing regionally specific morphological alterations associated with aging or a particular disease etiology. ³ , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹

Our team and others previously validated the Spatial Pattern of Atrophy for Recognition–Brain Age (SPARE‐BA) index, ³ , ⁶ , ¹² , ¹³ , ¹⁴ which assesses normative age‐related atrophy and can be used to evaluate how a given individual's brain morphology may deviate above or below the expected aging trajectory. The SPARE‐AD index was developed to evaluate atrophy patterns aligned with the stereotypical profile of Alzheimer's disease (AD). ¹⁵ Both SPARE‐BA and SPARE‐AD have been linked to cognitive decline and higher risk for clinical progression. ⁴ , ⁶ , ¹⁵ However, less is known about their sensitivity for detecting early cognitive changes and predicting progression among individuals with and without preclinical AD. ¹ SPARE‐BA has been shown to capture diffuse atrophy across the frontal lobe, inferior parietal cortex, superior temporal lobe, and insula, whereas SPARE‐AD is more localized to the entorhinal cortex, inferior temporal lobe, hippocampus, and amygdala. ⁶ The varied neuroanatomical patterns of SPARE‐BA and SPARE‐AD suggest they reflect distinct pathophysiological processes ⁶ ; however, further validation with other ADRD biomarkers is necessary.

Blood‐based biomarkers (BBMs) are rapidly transforming ADRD research and clinical care because they offer scalable, minimally invasive measures of disease‐related biology. In particular, phosphorylated tau 217 (p‐tau217), neurofilament light (NfL), and glial fibrillary acidic protein (GFAP) are among the best validated markers. ¹⁶ , ¹⁷ Plasma p‐tau217 is highly specific to cerebral amyloid beta (Aβ) and tau accumulation. ¹⁸ NfL, a marker of neuro‐axonal integrity, ¹⁹ and GFAP, a marker of reactive astrogliosis, ²⁰ , ²¹ dynamically change in response to acute and chronic neurological insults including, but not limited to, ADRD. Although these markers provide valuable information about molecular and cellular processes, they do not directly capture the spatial expression of neurodegeneration in the brain. It remains unknown whether MRI‐derived signatures such as SPARE‐BA and SPARE‐AD map onto distinct plasma biomarker profiles and whether they provide complementary diagnostic and prognostic information beyond plasma biomarkers alone.

In this study, we applied SPARE‐BA and SPARE‐AD in a large racially and ethnically diverse sample drawn from the Health and Aging Brain Study‐Health Disparities (HABS‐HD), ²² the South Texas Alzheimer's Disease Research Center (ADRC), ²³ and the Alzheimer's Disease Neuroimaging Initiative (ADNI). ²⁴ We sought to determine whether these MRI‐derived signatures showed distinct associations with plasma ADRD biomarkers, whether they discriminated between cognitively unimpaired (CU) and mild cognitive impairment (MCI) stages, and whether they improved prediction of clinical progression alone and in combination with plasma biomarkers in the overall sample and after stratification by Aβ positivity. We hypothesized that SPARE‐AD would align more strongly with plasma p‐tau217, whereas SPARE‐BA would align more strongly with markers of broader neurodegeneration and injury, and that combining MRI‐derived signatures with plasma biomarkers would improve detection of early cognitive impairment and prediction of clinical progression.

2. METHODS

2.1. Participants

Data for the current study were derived from the three aforementioned longitudinal cohort studies – HABS‐HD, the South Texas ADRC, and ADNI.

RESEARCH IN CONTEXT

1. Systematic review: A review of the literature was conducted using publicly available sources (e.g., PubMed). Machine‐learning‐derived neuroimaging signatures of aging and AD have been associated with cognitive decline and clinical conversion. However, less is known about their associations with plasma biomarkers and utility for detecting and predicting cognitive decline among ethnically and racially diverse individuals with and without preclinical AD.

2. Interpretation: Neuroimaging signatures of aging and AD displayed distinct associations with plasma biomarkers. The combination of neuroimaging signatures and plasma biomarkers optimized diagnostic accuracy and was most strongly associated with clinical progression.

3. Future directions: Expansion of this work with inclusion of a broader array of plasma biomarkers and neuroimaging sequences may further enhance early detection of cognitive impairment and disease monitoring for neurodegenerative diseases.

HABS‐HD is a longitudinal research study that conducts baseline and 24‐month longitudinal follow‐up visits consisting of a fasting blood draw, history and physical examination, cognitive assessments, questionnaires, brain MRI, and amyloid and tau positron emission tomography (PET) scans. ²² Participants include community‐dwelling non‐Hispanic White, Black/African American, and Hispanic adults recruited from the North Texas region. The protocol was approved by the University of North Texas Health Science Center Institutional Review Board (IRB), and all participants provided written informed consent prior to enrollment. HABS‐HD data undergo rigorous quality control and are released biannually through an online portal. This study utilized HABS‐HD data from the Wave 5 release. In addition, harmonized clinical diagnostic data were obtained from the Alzheimer's Disease Sequencing Project (ADSP). ²⁵

The South Texas ADRC is a longitudinal cohort study that enrolls participants at three sites in South Texas – University of Texas Health San Antonio, University of Texas Rio Grande Valley, and University of Texas Center at Laredo. ²³ All participants complete annual study visits consisting of the standardized National Alzheimer's Coordinating Center (NACC) Uniform Dataset, ²⁶ and a subset underwent brain MRI. Clinical diagnoses were adjudicated at formal consensus reviews following NACC guidelines. ²⁶ The protocol is approved at each participating site, and all participants provide written informed consent. The South Texas ADRC contributes data to NACC, where it is among the leading sites for Hispanic representation.

ADNI was launched in 2003 as public‐private partnership, led by Principal Investigator Michael W. Weiner, MD. ²⁴ The primary aim is to evaluate whether longitudinal clinical, neuroimaging, and other biological markers can predict clinical progression. Participants complete regular assessments, which include blood draws, physical and neurological examinations, cognitive testing, brain MRI, and amyloid and tau PET scans. These data are used by a neurologist or psychiatrist to assign a clinical diagnosis after each visit. ²⁷ The protocol is IRB approved at each participating site, and all participants provide written informed consent. Data used in this study were downloaded from adni.loni.usc.edu.

Across cohorts, this study applied the following inclusion criteria: (1) age 60 years or older, (2) diagnosis of CU or MCI at baseline, and (3) available T1‐structural brain MRI and plasma ADRD biomarker (p‐tau217, GFAP, NfL) data collected within 180 days of one another. Exclusion criteria include a diagnosis of dementia and a history of other neurological conditions such as stroke and Parkinson's disease.

2.2. Brain MRI processing and pattern classification

HABS‐HD, the South Texas ADRC, and ADNI performed T1‐weighted imaging on 3T MRI scanners using standardized protocols as previously described. ²² , ²⁸ , ²⁹ The T1‐weighted images were skull‐stripped, ³⁰ corrected for field bias, ³¹ and subjected segmentation using a fully automated multi‐atlas parcellation method to derive 145 regions of interest. ³² The SPARE‐BA and SPARE‐AD algorithms were then applied to the imaging data. SPARE‐BA and SPARE‐AD employed high dimensional support vector regressions to characterize the spatial pattern of volumetric changes associated with aging and AD, respectively ⁴ , ⁶ (Figure 1, Supplementary Material). The SPARE‐BA model was trained on 16,438 participants from 15 cohorts who were CU at baseline, spanning an age range of 18 to 102 years, with 60% female. The SPARE‐AD model was trained on 13,354 participants from 15 cohorts whose baseline diagnosis was either CU (N = 11,309) or dementia (N = 2045), with ages ranging from 60 to 102 years and a female proportion of 59%. The SPARE‐BA and SPARE‐AD models were tested on 18,815 participants and 21,915 participants, respectively, whose data were not used during model fitting or selection. Higher SPARE‐BA and SPARE‐AD scores indicate more advanced atrophy associated with aging and AD, respectively.

FIGURE 1.

Summary of relationships between chronological age and Spatial Pattern of Atrophy for Recognition of Brain Aging (SPARE‐BA) and Spatial Pattern of Atrophy for Recognition of Alzheimer's Disease (SPARE‐AD) scores in both the training and independent test datasets, together with brain regions contributing most strongly to the corresponding models. Training‐set scores were generated using five‐fold cross‐validation, such that the score for each participant was obtained from the model fitted without that participant's fold. Independent test‐set scores were generated for participants who were not used in model training. SHapley Additive exPlanations (SHAP) values were used to quantify the contribution of individual imaging features to model predictions relative to the expected model output. SHAP values were computed using the same model‐input feature order and preprocessing applied during prediction and were summarized by MUlti‐atlas region Segmentation utilizing Ensembles (MUSE) brain region to visualize regional contributions. (A) SPARE‐BA versus chronological age. SPARE‐BA predicted brain age showed a strong positive relationship with chronological age in both the training and independent test datasets. The solid blue line represents the fitted age‐related trend. The dashed lines indicate a ±5‐year range around the expected relationship. Observations above the upper dashed line represent relatively advanced brain aging, whereas observations below the lower dashed line represent relatively resilient brain aging. (B) SPARE‐AD versus chronological age. SPARE‐AD scores generally increased with chronological age in both the training and independent test datasets, while substantial interindividual variability was observed, particularly at older ages. The two dashed lines represent the lower and upper quadratic trend envelopes. These envelopes were calculated by fitting a shared monotonic second‐degree polynomial relationship between SPARE‐AD and chronological age and adding the 20th and 80th percentiles of the model residuals to the fitted trend, respectively. (C) SHAP‐based brain map for SPARE‐BA model. For SPARE‐BA, positive contributions increasing predicted brain age were strongest in the third ventricle, bilateral anterior insula, and right middle temporal gyrus, while negative contributions lowering predicted brain age were strongest in the right/left middle frontal gyrus, left supplementary motor cortex, medial superior frontal gyrus, and superior frontal regions. (D) SHAP‐based brain map for SPARE‐BA model. For SPARE‐AD, absolute SHAP values showed the greatest regional influence in the right and left inferior lateral ventricles, hippocampus, brain stem, amygdala, temporal pole, and opercular regions, indicating these regions contributed most strongly to the Alzheimer's disease dementia versus cognitively unimpaired decision score regardless of direction.

2.3. Aβ PET imaging

HABS‐HD conducted [18F]florbetaben (FBB) Aβ PET scans. ²² ADNI performed Aβ PET scans using one of three tracers: (1) FBB, (2) [18F]florbetapir (AV45), or (3) [18F]NAV4694 (NAV). Aβ PET image pre‐processing, tracer uptake quantification, and calculation of standardized uptake value ratios (SUVRs) were performed using standardized pipelines, as previously described. ³³ , ³⁴ For the current analyses, Aβ PET positivity or negativity data were downloaded from the HABS‐HD and ADNI database portals.

2.4. ADRD BBMs

In HABS‐HD, banked plasma specimens underwent assay of p‐tau217, NfL, and GFAP using Quanterix immunoassay kits on a Simoa HD‐X platform. P‐tau217 and GFAP were assayed using the latest Simoa Advantage PLUS Assay. ³⁵ In ADNI, p‐tau217 was quantified in plasma using Fujirebio chemiluminescence immunoassay reagents on the Lumipulse G1200 platform. ³⁶ Plasma levels of NfL and GFAP were assayed using the Lumipulse G1200 platform or the Quanterix duplex immunoassay kit on the Simoa HD‐X platform. In the South Texas ADRC, plasma p‐tau217, NfL, and GFAP were quantified using Quanterix immunoassay kits on the Simoa HD‐X platform. These assays were performed by the National Centralized Repository for Alzheimer's Disease and Related Dementias (NCRAD) Biomarker Assay Laboratory using standardized protocols. ³⁷

2.5. Apolipoprotein E genotyping

Across cohorts, apolipoprotein E (APOE) genotyping was conducted using polymerase chain reaction (PCR), as previously described. ³⁷ , ³⁸ , ³⁹ APOE ε4 carrier status was defined by the presence of one or more ε4 alleles.

2.6. Statistical analysis

Data normality was assessed using Q‐Q and kernel density plots. SPARE‐BA, SPARE‐AD, and ADRD BBMs were log‐transformed and then standardized to z‐scores by assay platform (for BBMs) based on Aβ‐negative CU participants. The demographic and clinical characteristics of the cross‐sectional and longitudinal samples were summarized using descriptive statistics. Associations between neuroimaging signatures and ADRD BBMs were evaluated using multivariate linear regression models adjusting for age, sex, APOE ε4, and cohort in the whole sample and with stratification by Aβ PET positivity. Partial correlations were also derived from Pearson's correlation models adjusting for the same covariates. To evaluate the ability of SPARE‐BA, SPARE‐AD, ADRD BBMs, and their combinations to discriminate between CU and MCI participants, area under the receiver operating characteristic curves (AUROC) were derived from logistic regression models conducted on the whole sample and then with stratification by Aβ PET positivity. Models were adjusted for age, sex, APOE ε4, and cohort. Model fit relative to the covariate‐only model was assessed using the Akaike information criterion (AIC). Least absolute shrinkage and selection operator (LASSO) methods with 10‐fold cross‐validation were applied to identify the combination of biomarkers that optimized diagnostic accuracy. Associations between SPARE‐BA, SPARE‐AD, ADRD BBMs, and their combinations with clinical progression were evaluated using separate Cox proportional hazard models with adjustment for age, sex, APOE ε4, and cohort. Time to event was days until clinical progression or last confirmed visit with the baseline diagnostic status. Model fit compared to the covariate‐only model was assessed using the C‐Statistic. LASSO with 10‐fold cross‐validation was applied to identify the combination of biomarkers that optimized predictions of clinical progression. The main analyses were performed on a combined sample of CU and MCI participants at baseline, and as a secondary analysis the sample was restricted to CU participants at baseline. In addition, analyses were repeated with stratification by race/ethnicity (Hispanic, Non‐Hispanic Black, Non‐Hispanic White). All analyses were validated using leave‐one‐subject‐out (LOSO) cross‐validation. ⁴⁰ All statistical tests were two‐sided, and the false discovery rate (FDR) was controlled at 5% using Benjamini‐Hochberg. ⁴¹ Statistical analyses were performed using Statsmodels version 0.14.5, Scikit‐learn version 1.7.1, and Lifelines version 0.30.3.

3. RESULTS

3.1. Participant characteristics

The current analyses included 1297 participants (mean age 71  ± 7 years, 42% female) from the HABS‐HD, South Texas ADRC, and ADNI studies (Table 1). Across cohorts, 27% (N = 354) of the sample identified as Hispanic, 21% (N = 272) as non‐Hispanic Black, and 47% (N = 609) as non‐Hispanic White. At baseline, 70% (N = 906) of the cohort was classified as CU, whereas the remainder met criteria for MCI (N = 391). Among the subset of 625 participants with Aβ PET imaging, 29% (N = 183) were Aβ positive, of which 52% (N = 96) were classified as CU and 48% (N = 87) as MCI at a baseline. A total of 394 participants (CU = 295, MCI = 99 at baseline) had longitudinal data available with a mean follow‐up duration of 3 ± 3 years. Of these, 154 participants (39%, CU = 107, MCI = 47 at baseline) had Aβ PET scans at baseline and 33% (N = 51, CU = 25, MCI = 26 at baseline) were Aβ positive. Across the follow‐up period, 16% (N = 46) of CU participants progressed to MCI and 17% (N = 17) with MCI converted to dementia. Of those who progressed, 22 participants (CU = 16, MCI = 6 at baseline) had Aβ PET scans at baseline with 54% (N = 12, CU = 7, MCI = 5 at baseline) classified as Aβ positive.

TABLE 1.

Participant demographics.

Characteristic	Whole sample N = 1297	Longitudinal sample N = 394
Age, years	71 ± 7	71 ± 6
Female, N (%)	757 (58%)	221 (56%)
CU / MCI, N (%)	906 (70%) / 391 (30%)	295 (75%) / 99 (25%)
Converters, N (%)	63 (5%)	63 (16%)
Race/Ethnicity, N (%)
Hispanic	354 (27%)	87 (22%)
Non‐Hispanic Black	272 (21%)	49 (12%)
Non‐Hispanic White	609 (47%)	220 (56%)
Other	62 (5%)	38 (10%)
Education, years	15 ± 4	15 ± 4
Presence of APOE ε4 allele, N (%)	353 (32%)	124 (32%)
Aβ PET scan, N (%)
Available	625 (48%)	154 (39%)
Positive	183 (29%)	51 (33%)
Cohort, N (%)
HABS‐HD	665 (51%)	243 (62%)
South Texas ADRC	186 (14%)	0 (0%)
ADNI	446 (35%)	151 (38%)

Note: Values represent mean ± standard deviation unless otherwise noted.

Abbreviations: Aβ, amyloid beta; ADNI, Alzheimer's Disease Neuroimaging Initiative; ADRC, Alzheimer's Disease Research Center; APOE, apolipoprotein E; CU, cognitively unimpaired; HABS‐HD, Health and Aging Brain Study–Health Disparities; MCI, mild cognitive impairment; PET, positron emission tomography; SPARE‐BA, Spatial Pattern of Atrophy for Recognition of Brain Age.

3.2. Associations of SPARE‐BA and SPARE‐AD with ADRD BBMs

In the adjusted linear regression models, SPARE‐BA and SPARE‐AD were positively associated with all three ADRD BBMs in the whole cohort and among Aβ PET‐positive participants (Table 2). In Aβ‐negative participants, SPARE‐BA and SPARE‐AD were only significantly associated with plasma NfL levels. In analyses stratified by race/ethnicity (Table S1), SPARE‐BA was consistently associated with higher NfL levels. SPARE‐AD was associated with higher NfL levels in non‐Hispanic Black adults and with all three BBMs in non‐Hispanic White adults.

TABLE 2.

Associations of neuroimaging signatures with plasma biomarkers.

Neuroimaging signature	P‐tau217 β (95% CI), FDR‐corrected p value	NfL β (95% CI), FDR‐corrected p value	GFAP β (95% CI), FDR‐corrected p value
Whole sample (N = 1297)
SPARE‐BA	0.092 (0.050 to 0.134), p < 0.001 *	0.115 (0.084 to 0.147), p < 0.001 *	0.045 (0.012 to 0.078), p = 0.007 *
SPARE‐AD	0.126 (0.097 to 0.156), p < 0.001 *	0.067 (0.044 to 0.089), p < 0.001 *	0.059 (0.036 to 0.082), p < 0.001 *
Aβ PET positive (N = 183)
SPARE‐BA	0.194 (0.070 to 0.319), p = 0.002 *	0.160 (0.071 to 0.249), p = 0.002 *	0.107 (0.019 to 0.195), p = 0.018 *
SPARE‐AD	0.240 (0.164 to 0.316), p < 0.001 *	0.085 (0.026 to 0.144), p = 0.006 *	0.092 (0.034 to 0.149), p = 0.004 *
Aβ PET Negative (N = 442)
SPARE‐BA	−0.014 (−0.070 to 0.042), p = 0.63	0.088 (0.041 to 0.136), p = 0.001 *	0.023 (−0.022 to 0.070), p = 0.39
SPARE‐AD	0.041 (−0.006 to 0.089), p = 0.13	0.073 (0.033 to 0.113), p = 0.001 *	0.038 (−0.001 to 0.077), p = 0.11

Note: β (95% CI) and FDR‐corrected p values derived from linear regression models examining the associations between neuroimaging signatures with plasma biomarkers adjusted for age, sex, apolipoprotein E ε4, and cohort in the whole sample and stratified by Aβ PET positivity

Abbreviations: CI, confidence interval; GFAP, glial fibrillary acidic protein; NfL, neurofilament light; PET, positron emission tomography; p‐tau217, phosphorylated tau 217; SPARE‐AD, Spatial Pattern of Atrophy for Recognition of Alzheimer's disease; Aβ, amyloid beta; SPARE‐BA, Spatial Pattern of Atrophy for Recognition of Brain Age.

^*FDR‐corrected p < 0.05.

The partial correlations between the neuroimaging signatures and ADRD BBMs in the whole sample were modest (r = 0.07 to 0.23) (Figure 2). The strongest partial correlation was observed between SPARE‐AD and p‐tau217. SPARE‐BA displayed a higher partial correlation with plasma NfL as compared to p‐tau217 and GFAP. Analyses stratified by Aβ positivity accentuated SPARE‐BA and SPARE‐AD's varied pattern of associations with ADRD BBMs. SPARE‐AD demonstrated a moderate partial correlation with p‐tau217. All other partial correlations were modest with SPARE‐BA most strongly associated with NfL. In Aβ‐negative participants, the strength of the associations was attenuated. Only the partial correlations between SPARE‐BA and SPARE‐AD with plasma NfL were statistically significant.

FIGURE 2.

Heatmap of partial correlations between neuroimaging signatures and standardized plasma biomarkers. Pearson's r derived from partial correlations adjusting for age, sex, apolipoprotein E ε4, and cohort in the whole sample are displayed in (A) and results stratified by Aβ PET positive and negative participants are displayed in (B) and (C), respectively. SPARE‐BA,  Spatial Pattern of Atrophy for Recognition of Brain Age; SPARE‐AD, Spatial Pattern of Atrophy for Recognition of Alzheimer's disease; p‐tau217, phosphorylated tau 217; NfL, neurofilament light; GFAP,  glial fibrillary acidic protein; Aβ,  amyloid beta, PET, positron emission tomography.

3.3. Diagnostic discriminability of neuroimaging signatures and ADRD BBMs

In the fully adjusted analyses, SPARE‐AD, but not SPARE‐BA, discriminated between CU and MCI groups at baseline (Table 3). Plasma p‐tau217, NfL, and GFAP also individually discriminated between diagnostic groups and displayed similar AUROCs with one another and SPARE‐AD alone. The combination of SPARE‐AD and all ADRD BBMs displayed the highest diagnostic accuracy and best model fit as characterized by AIC in the whole sample and across all racial/ethnic groups (Table S2).

TABLE 3.

Area under the receiver operating characteristics curve (AUROC) for cognitively unimpaired versus mild cognitive impairment.

Biomarker	Whole sample (N = 1297) AUROC (95% CI), FDR‐corrected p value [AIC]	Aβ PET positive (N = 183) AUROC (95% CI), FDR‐corrected p value [AIC]	Aβ PET negative (N = 442) AUROC (95% CI), FDR‐corrected p value [AIC]
Covariates only	0.622 (0.587 to 0.656) [1759.6]	0.692 (0.623 to 0.764) [246.6]	0.638 (0.585 to 0.694) [604.0]
SPARE‐BA	0.627 (0.590 to 0.659), p = 0.14, [1760.3]	0.719 (0.647 to 0.784), p = 0.20 [244.6]	0.642 (0.592 to 0.693), p = 0.65 [605.8]
SPARE‐AD	0.655 (0.618 to 0.689), p = 0.005 * [1708.9]	0.785 (0.713 to 0.846), p = 0.004 * [219.6]	0.649 (0.594 to 0.703), p = 0.46 [601.7]
P‐tau217	0.647 (0.612 to 0.680), p = 0.006 * [1721.8]	0.804 (0.736 to 0.865), p = 0.004 * [218.1]	0.641 (0.583 to 0.692), p = 0.65 [605.9]
NfL	0.644 (0.609 to 0.680), p = 0.013 * [1735.2]	0.745 (0.670 to 0.813), p = 0.036 * [238.8]	0.661 (0.607 to 0.714), p = 0.06 [599.6]
GFAP	0.638 (0.605 to 0.672), p = 0.006 * [1743.9]	0.741 (0.674 to 0.811), p = 0.036 * [238.8]	0.642 (0.589 to 0.695), p = 0.65. [605.5]
SPARE‐BA + ADRD BBMs	0.659 (0. 624 to 0.692), p = 0.005 * [1720.34]	0.812 (0.748 to 0.873), p = 0.004 * [223.1]	0.675 (0.621 to 0.724), p = 0.018 * [601.5]
SPARE‐AD + ADRD BBMs	0.670 (0.636 to 0.701), p = 0.005 * [1687.7]	0.826 (0.763 to 0.882), p = 0.004 * [211.8]	0.683 (0.631 to 0.736), p = 0.018 * [598.7]

Note: AUROC (95% CI) and FDR‐corrected p‐values derived from logistic regression models examining associations between neuroimaging signatures, plasma biomarkers, and their combinations with diagnostic stage adjusted for age, sex, apolipoprotein E ε4, and cohort in the whole sample and stratified by Aβ PET positivity. AIC compares model fit relative to covariates alone.

Abbreviations: Aβ, amyloid beta; ADRD BBM, Alzheimer's disease and related dementias blood‐based biomarker; AIC, Akaike information criterion; AUROC, area under the receiver operating characteristic curve; GFAP, glial fibrillary acidic protein; CI, confidence interval; NfL, neurofilament light; PET, positron emission tomography; p‐tau217, phosphorylated tau 217; SPARE‐AD, Spatial Pattern of Atrophy for Recognition of Alzheimer's disease; SPARE‐BA, Spatial Pattern of Atrophy for Recognition of Brain Age.

^*FDR‐corrected p < 0.05.

In analyses stratified by Aβ PET positivity (Table 3), SPARE‐AD and ADRD BBMs individually discriminated between CU and MCI participants. Plasma p‐tau217 was the strongest individual predictor of diagnostic group, but the combination of SPARE‐AD and ADRD BBMs obtained the highest diagnostic accuracy and best model fit. In Aβ‐negative participants, no individual neuroimaging signature or ADRD BBM discriminated between groups. The combination of SPARE‐AD and ADRD BBMs resulted in the highest diagnostic accuracy and strongest model fit. In analyses further stratified by race/ethnicity (Table S2), all models failed to achieve statistical significance. However, the combination of SPARE‐AD and ADRD BBMs similarly achieved the best model fit for Aβ PET‐positive participants across groups. For Aβ PET‐negative participants, the combination of SPARE‐AD and SPARE‐BA with ADRD BBMs yielded the best model fit across groups.

For the combined SPARE‐BA and ADRD BBM model in the whole sample, LASSO retained p‐tau217, NfL, and GFAP in the whole sample (Table S3). Among Aβ‐positive participants, SPARE‐BA, p‐tau217, and GFAP were retained, whereas in Aβ‐negative participants, only NfL and GFAP were retained. For the combined SPARE‐AD and ADRD BBM model, LASSO retained SPARE‐AD, p‐tau217, NfL, and GFAP in the whole sample and in analyses stratified by Aβ status.

3.4. Cox proportional hazards models for clinical conversion

SPARE‐BA, SPARE‐AD, and all ADRD BBMs were individually associated with clinical progression in the whole sample (Figure 3, Table 4) and across racial/ethnic groups in stratified analyses (Table S4). In the whole sample (Table 4), SPARE‐BA yielded the highest hazard ratio with a one standard deviation unit increase associating with a two‐fold increased risk of clinical conversion. The combination of SPARE‐BA and ADRD BBMs had the best model fit with only SPARE‐BA and p‐tau217 maintaining statistical significance. LASSO was optimized with inclusion of all biomarkers. The same findings were observed in Hispanic and non‐Hispanic White participants in stratified analyses and observations were limited to two cases in non‐Hispanic Black adults (Table S4).

FIGURE 3.

Cox proportional hazards models for clinical progression adjusting for age, sex, APOE ε4, and cohort. (A) Results for each individual biomarker. (B) Results for combined SPARE‐BA and ADRD BBM model. (C) Results for combined SPARE‐AD and ADRD BBM model. SPARE‐BA, Spatial Pattern of Atrophy for Recognition of Brain Age; SPARE‐AD, Spatial Pattern of Atrophy for Recognition of Alzheimer's disease; p‐tau217, phosphorylated tau 217; NfL, neurofilament light; GFAP, glial fibrillary acidic protein; ADRD BBM,  Alzheimer's disease and related dementias blood‐based biomarker.

TABLE 4.

Cox proportional hazard models for clinical progression.

	Whole sample (Cases/N = 63/394)			CU at baseline (Case/N = 46/295)
Biomarker	HR (95% CI), FDR‐corrected p value	C‐Statistic	LASSO mean N selected per fold	HR (95% CI), FDR‐corrected p value	C‐statistic	LASSO mean N selected per fold
Covariates only		0.682			0.718
SPARE‐BA	2.074 (1.428 to 3.012), p < 0.001 *	0.733		1.804 (1.156 to 2.813), p = 0.056	0.742
SPARE‐AD	1.472 (1.243 to 1.742), p < 0.001 *	0.743		1.265 (0.996 to 1.607), p = 0.12	0.738
P‐tau217	1.877 (1.436 to 2.456), p < 0.001 *	0.730		1.720 (1.222 to 2.421), p = 0.031 *	0.747
NfL	1.486 (1.124 to 1.966), p = 0.016 *	0.708		1.376 (0.953 to 1.988), p = 0.12	0.719
GFAP	1.630 (1.227 to 2.165), p = 0.003 * .	0.721		1.740 (1.200 to 2.523), p = 0.031 *	0.745
SPARE‐BA + ADRD BBMs		0.759	4		0.771	4
SPARE‐BA	1.692 (1.155 to 2.480), p = 0.016 *		1	1.722 (1.094 to 2.710), p = 0.085		1
P‐tau217	1.578 (1.140 to 2.183), p = 0.016 *		1	1.536 (1.035 to 2.278), p = 0.12		1
NfL	1.026 (0.724 to 1.453), p = 0.90		1	0.943 (0.610to1.458), p = 0.84		1
GFAP	1.120 (0.786 to 1.597), p = 0.67		1	1.415 (0.958 to 2.0981), p = 0.12		1
SPARE‐AD + ADRD BBMs		0.748	4		0.763	4
SPARE‐AD	1.295 (1.075 to 1.559), p = 0.016 *		1	1.205 (0.950 to 1.530), p = 0.16		1
P‐tau217	1.469 (1.047 to 2.062), p = 0.051		1	1.417(0.953 to 2.108), p = 0.12		1
NfL	1.044 (0.742 to 1.468), p = 0.90		1	1.043 (0.682 to 1.596), p = 0.84		1
GFAP	1.143 (0.799 to 1.636), p = 0.64		1	1.467 (0.982 to 2.192), p = 0.12		1

Note: Hazard ratios (95% CI) and FDR‐corrected p values derived from Cox proportional hazards models for clinical progression adjusted for age, sex, apolipoprotein ε4, and cohort in the whole sample and in CU participants at baseline. Model fit relative to the covariate‐only model was assessed using the C‐statistic. Selected frequency of each biomarker divided by the total number of trials derived by least absolute shrinkage and selection operator (LASSO) methods.

Abbreviations: ADRD BBM, Alzheimer's disease and related dementias blood‐based biomarker; CI, confidence interval; CU, cognitively unimpaired; GFAP, glial fibrillary acidic protein; HR, hazard ratio; LASSO, least absolute shrinkage and selection operator; NfL, neurofilament light; n, number; p‐tau217, phosphorylated tau 217; SPARE‐AD, Spatial Pattern of Atrophy for Recognition of Alzheimer's disease; SPARE‐BA, Spatial Pattern of Atrophy for Recognition of Brain Age.

^*FDR‐corrected p < 0.05.

In analyses restricted to CU participants at baseline (Table 4), only plasma p‐tau217 and GFAP individually associated with conversion to MCI. For both plasma p‐tau217 and GFAP, a one standard deviation unit increase was associated with a 1.7‐fold higher risk of progression to MCI. In analyses further stratified by race/ethnicity, SPARE‐BA, p‐tau217, and GFAP were significantly associated with increased risk of conversion to MCI across groups (Table S4). In the whole sample (Table 4), the combination of SPARE‐BA and ADRD BBMs evidenced the strongest model fit, and LASSO was optimized with the inclusion of SPARE‐BA and all plasma biomarkers.

4. DISCUSSION

This study extends previous work by deriving validated machine‐learning neuroimaging signatures of aging (SPARE‐BA ⁶ ) and AD (SPARE‐AD ¹⁵ ) in an ethnically and racially diverse sample composed of participants from clinical and population‐based longitudinal cohort studies. To better understand the overlapping and shared biological substrates captured by neuroimaging signatures, we evaluated their associations with plasma ADRD biomarkers. Relative to SPARE‐BA, SPARE‐AD displayed a stronger partial correlation with p‐tau217, highlighting its superiority for detecting atrophy due to AD pathological processes. The combination of SPARE‐AD and ADRD BBMs best discriminated between the CU and MCI diagnostic stages, whereas the combination of SPARE‐BA and ADRD BBMs most strongly associated with clinical progression. Importantly, these findings were replicated across racial/ethnic groups. In the whole sample, a one standard deviation unit increase in SPARE‐BA was individually associated with an approximate two‐fold increased risk of clinical conversion over the mean 3‐year follow‐up period. The findings have potential implications for optimizing stratification in clinical trials and future ADRD treatment decisions in asymptomatic at‐risk individuals. ⁴²

Biofluid markers of ADRD have yielded important insights into the pathophysiological processes underlying neurodegenerative diseases, ¹⁶ , ¹⁷ and markers such as p‐tau217 have demonstrated high specificity for AD pathology. ¹⁸ Neurodegeneration is not specific to a singular etiology, but machine‐learning algorithms, such as SPARE‐BA and SPARE‐AD, leverage the stereotypical progression of pathology in an effort to improve these discriminations. ³ , ⁴ , ⁵ , ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ , ¹¹ In our study, SPARE‐AD and plasma p‐tau217 demonstrated the strongest partial correlation, which was further accentuated in Aβ PET‐positive participants. The results highlight the improved capability of SPARE‐AD, relative to SPARE‐BA, for detecting atrophy patterns associated with AD pathology. Our results align with previous work demonstrating that an aggregate index of AD‐related brain atrophy most strongly associated with Alzheimer's Disease Neuropathologic Change (ADNC), particularly tau pathology, at autopsy. ²⁹ In our study, SPARE‐BA and SPARE‐AD demonstrated modest correlations with plasma GFAP and NfL in the whole cohort and in Aβ PET‐positive participants, suggesting they capture related but distinct aspects of aging and ADRD. Among Aβ‐negative participants, SPARE‐BA and SPARE‐AD were solely associated with plasma NfL, a marker of neuro‐axonal injury. ¹⁹ Our results are aligned with prior studies demonstrating associations between accelerated brain aging and circulating NfL levels. ⁴³ , ⁴⁴

SPARE‐AD, but not SPARE‐BA, discriminated between CU and MCI participants, further demonstrating its enhanced utility as a biomarker for ADRD. SPARE‐AD's diagnostic discriminability was stronger in analyses restricted to Aβ‐positive participants, suggesting that it advances with cognitive decline in early symptomatic AD. Plasma p‐tau217, NfL, and GFAP also discriminated between CU and MCI participants. Consistent with its specificity for AD pathology, ¹⁸ plasma p‐tau217 had the highest individual diagnostic accuracy among Aβ positive participants. However, the combination of SPARE‐AD and ADRD BBMs demonstrated the best model fit in the whole sample and in analyses stratified by Aβ positivity. The results indicate that elevations in p‐tau217, NfL, and GFAP, coupled with more advanced atrophy in regions vulnerable to AD, is associated with a greater likelihood of early symptomatic cognitive impairment, particularly among Aβ‐positive participants. This pattern of results was consistent across Hispanic, Non‐Hispanic Black, and non‐Hispanic White participants. It is important to note that the diagnostic discriminability of the combination of SPARE‐AD and ADRD BBMs was modest in the whole sample and only achieved strong discriminability in Aβ‐positive participants, suggesting the need for further optimization among those without preclinical AD.

No single neuroimaging or biofluid marker was able to discriminate between the CU and MCI stages in Aβ PET‐negative participants. These findings further highlight the challenge of discriminating between clinical stages when disease‐specific biomarkers are lacking. The combination of SPARE‐AD with ADRD BBMs resulted in the highest diagnostic accuracy; however, the overall model fit was acceptable to good, indicating that further improvements are necessary for clinical implementation. As cognitive decline is driven by heterogeneous processes, ⁴⁵ it is not surprising that inclusion of neuroimaging and fluid biomarkers improved the discriminability of cognitive impairment in adults without AD pathology.

In our study, SPARE‐AD was superior to SPARE‐BA for discriminating between CU and MCI participants cross‐sectionally. In contrast, SPARE‐BA was more strongly associated with longitudinal clinical progression than SPARE‐AD. It is possible that the diffuse atrophy pattern captured by SPARE‐BA may better predict risk for incident cognitive decline, whereas SPARE‐AD may be more sensitive for detecting early symptomatic cognitive impairment. Moreover, individuals with more extensive atrophy than expected for chronological age may have heightened vulnerability for developing a cognitive disorder in the future. ⁴⁶ Our results align with prior studies demonstrating that accelerated brain aging is associated with faster cognitive decline. ⁴ , ⁶ , ⁴⁴ , ⁴⁷ In addition, older brain age is observed not only in AD ⁴³ but also in other diseases such multiple sclerosis and major depression. ⁴⁸ , ⁴⁹ , ⁵⁰ In our sample, only 22 participants who converted had an Aβ PET scan at baseline, and of those, 54% (N = 12) were classified as Aβ positive. Therefore, it is likely that clinical progression in our sample was driven by multiple factors beyond AD. Plasma p‐tau217, NfL, and GFAP were also independently associated with increased risk of clinical progression. SPARE‐BA individually had the highest hazard ratios with a one standard deviation unit increase associated with a two‐fold increased risk of clinical progression. However, the combination of SPARE‐BA and ADRD BBMs had the best model fit and LASSO retained SPARE‐BA and all plasma biomarkers to optimize associations with clinical progression.

When the sample was restricted to CU participants at baseline, plasma p‐tau217 and GFAP were the only biomarkers that individually associated with progression to MCI. Our results align with prior studies demonstrating associations between plasma p‐tau217 and incident cognitive decline and clinical conversion. ⁵¹ , ⁵² , ⁵³ Even among CU participants, plasma p‐tau217 has been associated with increased Aβ and tau PET binding. ⁵⁴ , ⁵⁵ , ⁵⁶ , ⁵⁷ GFAP reflects astrocytic activation and increases with cerebral Aβ accumulation. ⁵⁸ CU adults with Aβ positivity demonstrate elevations in plasma GFAP, ⁵⁹ , ⁶⁰ and multiple studies have demonstrated that higher GFAP levels associate with cognitive decline and clinical conversion. ² , ²¹ , ⁴² , ⁶¹ , ⁶² , ⁶³ While p‐tau217 and GFAP were the strongest individual predictors, model fit was optimized with the combination of SPARE‐BA and ADRD BBMs, and LASSO retained SPARE‐BA and all plasma ADRD biomarkers. Our findings demonstrate that the prognostic utility of plasma ADRD biomarkers for cognitive impairment is further enhanced by the spatially bound information provided by SPARE‐BA.

This study has several strengths, including the large ethnically and racially diverse sample, the integration of multimodal biomarkers, and the inclusion of both cross‐sectional and longitudinal analyses. However, the limitations of the study also bear consideration. This study combined data across multiple cohort studies, resulting in data derived from multiple scanners, neuroimaging sequences, and ADRD assay kits and platforms. All measures were standardized to z‐scores in each cohort and by each assay platform for BBMs prior to analyses. Nonetheless, variances across protocols may contribute to heterogeneity in the results. While the overall sample size was large, the analyses stratified by race/ethnicity, particularly when used in combination with Aβ PET status, may have been underpowered and require further validation. In addition, kidney functioning and body mass index have been shown to influence ADRD biomarker levels, but, due to a lack of availability of these measures across cohorts, we were unable to adjust for these potential confounds. It is also important to note that data from a subset of participants in these analyses, combined with data from 12 other cohorts, were used to derive the SPARE‐BA and SPARE‐AD signatures, which could contribute to bias. Finally, Aβ PET data were only available on a subset of participants, reducing the sample size available for stratified analyses and limiting confirmation of AD pathology on all participants. Therefore, future studies with additional biomarker data, as well as the inclusion of multiple neuroimaging sequences, will be important for further evaluating early predictors of cognitive decline associated with specific etiologies.

In conclusion, SPARE‐AD and SPARE‐BA demonstrated distinct associations with ADRD BBMs, disease stages, and clinical progression, suggesting that they capture atrophy patterns associated with distinct pathophysiological processes. SPARE‐AD was more strongly correlated with plasma p‐tau217 than SPARE‐BA, indicating that it better assesses atrophy patterns associated with AD pathology. ¹⁸ The combination of SPARE‐AD and ADRD BBMs improved diagnostic accuracy for CU versus MCI participants, whereas the combination of SPARE‐BA and ADRD BBMs optimized associations with clinical progression. The integration of neuroimaging signatures and biofluid markers may improve detection and prediction of cognitive impairment, aiding decisions about clinical trial stratification and treatment initiation as therapeutics target earlier disease stages.

CONFLICT OF INTEREST STATEMENT

Mitzi M. Gonzales owns stock in AbbVie. Vanessa M. Young is the Communications Chair of the Technology and Dementia Professional Interest Area PIA and a member of the Sex Differences in Sleep and Alzheimer's Disease Work Group of the Alzheimer's Association. Jon B. Toledo has performed consulting work not related to this work with GLG Consulting, Spear Bio, Newell, and Acadia Pharmaceuticals. Jon B. Toledo is Clinical Trials Deputy Section Editor for the Alzheimer's & Dementia journal. Timothy J. Hohman has received consulting fees from Circular Genomics and serves on the scientific advisory board of and receives stock from Vivid Genomics outside of the current work. Timothy J. Hohman is also a Deputy Editor for Alzheimer's & Dementia: Translational Research & Clinical Interventions and a Senior Associate Editor for the Alzheimer's & Dementia journal. SS has a financial or non‐financial interest in Novo Nordisk outside of the current work. Gladys Maestre has served on the advisory board of and received stock from Fundaconciencia, LLC, and Appremend outside of the current work. Ngoc Huynh Ho, David H. Wang, Tanweer Rashid, Sachintha R. Brandigampala, Mengjin Dong, David Martinez, Sokratis Charisis, Christos Davatzikos, Tiffany F. Kautz, Ashley LaRoche, Kristine Yaffe, and Mohamad Habes report no disclosures. Author disclosures are available in the Supporting Information.

CONSENT STATEMENT

The HABS‐HD, ADNI, and South Texas ADRC study protocols were IRB approved at each participating site and conducted in accordance with the ethical standards of the Declaration of Helsinki. All participants or their legally authorized representatives provided written informed consent prior to enrollment.

Supporting information

Supporting Information: alz71889‐sup‐0001‐SuppMat.docx

Supporting Information: alz71889‐sup‐0002‐ICMJE.pdf