Clinical impact of the Alzheimer's Disease Neuroimaging Initiative: A review of studies using ADNI data (2023 to June 2025)

Abstract

Alzheimer's disease (AD) continues to pose a major public health challenge. Since its launch in 2004, the Alzheimer's Disease Neuroimaging Initiative (ADNI) has played a pivotal role in advancing the field by providing a comprehensive, open‐access, longitudinal dataset that integrates neuroimaging, biomarker, genetic, and clinical data related to AD. We used standard literature search methods to identify ∼1830 publications from 2023 to mid‐2025 that used ADNI data or samples. This review highlights key ADNI studies with direct clinical applications. We describe how these have impacted the development and validation of plasma biomarkers, improved clinical trials, assessed AD therapies, and developed methods for diagnosis and prediction using clinical, fluid, and imaging biomarkers. These contributions are underlain by an improved understanding of biological mechanisms of disease progression in AD and highlight ADNI's central role in advancing translational research and accelerating progress toward more effective, individualized care for patients with AD.

Highlights

• Review of ∼1830 recent Alzheimer's Disease Neuroimaging Initiative (ADNI) papers identified key studies with clinical impact.

• ADNI data advanced plasma biomarker development and clinical validation.

• ADNI data improved trial design and evaluation of Alzheimer's therapies.

• Multimodal biomarkers improved the diagnosis and prediction of disease progression.

1. INTRODUCTION

Alzheimer's disease (AD) is a progressive neurodegenerative condition with few available therapies, making early and accurate diagnosis vital for clinical management. The Alzheimer's Disease Neuroimaging Initiative (ADNI), launched in 2004, is a groundbreaking public‐private partnership established to develop and characterize biomarkers for use in clinical trials. ¹ Over two decades, ADNI has made substantial contributions to the field: its data supported clinical trials that led to the approval of lecanemab and donanemab; informed the development of the National Institute on Aging‐Alzheimer's Association (NIA‐AA) diagnostic criteria for AD; and promoted global collaboration through Worldwide ADNI. ² ADNI has influenced other major research initiatives, such as the Dominantly Inherited Alzheimer's Network (DIAN), ³ which adopted ADNI's imaging and biomarker protocols to directly compare autosomal dominant AD (ADAD) and late‐onset AD (LOAD) and formed the basis for standardizing data acquisition across the NIA's Alzheimer's Disease Research Centers program through the Standardized Centralized Alzheimer's and Related Dementias Neuroimaging (SCAN) project. ⁴ In 2024, ADNI marked its 20th anniversary with a special issue in Alzheimer's & Dementia, comprising 72 articles, including articles from across the Clinical, Magnetic Resonance Imaging (MRI), Positron Emission Tomography (PET), Biomarker, Genetics, Informatics, Biostatistics, and Engagement Cores, collectively reflecting how ADNI has shaped the field over two decades.

Central to ADNI's impact is its commitment to open science. ADNI data have been downloaded over 405 million times by more than 26,000 researchers across 169 countries, ⁵ resulting in over 7000 publications that used ADNI data and/or samples as the primary study resource or for external validation. Since 2012, ADNI's clinical contributions−spanning biomarker characterization and validation, outcome measure evaluation, and clinical trial design have been summarized in a series of approximately biannual reviews, ⁶ , ⁷ , ⁸ , ⁹ , ¹⁰ each providing an unbiased overview of ADNI papers from the broader scientific community, complementing ADNI Core‐focused reviews and overview narratives.

The present review continues this series. Due to the continued growth in the number and scope of ADNI publications, this iteration has been split into two complementary reviews: the present paper focuses on improvements to clinical trials and advances in diagnosis and prediction, while a companion review (Veitch et al., manuscript in preparation) covers publications enhancing our understanding of AD disease progression. From the beginning of 2023 until mid‐2025, ∼1830 ADNI studies were published, of which 403 had direct clinical applications. From these, we selected 101 of the most impactful studies, which we cover in the main text. This review describes ADNI's impact in the following key areas: (1) the development and validation of AD plasma biomarkers; (2) improvements to clinical trials, including the recruitment and selection of clinical trial participants and measurement of treatment effects; and (3) advances in clinical care, including assessments of current and emerging AD therapies, and development of methods for diagnosis and prediction. Together, these studies demonstrate the contributions of ADNI as an important data resource for advancing translational research in AD.

2. REVIEW METHODOLOGY

A total of ∼1830 ADNI publications were identified through searches of PubMed, Web of Science, and Google Scholar using the terms “ADNI” or “Alzheimer's Disease Neuroimaging Initiative.” These were reviewed by S.K. and D.V. and excluded if they were reviews, conference proceedings, or preprints, did not use ADNI data/samples, or had an electronic publication date outside the predefined search timeframe (01/01/2023–06/30/2025). Exceptions to the date criterion were limited to a small number of foundational methodological and biomarker studies that remain directly relevant to describing the selected ADNI studies in this review. In addition, several articles and reviews authored by ADNI Core leaders and included in the Alzheimer's & Dementia 2024 ADNI Special Issue are briefly discussed in the Introduction. Research articles reporting primarily clinical applications to advance biomarker validation and standardization, improve participant selection and outcome measures for clinical trials, or develop tools for diagnosis and prediction of AD applicable to real‐world care were included in this clinical review. Research articles primarily addressing mechanistic questions and improving our understanding of the underlying biology of AD were included in the companion research review (Veitch et al, manuscript in preparation). Using these criteria, 101 ADNI papers with a clinical focus published in journals with a 2024 Journal Impact Factor of 6 or higher were included in the main text. The selection process is summarized in Figure 1. Additional studies that may be of interest to readers are provided in Table S1.

FIGURE 1.

A flow diagram of the identification and selection of ADNI publications for this review. Approximately ∼1830 ADNI‐related records were identified through PubMed, Web of Science, and Google Scholar searches (January 1, 2023–June 30, 2025). After exclusion of reviews, preprints, and non‐peer‐reviewed articles and conference proceedings, 403 publications with a clinical focus were assessed for eligibility. Of these, 101 papers published in journals with a 2024 Journal Impact Factor > 6 were included in the main analysis. ADNI, Alzheimer's Disease Neuroimaging Initiative.

3. ADNI'S IMPACT ON THE DEVELOPMENT OF AD BLOOD TESTS

ADNI has been instrumental in developing key cerebrospinal fluid (CSF), mMRI, and PET biomarkers for AD. ¹⁰ However, these methods can be invasive, expensive, and not widely available for routine clinical use, while blood‐based biomarkers offer less invasive and more scalable alternatives. The recent FDA marketing clearance (May 2025) of the Lumipulse G p‐tau217/β‐amyloid 1‐42 Plasma Ratio test, designed to diagnose symptomatic patients aged 55 and above, marks the first United States Food and Drug Adminstration (FDA)‐approved blood test for AD diagnosis. ¹¹ Its approval represents the culmination of a decade‐long effort to develop and characterize ultrasensitive plasma assays for clinical use, a process in which the ADNI Biomarker Core has played a pivotal role. ¹² However, several challenges remain in the implementation of additional plasma biomarkers of amyloid‐β (A), tau (T), or neurodegeneration (N) for clinical use. These include the lack of large‐scale data on autopsy validation, the limited specificity of neurodegeneration markers like neurofilament light chain (NfL), and variability in assay techniques across studies. Here, we summarize recent ADNI studies on diagnostic blood tests comprising AT(N) plasma biomarkers. The utility of these biomarkers in enriching trial populations and as surrogate endpoints is discussed in Sections 4.1 and 4.2, respectively. Their diagnostic performance relative to CSF and neuroimaging AT(N) biomarkers is presented in Section 5.2.4. Exploratory blood‐based biomarkers, including plasma proteomics, lipidomics, miRNAs and epigenetic markers that are not yet in late‐stage clinical validation are covered in Section 5.2.3.

Plasma amyloid β (Aβ) 42/40 as a biomarker of Aβ was the first ultrasensitive AD assay described. ¹³ Although its ability to predict other AD hallmarks is limited, recent ADNI studies suggest plasma Aβ42/Aβ40 may be leveraged in different ways. First, it may rule out Aβ positivity in clinical trial prescreening. In the BioFINDER and ADNI cohorts, screening with plasma Aβ42/Aβ40 effectively ruled out Aβ pathology in low‐prevalence settings. However, due to the narrow dynamic range of Aβ42/Aβ40, its performance diminished in moderate‐prevalence environments where subtle differences in Aβ42/Aβ40 ratios become harder to detect, increasing the risk of misclassification. ¹⁴ Second, the well‐characterized discordance between plasma Aβ42/Aβ40, which measures soluble Aβ, and Aβ positron emission tomography (Aβ PET), which measures Aβ deposition, may hold diagnostic potential in early‐stage disease. Individuals who were plasma Aβ42/Aβ40 positive but Aβ PET negative (plasma+/PET–) showed faster Aβ accumulation in key brain regions, such as the superior parietal cortex (as measured by Aβ PET), compared with those who were negative by both measures (plasma–/PET–). ¹⁵ This suggests that plasma Aβ42/Aβ40 abnormalities may precede detectable changes on Aβ PET, positioning it as a potential marker of early Aβ pathology. Measurements of tau biomarkers, neurodegeneration biomarkers, and cognitive impairment in individuals stratified by plasma/PET positivity support this conclusion. Plasma+/PET+ individuals had a higher tau burden (as measured by tau‐PET) and poorer cognitive performance compared to plasma–/PET– individuals, indicating a more advanced stage of disease progression. ¹⁶

It is important to note that the diagnostic performance of the Aβ42/Aβ40 ratio is limited by the relatively modest difference in values (8%–15% across studies) observed between Aβ PET positive and Aβ PET negative individuals. Because the separation between groups is small, reliable application of this ratio requires exceptionally high analytical precision, minimal lot‐to‐lot reagent variability and ultrasensitive assays to ensure reproducible performance. These challenges are less prominent in plasma phospho‐tau biomarkers, which are indicative of multiple AD features. Since the initial reports of plasma p‐tau181 ¹⁷ and, more recently, p‐tau217, ¹⁸ as biomarkers of AD, there has been a surge of publications in this area, with ADNI making major contributions.

Although plasma p‐tau181 levels were shown to be a reliable biomarker for AD in a 2021 study of over 1000 ADNI participants, ¹⁹ plasma p‐tau217 holds even greater potential as a “jack of all trades” AD biomarker. It outperformed plasma p‐tau181, p‐tau231, glial fibrillary acidic protein (GFAP), and NfL in predicting Aβ positivity (CSF or PET) in a large, multicohort study of nearly 7000 individuals with CI across six countries (ADNI and Alzheimer's Disease Research Center [ADRC] cohorts). In this study, it achieved positive predictive values (PPVs) above 95% for ruling in Aβ pathology and negative predictive values (NPVs) between 90% and 99% for ruling out Aβ pathology in non‐AD dementia syndromes. ²⁰

Furthermore, plasma p‐tau217 was cross‐sectionally associated with regional cortical Aβ in preclinical ADNI and A4 participants ²¹ and with temporoparietal cortical tau deposition both cross‐sectionally and longitudinally. ²⁰ These studies extended previous work demonstrating its ability to outperform other plasma biomarkers in predicting tau PET positivity. ²² Similarly, the ratio of phosphorylated to non‐phosphorylated plasma p‐tau217 (p‐tau217R) predicted regional brain tau deposition in A+ early AD participants from the Clarity AD study and ADNI. ²³ Plasma p‐tau217 outperformed other plasma biomarkers (Aβ42/Aβ40, p‐tau181, GFAP, and NfL) for early diagnosis and for monitoring disease progression and distinguished AD patients from controls. It is frequently observed in AD and was associated with early pathological changes, which often precede dementia, including the onset of tau aggregation and mild cognitive impairment (MCI). ²¹

A systematic head‐to‐head evaluation of commercial AD blood tests from C2N Diagnostics, Fujirebio Diagnostics, ALZPath, Janssen, Roche Diagnostics, and Quanterix in ADNI participants found that plasma p‐tau217 consistently demonstrated the highest classification accuracy for Aβ PET status, tau‐PET status, cortical thickness status and cognitive impairment status across different platforms. ²² Plasma p‐tau217 levels were strongly correlated with Aβ PET, tau PET, and cortical thickness, outperforming other biomarkers such as plasma Aβ42/Aβ40 and NfL (Figure S1). The multifaceted prognostic and diagnostic abilities of plasma p‐tau217, particularly its ability to not only detect Aβ and tau positivity but to infer regional accumulation of these pathologies, may be of particular significance in the clinic. Additionally, use of plasma p‐tau217 may enable more efficient enrollment of preclinical participants in clinical trials, significantly reducing the need for confirmatory Aβ and tau PET scans. ²³

Implementing plasma p‐tau217 into widespread clinical practice requires population‐specific cutoffs, which can be difficult to determine due to the requirement for costly or invasive PET or CSF Aβ status as gold standards. Current uncertainty around the high rate of false positives reported with the Lumipulse assay may stem in part from quality control problems or cutoffs that do not appear to conform to amyloid PET in confirmatory studies. ²⁴ A novel modeling approach overcame this issue in ADNI participants, outperforming the FDA‐approved cut‐off in terms of NPV and PPV. ²⁵ This method may represent a cost‐effective solution to the cutoff determination issue and may be useful in low‐ and middle‐income countries where gold‐standard Aβ determination is inaccessible.

At the current time, several major limitations affect most studies referenced in the preceding paragraph and others performed in this field. First, most have been performed on cohorts of participants from academic medical centers, which differ demographically from the general population in race, ethnicity, education, and socioeconomic status, factors that may affect test performance. Second, in many studies, plasma samples were collected, frozen, and analyzed as a large batch. This reduces the problem caused by so‐called “lot‐to‐lot variation” of the antibodies or other components of the test kits, but contrasts with the clinical environment, where samples are analyzed on a day‐to‐day or week‐to‐week basis as blood is collected from the patients. Third, the reports compared plasma test performance with PET scans or CSF measures, which may or may not reflect how PET or CSF are performed and analyzed in the community. Fourth, test performance is generally much better in symptomatic impaired patients (MCI and dementia) because the prevalence of positivity is higher, and on average, the overall signal is stronger. Finally, different companies have developed tests with different instruments, antibodies, and methods (immunoassays vs. mass spectroscopy).

Ratios of biomarkers overcome issues of variability associated with preanalytical factors as well as physiological confounders and are therefore preferred over absolute measures. Several ADNI studies consistently reported superior performance of a p‐tau to Aβ42 ratio in predicting Aβ status across a variety of biomarker platforms. In the ALZAN cohort and ADNI, p‐tau217/Aβ42 outperformed both p‐tau181/Aβ42 and p‐tau217 alone in predicting CSF Aβ42 positivity (area under the curve [AUC] 0.93) ²⁶ (Figure 2). This ratio also outperformed other plasma biomarkers in predicting Aβ PET positivity ²² and was effective in both Black/African–American (AUC 0.88) and White (AUC 0.91) ADNI participants. ²⁷ Both the C2N PrecivityAD2 p‐tau217 assay and the Aβ42/Aβ40 combined model (AUC 0.93), as well as Fujirebio Lumipulse p‐tau217 and Aβ42/Aβ40 combined (AUC 0.91), accurately distinguish Aβ PET+ from Aβ PET– individuals, ²² supporting consistency across diagnostic platforms. However, it has been suggested that although using Aβ42 in the denominator reduces the size of the indeterminant zone, the greater variability of Aβ42 (due to diurnal variation, and possibly pre‐analytic handling issues) may lead to greater variability of the p‐tau217/Aβ42 ratio. Due to these issues and the others outlined above, considerable work remains to be done before the blood tests are ready for widespread clinical use.

FIGURE 2.

ROC curves of plasma biomarkers according to amyloid status in ALZAN and ADNI cohorts. ROC curves for Aβ+ detection in ALZAN (A–C) and ADNI (D–F) using Aβ40, Aβ42, and Aβ42/Aβ40 in combination with p‐tau181 (A) or p‐tau217 (B, D, E). The corresponding AUCs) with 95% confidence intervals are shown in (C) and (F). In ADNI, biomarkers were measured using either an immunoassay approach (Fuji) (D) or mass spectrometry (C2N) (E). Reproduced under open access from Lehmann et al. ²⁶ ADNI, Alzheimer's Disease Neuroimaging Initiative; Aβ, amyloid‐β; AUC, area under the curve; p‐tau, phosphorylated tau; s, areas under the curve; ROC, receiver operating characteristics.

AT(N) classification, ²⁸ originally intended for research, is now being applied to clinical settings, making a specific plasma biomarker of the third category, neurodegeneration, of great interest. Plasma NfL is an established biomarker of non‐AD specific neurodegeneration. However, its prognostic utility may be influenced by commonly comorbid cardiometabolic risk factors, now recognized as exacerbating disease progression via a variety of mechanisms (Veitch et al., manuscript in preparation). Higher plasma NfL levels were associated with faster cognitive decline (Alzheimer's Disease Assessment Scale–Cognitive Subscale; ADAS‐Cog), hippocampal atrophy, and increased white matter hyperintensity (WMH) volumes over 62.5 months in ADNI cognitively unimpaired (CU) and MCI participants. ²⁹ Aβ deposition, hypertension, and type 2 diabetes mellitus (T2DM) strengthened the association between plasma NfL and cognitive decline, while Aβ deposition, hypertension, and obesity strengthened the association between plasma NfL and hippocampal atrophy. These vascular and metabolic factors may therefore influence the prognostic utility of plasma NfL and should be integrated into neurodegeneration models.

Plasma N‐terminal tau (NT1‐tau) may serve as an alternative AD‐specific biomarker of neurodegeneration. ³⁰ In ADNI and a Chinese cohort, plasma NT1‐tau became prominently abnormal in symptomatic AD and was correlated with both Aβ plaques and Aβ‐dependent tau deposition. Moreover, it predicted faster Aβ accumulation and greater tau aggregation and was more highly correlated with brain atrophy and cortical thinning than NfL.

Given the superior performance of these plasma biomarkers, the ADNI Biomarker Core has incorporated plasma p‐tau217 into the ADNI‐4 study, alongside other key plasma biomarkers (p‐tau181, Aβ42/Aβ40, GFAP, and NfL). The team works closely with assay developers and industry partners, including Roche Diagnostics, Fujirebio, Quanterix, Janssen, and C2N Diagnostics, to validate and implement new assays. At the same time, they are conducting longitudinal studies to track how biomarkers change over time and how these changes relate to clinical outcomes, with the goal of improving their use in diagnosing, monitoring, and ultimately managing AD. ¹²

4. ADNI'S IMPACT ON CLINICAL TRIALS

FDA approval of the anti‐amyloid therapies, aducanumab, lecanemab, and donanemab, for clinical use is a testament to the contributions ADNI data have made in overcoming the myriad challenges facing AD clinical trials. These include high costs, lengthy timelines, identifying suitable participants, and the common presence of co‐pathologies. Variations in disease onset, progression, underlying biological mechanisms, and clinical presentations complicate the development of universal treatments and further hinder the design of AD clinical trials. Innovative, data‐driven methods can help overcome these barriers and substantially accelerate the development of effective treatments. ADNI, structured as a simulated clinical trial, offers a rich dataset for testing new approaches to address key criteria of clinical trial design, including recruitment, subject selection, surrogate markers, controls, endpoints, and outcome measures.

4.1. Selection of trial participants

Numerous strategies for selecting trial participants most likely to show cognitive decline during trial duration and to benefit from a given treatment have been described, all seeking a balance between cost and efficacy. Digital recruitment and screening strategies can lower costs, increase efficiency, and address the ongoing challenge of recruiting cohorts that are representative of the broader population. An unsupervised digital platform designed to engage a broad range of population groups recruited and assessed an ADNI online study cohort in which 42% of participants came from groups previously underrepresented in medical research. ³¹ However, issues with completion rates and engagement among these groups suggest the need for strategies to enhance digital literacy and accessibility.

The optimum selection strategy may vary depending on the trial population and mechanism of action of the disease‐modifying therapy being tested. Several novel approaches have been described for early AD trial populations. The application of machine learning (ML) techniques may uncover additional information from a range of modalities, including the bedrock of AD assessment, cognitive and functional assessments. Deep learning applied to trajectories of Mini‐Mental State Examination (MMSE), Clinical Dementia Rating–Sum of Boxes (CDR‐SB), and Functional Activities Questionnaire (FAQ) identified “slow” and “fast” progressing subgroups within ADNI early AD participants. ³² The model, validated in the National Alzheimer's Coordinating Center (NACC) ³³ cohort, predicted progression subgroups with an AUC of 0.70. Similarly, ML models developed in the placebo arm of the EXPEDITION3 trial and externally validated in ADNI combined demographics, Apolipoprotein E (APOE), structural MRI (sMRI) measures, and cognitive measures (ADAS‐Cog11, MMSE). These models significantly increased the PPV of detecting participants who demonstrated a clinically meaningful cognitive decline over the trial period. ³⁴ Both approaches were estimated to significantly lower trial sample sizes. In this participant group (A+ early AD), sample size reductions were also achieved by subject selection based on a combination of MRI and clinical data ³⁵ and a combination of plasma p‐tau181 positivity and high genetic risk beyond the APOE ε4 allele, operationalized using a polygenic hazard score (PHS). ³⁶ Stratification using a PHS alone identified a high‐risk group characterized by faster tau accumulation and cognitive decline and significantly reduced the required sample sizes and trial costs. ³⁷

In preclinical and prodromal AD subjects, similar combinations of modalities predicted progression. These included combinations of sMRI and plasma biomarkers (Aβ42/Aβ40 and p‐tau181) ³⁸ and of sMRI biomarkers, plasma biomarkers (p‐tau181 and NfL), and cognitive assessments (ADNI‐MEM composite and CDR‐SB) in A+ participants. ³⁹ In the latter study, the best models achieved AUC values of 0.78–0.93 for MCI to dementia progression and 0.65–0.73 for CU to MCI progression. ³⁹ The lower accuracy of prediction reflects the increased difficulty of this prognostic challenge earlier in disease progression, where heterogeneity of disease course is greater.

As tau accumulation depends on antecedent Aβ deposition and is closely related to cognitive decline (Veitch et al., manuscript in preparation), identification of participants likely to quickly accumulate tau over the trial period could reduce required sample sizes and costs. Tau PET outperformed sMRI, plasma, and CSF biomarkers (Aβ42/Aβ40, p‐tau217, NfL), and APOE status in predicting cognitive decline over 2 years in patients with amnestic MCI and mild dementia. ⁴⁰ Three distinct tau accumulation profiles−stable, moderate‐accumulator, and fast‐accumulator—were identified across the cognitive spectrum using longitudinal tau PET data from ADNI, Avid Pharmaceuticals, and the Harvard Aging Brain Study (HABS). ⁴¹ To reduce the cost of longitudinal tau PET scans in a trial population, moderate and fast tau accumulators were identified from baseline tau PET levels in combination with Aβ positivity and clinical variables. Pre‐screening for fast tau accumulation and Aβ positivity reduced sample size requirements by 46%–77% (Figure 3).

FIGURE 3.

Sample size reduction using subject selection with longitudinal tau PET. Required treatment effect and sample size estimates were analyzed for an 18‐month early AD trial with four enrichment schemes: (1) Aβ pathology with variable tau progression (black), (2) Aβ pathology and stable tau (green), (3) Aβ pathology with moderate tau accumulation (blue), and (4) Aβ pathology and rapid tau accumulation (red). The curves illustrate the necessary treatment effects and sample sizes to achieve 80% power for outcome measures such as ADAS‐Cog, CDR‐SB, PACC, and tau burden in the metaROI. Reproduced under open access from. ⁴¹ Aβ, amyloid‐β; AD, Alzheimer's disease; ADAS‐Cog, Alzheimer's Disease Assessment Scale–Cognitive Subscale; CDR‐SB, Clinical Dementia Rating–Sum of Boxes; meta‐ROI, meta–region of interest; PACC, Preclinical Alzheimer Cognitive Composite; PET, positron emission tomography.

However, any reductions in sample size achieved through tau PET and MRI scans must be balanced against the additional cost of the scans. Reducing the number needed to enroll a trial population can increase trial efficiency. A two‐step process first applied ML algorithms on demographic, clinical and genetic data from ADNI participants across the cognitive spectrum to select those to undergo MRI and Aβ PET screening. ⁴² A second step applied an ML model to imaging scans. This funnel approach predicted clinical progression (2‐year CDR‐SB) with an AUC of 0.84, substantially reduced the need for costly Aβ PET scans, and required sample sizes by 55%.

Individual variability in the rate of cognitive decline may introduce bias in the allocation of slow and fast decliners to placebo and treatment groups, which can distort estimates of treatment effects. An artificial intelligence (AI)‐based stratified randomization method reduced allocation bias in AD clinical trials by predicting cognitive decline using MRI and demographic data. ⁴³ Simulations using the ADNI dataset showed a 22% reduction in allocation bias and a 37% decrease in required sample size.

Mixed pathologies are more common than pure AD pathologies in older adults. Excluding participants with co‐pathologies is therefore critical to minimizing confounding factors in AD clinical trials. Common non‐AD neuropathologies include transactive response DNA‐binding protein 43 (TDP‐43) proteinopathy, Lewy body disease (LBD), and cerebral amyloid angiopathy (CAA), for which reliable biomarkers are lacking or require further validation. To circumvent this issue, a model using MRI signatures of these co‐pathologies was developed and validated in an autopsy cohort ⁴⁴ (Figure S2). When applied to ADNI CU and MCI participants, the model accounted for up to 25% of the variance in cognitive decline attributable to non‐AD causes. This approach helped to identify and exclude individuals with likely mixed pathologies from trials.

4.2. Measuring treatment effects

A major challenge in conducting clinical trials is measuring a treatment effect within the duration of the trial in the selected population. Recent ADNI studies have focused on improving traditional cognitive assessments and identifying surrogate biomarkers that reflect underlying AD biology, such as MRI, tau PET imaging and AT(N) plasma biomarkers. These innovations hold promise for streamlining trial protocols, reducing participant burden, and increasing sensitivity to detect treatment effects, particularly in early disease stages.

Traditional cognitive outcome measures such as the CDR‐SB and ADAS‐Cog have been considered the gold standard for AD cognitive assessment but are time‐consuming and resource‐intensive and may not be sensitive enough to detect early cognitive changes in primary prevention trials. Psychometric analysis of these tests can improve their sensitivity to detect a treatment effect. Individual items of the ADAS‐Cog varied in their reliability and measurement error. ⁴⁵ The five language items and the memory items, “word recall” and “delayed word recall,” showed the highest reliability and the lowest measurement error, effectively capturing 35%–45% of cognitive deficits and/or reliable fluctuations in cognitive performance from visit to visit (24%–37%) (Figure 4). The study suggests that variability linked to between‐subject trait differences may complicate the detection of treatment effects when relying on less robust ADAS‐Cog items in clinical trials.

FIGURE 4.

Visual representation of the variance in the components of the ADAS‐Cog evaluation at 24‐months. The variance sources are categorized as follows: consistency (including both the trait effect + cumulative effects), inconsistency (occasion‐specific effects), and measurement error. These sources collectively account for 100% of the total variance. The components assessed include: Q1, word recall; Q2, commands; Q3, constructional praxis; Q4, delayed word recall; Q5, naming; Q6, ideational praxis; Q7, orientation; Q8, word recognition; Q9, remembering test instructions; Q10, comprehension of spoken language; Q11, word finding difficulty; Q12, spoken language ability. Reproduced under open access from. ⁴⁵ ADAS‐Cog, Alzheimer's Disease Assessment Scale–Cognitive Subscale; Q1–Q12, individual ADAS‐Cog items.

The quest for a surrogate outcome measure that reflects underlying AD pathophysiology and circumvents the need for cognitive outcome measures has suggested that tau PET may be a promising candidate, given its strong association with neurodegeneration and cognitive decline. However, the selection of meaningful ROIs linked to cognitive decline and technical considerations, such as the choice of a reference region, can influence the power of tau PET to detect a treatment effect. The use of individualized regions of interest (ROIs) in the entorhinal cortex, hippocampus, and amygdala for measuring annual percentage changes in tau PET SUVR reduced sample size requirements by 15.9%–72.1% compared to group‐level ROIs in A+ participants across the cognitive spectrum. ⁴⁶ Ratio‐based tau PET biomarkers that were independent of a set reference region, discovered using an ML tool, reduced sample size estimates by up to 82% and errors by 65% over existing biomarkers. ⁴⁷ These improvements may help to shift the cost balance toward the use of tau PET as a surrogate endpoint, but the issue of scanner and tracer accessibility remains. MRI offers a more widespread neuroimaging alternative, crucial for tracking neurodegeneration and predicting future decline. The efficacy of MRI biomarkers in clinical trials was compared in ADNI MCI and mild AD participants. ⁴⁸ Ventricular and hippocampal volumes were the most reliable biomarkers for tracking AD progression over time. However, their clinical validity varied between MCI and dementia groups, suggesting that disease stage should be considered when using these biomarkers in clinical trials.

While tau PET, sMRI, and cognitive tests detect pathological changes occurring later in disease progression, plasma biomarkers can reflect earlier changes and therefore are of interest as the focus of disease‐modifying therapies shifts to primary prevention. Plasma biomarkers may serve as surrogate endpoints in clinical trials evaluating therapeutic efficacy in preclinical AD. Longitudinal changes in plasma p‐tau181 and NfL over a 24‐month period reduced the sample sizes required by up to 85% and 63%, respectively, compared to a 12‐month period in CU ADNI participants. ⁴⁹ Enrolling CU individuals with intermediate Aβ levels proved to be the most cost‐effective strategy, as it improved trial efficiency while allowing for the monitoring of treatment effects. However, while the estimated biomarker costs of a clinical trial were lower with these plasma biomarkers compared to neuroimaging surrogates, they necessitated a higher sample size with the greater associated costs of enrolling and monitoring participants, resulting in a higher overall estimated trial cost. These sorts of simulations with ADNI data help researchers to sift through the pros and cons of different outcome measures.

As neuroinflammation has been documented as playing an integral role in the pathological response to Aβ deposition (Veitch et al., manuscript in preparation), plasma biomarkers of this response may have utility in clinical trials. GFAP, a marker of astroglial activation, may be of use as a secondary endpoint in clinical trials, not only in A+ CU and cognitively impaired (CI) individuals, but also in A– CI individuals with a high vascular burden. ⁵⁰ In these participants from three cohorts, including ADNI, plasma GFAP increased over time and was associated with worsening CDR‐SB. Astroglial activation can also occur in response to cerebrovascular injury, making plasma GFAP a useful secondary endpoint in clinical trials of therapies targeting mechanisms beyond anti‐amyloid approaches.

It is important to note that although clinical trials may detect a treatment effect with these cognitive, neuroimaging, or plasma biomarker outcomes, these may not translate into a perceived meaningful difference for individuals with MCI or AD. Simulated clinical trials in ADNI participants estimated “time saved,” that is, the amount of additional time before progression to the next disease stage, for different disease‐modifying therapies’ efficacies. ⁵¹ These findings translate the oft‐cited minimal clinically important difference of 0.98 on the CDR‐SB outcome measure in individuals with MCI into a delay of 11 months over an 18‐month period. In comparison, the time saved for donanemab was approximately 7 months, suggesting that currently defined minimal clinically important differences may not align with patient hopes and expectations. This is consistent with a study that reported a hypothetical effect of a 20% slowing in CDR‐SB translated into only 10%–20% reductions in functional and behavioral changes over 5 years in A+ MCI and early AD individuals. ⁵² It is likely that a perceived clinical difference in prevention trials would be even more difficult to attain.

Beyond testing subject selection strategies and outcome measures, ADNI cognitive, functional, and imaging data were integrated in a longitudinal disease progression model to dynamically assess disease progression in mild to moderate AD. ⁵³ The model, validated using placebo data from Phase 2 clinical trials ABBY and BLAZE, supported clinical trial simulations and drug effect evaluations. It could enhance clinical trial design by simulating hypothetical drug effects, interpolating missing data, and assessing in‐sample information.

5. ADNI'S IMPACT ON CLINICAL CARE

ADNI studies have had multiple impacts on the development of approaches applicable to clinical care. With the approval of disease‐modifying therapies and the emergence of other therapeutic approaches, ADNI studies have assessed their effects, aiding clinicians in deciding whether these options are appropriate for their patients. The contribution of ADNI studies to understanding multifactorial factors to disease progression (Veitch et al., manuscript in preparation) has led to the development of novel methods for diagnosis or prediction that show promise in revolutionizing the efficacy of less invasive or lower‐cost approaches.

5.1. Assessments of current and emerging AD therapies

Recent ADNI studies have examined the effects of actual and hypothetical disease‐modifying treatments and therapeutic strategies on disease progression. Several studies leveraged ADNI imaging data and cognitive assessments to model the effects of anti‐amyloid therapies. ⁵⁴ , ⁵⁵ , ⁵⁶ These studies used a framework that incorporated biological mechanisms linking Aβ deposition to tau accumulation, MTL thickness, and finally changes in cognition (as measured by the CDR‐SB), ⁵⁴ to simulate the long‐term effects of anti‐amyloid therapies. Natural AD progression data from ADNI were used to validate biomarker trajectories and the relationship between biomarkers and clinical decline.

Model simulations closely matched outcomes from clinical trials, including the aducanumab EMERGE study and predicted a substantial reduction in Aβ plaques over time, indicating that longer treatment durations may yield greater clinical benefits. However, clinical benefits may not be equal in both sexes, with sex subgroup analysis of the Phase 3 CLARITY AD lecanemab trials suggesting greater effectiveness in men than women. As this trial was insufficiently powered for this subgroup analysis, it was unclear whether these results could be attributed to heterogeneity in the rate of cognitive decline or reflected pre‐existing sex differences. Trial simulations in ADNI participants using CLARITY AD inclusion criteria suggested that the reported sex differences were unlikely to be random and that lecanemab was likely less effective in women than men. ⁵⁵

Pressing issues facing the clinical use of lecanemab and donanemab are the critical side effects of amyloid‐related imaging abnormalities (ARIA) and space limitations in infusion facilities. ARIA has been linked to cerebral amyloid angiopathy (CAA), and the identification of patients with CAA pathology may enable prediction of ARIA risk. CAA, confirmed at autopsy, can be detected in only a subset of patients using MRI. ⁵⁷ A study of autopsy‐confirmed ADNI participants with MRI‐negative CAA reported that CSF Aβ42 outperformed other CSF AD biomarkers and Aβ PET in predicting CAA. This approach may offer a means of identifying high ARIA risk in these patients seeking anti‐amyloid antibody therapy. Limited infusion facility space, combined with a relatively narrow range of cognitive eligibility requirements MMSE and Clinical Dementia Rating‐Global Score (CDR‐GS), can mean that the window of opportunity for disease‐modifying therapy is missed. An analysis of ADNI and NACC longitudinal data found that participants with higher baseline CDR‐GS or MMSE had a lower risk of becoming ineligible for treatment within 12 months. ⁵⁶ The study estimated 25% of participants with a baseline CDR‐GS of 0.5 would become ineligible for treatment within a year compared to 50% of those with a CDR‐GS of 1. These results may guide the prioritization of patients at higher risk of ineligibility.

Despite FDA approval of anti‐amyloid therapies, their use remains limited due to the restricted appropriate use recommendations, cost barriers, potential side effects such as ARIA, and perceived lack of clinical effect. The frontline treatment for symptomatic patients is still cholinesterase inhibitors (ChEI); however, studies of the long‐term effects of ChEI therapy have reported inconsistent results. A comprehensive longitudinal analysis in ADNI MCI participants dichotomized by Aβ status revealed that ChEI therapy did not slow cognitive decline, and actually accelerated it, independently of Aβ positivity, and additionally accelerated the rate of MCI to dementia progression (Figure 5). ⁵⁸ In A‐ ADNI MCI participants, ChEI use was similarly associated with faster cognitive decline, and reduced baseline and longitudinal atrophy. ⁵⁹ The authors call for an urgent reevaluation of the use of ChEI in amnestic MCI.

FIGURE 5.

Effects of ChEIs on progression from MCI to dementia and cognitive trajectories with IPTW adjustment. Kaplan–Meier survival curves illustrate progression from MCI to dementia in the (A) unadjusted analysis, (B) following IPTW adjustment, with numbers at risk shown. (C) Longitudinal trajectories of cognitive measures including CDR‐SB score changes, (D) MMSE score changes, and (E) ADAS‐Cog13 score changes over time, adjusted using IPTW‐weighted LME models. Blue lines represent non‐users, whereas orange lines denote ChEI users; shaded areas represent 95% confidence intervals. All analyses were adjusted for age, gender, education, and APOE ε4 status using IPTW. β coefficients and p‐values reflect between‐group differences in rates of cognitive decline. Reproduced under open access from. ⁵⁸ ADAS‐Cog13, Alzheimer's Disease Assessment Scale–Cognitive Subscale (13‐item version); APOE ε4, apolipoprotein E ε4; CDR‐SB, Clinical Dementia Rating Sum of Boxes; ChEI, cholinesterase inhibitor; IPTW, inverse probability of treatment weighting; LME, linear mixed effects; MCI, mild cognitive impairment; MMSE, Mini‐Mental State Examination.

A growing interest in modifiable risk factors for AD suggests that controlling them may help slow disease progression. The breadth and depth of ADNI data were used in conjunction with data from the Hellenic Longitudinal Investigation of Aging and Diet (HELIAD) study to develop a system dynamics model describing the interactions among the multitude of neurobiological and psychosocial variables that contribute to sporadic AD. ⁶⁰ Simulated interventions on 15 modifiable risk factors identified sleep quality and depressive symptoms as having the strongest effects on cognitive decline (Figure 6). Sleep is critical for the clearance of waste products such as toxic Aβ from the brain via the glymphatic system, and depressive symptoms have also been associated with Aβ deposition, supporting the validity of these findings (Veitch et al., manuscript in preparation). Management of diabetes, hyperlipidemia, and hypertension, and more physical activity were modeled to have a lesser effect, ⁶⁰ but additional ADNI studies suggest that targeting these would also be beneficial. Participants with mild AD dementia taking a combination of diabetes, lipid‐lowering, and antihypertensive medications, and non‐steroidal anti‐inflammatory drugs, termed QuadRx, had significant delays in cognitive decline. In ADNI participants, QuadRx was associated with a 60% delay in decline in MMSE over 2 years, and in NACC participants, it was associated with a 47% delay over 10 years. ⁶¹ ADNI amnestic MCI (aMCI) participants served as a control group for the EXERT clinical trial, which tested the effects of moderate‐high versus low‐intensity exercise on 12‐month trajectories of cognition (ADAS‐Cog‐Exec). Both exercise groups showed significantly slower cognitive decline than the usual care ADNI group. ⁶² Despite the different approaches taken by these studies, they all point to the potential impact of treating modifiable risk factors on mitigating cognitive and functional decline and extending independent living.

FIGURE 6.

The effects of simulated interventions on 15 modifiable risk factors on ADAS‐Cog13 over 5 years. Simulations with and without interventions were performed using 1000 posterior samples and are shown with the addition or subtraction of one standard deviation from the baseline for each risk factor. The dotted line indicates the absence of an effect. Reproduced under open access from. ⁶⁰ ADAS‐Cog13, Alzheimer's Disease Assessment Scale–Cognitive Subscale (13‐item version).

5.2. ADNI's contributions to diagnosis and prediction

AD progression is highly heterogeneous due to multiple contributing factors that modulate the stereotypical cascade of events (Veitch et al., manuscript in preparation). Prediction of future decline remains a greater challenge than diagnosis of clinical stages, particularly at the MCI stage, where there is significant regression to CU and reprogression back again; reversion, stability, and progression in MCI participants are associated with distinctly different prognoses. ⁶³ , ⁶⁴

Accurate and robust prediction models are essential for enabling personalized care and improving the design and efficacy of clinical trials. With the accumulation of longitudinal data across multiple modalities from participants across the cognitive spectrum, recent studies have continued to use ADNI as a cohort for developing or validating a wide range of diagnostic and prognostic approaches. Notably, the application of ML frameworks that enable models to consider complex relationships in the data has resulted in significant gains in prognostic accuracy. These AI approaches are particularly powerful when applied to high‐dimensional data, such as imaging or ‘omics data. The ADNI dataset has also enabled the systematic comparison of different ML approaches. ⁶⁵

5.2.1. Clinical variables

Clinical variables, including cognitive and functional tests, APOE genotyping, demographics, and routine assessments such as cardiovascular risk and depression screening, are the most accessible and cost‐effective diagnostic and prognostic tools available. The Everyday Cognition (ECog‐12) scale, reported by both the participant and their study partner, measures subjective changes in cognition and instrumental activities of daily living compared to 10 years before. Study partner‐reported ECog‐12 scores distinguished ADNI CU from CI participants (CI; MCI and AD) more effectively than self‐reported scores (AUC 0.78 vs. 0.70). ⁶⁶ As self‐awareness of cognitive and functional deficits decreases with worsening disease stage, the disparity between self‐ and study partner‐reported scores may provide the basis for a low‐cost and scalable screening method. A model developed in ADNI participant–study partner dyads and validated in an external cohort included ECog discordance metrics, demographics, and a depression score. This achieved high accuracy (AUC 0.87) and specificity (0.97) in distinguishing CU from MCI participants, suggesting it is a practical method for ruling out MCI in clinical settings. ⁶⁷

Psychometric analysis of standard cognitive tests can improve their prognostic ability. Process scoring and latent modeling of a word list recall test from ADAS‐Cog yielded a cognitive biomarker that predicted CDR‐SB scores at 36 months substantially better than traditional ADAS‐Cog metrics. Its high NPV indicated it could be useful for screening out those unlikely to experience cognitive decline. ⁶⁸ Statistical modeling of CDR‐SB scores in Australian Imaging, Biomarkers and Lifestyle Study of Ageing (AIBL) participants led to the development of a simple, cognition‐based method for predicting age at onset of MCI or AD, the Florey Dementia Index. This tool, validated in ADNI, could aid clinicians in managing patients or prioritizing them for disease‐modifying therapies. ⁶⁹

Sporadic early onset AD (EOAD; onset before 65 years) may be cognitively distinguishable from LOAD (onset after 65 years), despite sharing a similar etiology. Instead of the typical early memory deficits observed in ADNI participants, participants from the Longitudinal Early‐Onset Alzheimer's Disease Study (LEADS) showed greater impairments in visuospatial skills, executive functioning, and processing speed/attention. ⁷⁰ Non‐amnestic deficits in younger patients may therefore be an important diagnostic feature of EOAD.

Application of ML approaches to clinical data yielded further improvements. In ADNI participants dichotomized by CSF p‐tau181/Aβ42 into AD‐like CSF+ and control CSF–, ML applied to standard clinical data generated a model comprising demographic and neuropsychological variables. ⁷¹ This predicted diagnostic conversion over 3 years with an accuracy comparable to CSF AD biomarkers. ⁷¹ Important predictors were sex, Rey Auditory Verbal Learning Test (RAVLT) percent forgetting, FAQ, and ADAS‐Cog13. Beyond solely neuropsychological tests lies a wealth of readily available clinical data reflecting other contributors to disease progression that may be exploited for diagnostic purposes. ML algorithms applied to 21 clinical variables from CU and AD participants from ADNI and AIBL diagnosed AD with AUCs up to 0.75, driven predominantly by APOE ε4 genotype, aided by measures such as urea nitrogen, hemoglobin, and platelets. ⁷² The accessibility of APOE genotyping supports its incorporation into routine diagnostic clinical use. Consideration of additional APOE genotypes, particularly ε2, may improve diagnostic accuracy. ⁷³ While applying psychometric and ML approaches to clinical data has enabled improvements in predicting disease status and future decline to the point where some models rival the more invasive CSF biomarkers, these data alone are insufficient for some applications. For example, cognitive measures added no additional utility beyond demographic and genetic measures in predicting Aβ status in three cohorts. ⁷⁴

5.2.2. Imaging

Aβ deposition occurs early in disease progression before the onset of cognitive impairment and may have diagnostic utility in early‐stage disease. A deep learning model that selected radiomics features from Aβ PET scans from ADNI CU and MCI participants and the EudraCT 2015‐001184‐39 trial ⁷⁵ accurately diagnosed MCI (AUC of 90%) and significantly improved diagnostic performance compared to traditional standardized uptalke value ratio (SUVR) measures. However, not all participants with Aβ pathology will decline. The spatiotemporal relationship between Aβ and tau deposition and the deleterious effects of their concomitant presence (Veitch et al., manuscript in preparation) are not only important for understanding disease progression but also for predicting decline and selecting individuals for anti‐amyloid therapies. An ML model trained on ADNI imaging and cognitive data and validated in HABS reported that Aβ binding in frontal and striatal regions was predictive of a variety of measures of progression: entorhinal tau binding, hippocampal atrophy, and decline in Preclinical Alzheimer's Cognitive Composite (PACC) scores. ⁷⁶ Conversely, discordance between Aβ status measured by Aβ PET and CSF Aβ42, occurring in around 10% of ADNI MCI participants, was associated with a lack of progression; the cognitive trajectories of ADNI participants who were either CSF+/PET– or CSF–/PET+ in their Aβ status did not differ significantly from participants with no evidence of Aβ pathology (CSF–/PET–). ⁷⁷ These individuals may have other pathologies, and their eligibility for anti‐amyloid treatment should be confirmed with additional testing. The widespread implementation of AD plasma assays may supplant CSF assays, and additional studies are required to determine whether a similar issue exists with respect to Aβ PET.

Tau‐PET imaging holds considerable promise for diagnosing AD and assessing treatment effects. One hurdle to its use in multicohort studies is the difficulty in comparing results from different tracers. Much like the development of the Centiloid system for Aβ PET, ⁷⁸ ADNI data contributed to the construction of a universal scale, CenTauR, from the standardization of quantitative tau PET across six tracers. ⁷⁹ The CenTauR scale accurately distinguished A+ AD from A– CU individuals, and T‐ individuals from those with tau PET scans indicative of limbic predominant, hippocampal sparing, and typical AD subtypes.

Fluorodeoxyglucose PET (FDG‐PET) imaging measures impaired glucose metabolism in the brain, reflecting early neurodegeneration. Distinct patterns of hypometabolism are associated with antecedent tau deposition and with dementias of other etiologies. As tau PET is not currently widely accessible for use as a biomarker for AD diagnosis and as a treatment outcome measure, the relationship between tau deposition and hypometabolism may be used to impute regional tau accumulation from FDG‐PET scans. ML techniques applied to ADNI FDG‐PET data outperformed models based on Aβ PET and MRI for imputing tau‐PET images. ⁸⁰ FDG‐PET may also be a useful diagnostic tool in differentiating AD from related dementias. Analysis of antemortem FDG‐PET from patients with autopsy‐confirmed AD or limbic age‐related TDP‐43 encephalopathy neuropathologic change (LATE‐NC) revealed distinct patterns of hypometabolism ⁸¹ (Figure 7). A temporolimbic pattern of hypometabolism was associated with LATE‐NC, whereas AD was characterized by temporoparietal hypometabolism. Beyond differentiating between AD and LATE‐NC, FDG‐PET images form the basis of a clinical decision support system, StateViewer, that differentiates between nine neurodegenerative syndromes. ⁸² StateViewer uses an ML framework to detect patterns of atrophy associated with AD and related disorders and tripled the odds of having a diagnosis concordant with pathology.

FIGURE 7.

FDG‐PET imaging of impaired glucose metabolism in AD, LATE‐NC and mixed patient groups. The average Z‐score maps reflect hypometabolism in clinically diagnosed AD dementia patients, categorized as AD‐like (A), clearly LATE‐NC–like (B), or mixed (C) by an automated pattern‐matching algorithm. Average Z‐scores represent Glass’ Δ effect size. Reproduced under open access from. ⁸¹ AD, Alzheimer's disease; FDG‐PET, fluorodeoxyglucose positron emission tomography; LATE‐NC, limbic age‐related TDP‐43 encephalopathy neuropathologic change; Δ, Glass’ delta (standardized effect si.

Progressive changes in brain functional connectivity, as measured by resting‐state functional MRI (rs‐fMRI), follow glucose hypometabolism and precede atrophy during AD progression and may be effective in predicting MCI to dementia conversion. However, the heterogeneity of functional disconnection makes the selection of stable and objective functional connectivity biomarkers a challenging task. Differences in functional connectivity patterns in CU, MCI, and AD individuals from three large cohorts, including ADNI, identified key brain networks as significant differentiators between CU and MCI/AD participants. ⁸³ Regions, including the sensorimotor, visual, and default mode networks, mediated the effects of Aβ deposition and brain glucose metabolism on cognitive decline and strongly predicted conversion from MCI to dementia. Individual‐specific functional connectivity may further enhance diagnostic and prognostic capability in both APOE ε4 carriers and non‐carriers. ⁸⁴

Stereotypical patterns of atrophy, measured by sMRI, are the final neurodegenerative step and are most closely linked to cognitive decline and diagnostic progression. sMRI has therefore long been employed as a powerful diagnostic and prognostic tool. ML techniques offer a powerful way to select the most discriminative regions for a particular diagnostic challenge. Still, the high dimensionality of sMRI data requires large datasets to overcome statistical bias, which can limit model generalizability and robustness. ADNI, alone and as a part of large neuroimaging consortia such as ENIGMA and iSTAGING, has provided a wealth of sMRI data for testing ML models. Some of these now achieve impressive diagnostic accuracy in detecting AD; Mmadumbu et al. reported an accuracy of over 96% in distinguishing ADNI AD from CU participants. ⁸⁵ Accuracies may differ between males and females, ⁸⁶ possibly reflecting sex differences in disease progression, and so sex differences should be considered when applying diagnostic algorithms to the wider population.

The greater challenges of predicting CU to MCI progression, ⁸⁷ , ⁸⁸ MCI to CU reversion, ⁸⁷ and MCI to dementia progression have also benefited from the application of ML methods to ADNI sMRI data to the point where reasonably accurate (AUCs in the range of 0.70–0.90) predictions are possible up to 7 years. ⁸⁹

An emerging use of sMRI data is for the estimation of “brain age gap.” This is defined as the difference between predicted brain age (the age of an individual's brain structure and function) and chronological age. Both the structurally and neuropsychologically defined brain age gaps track brain aging and predict AD progression. ADNI AD participants had a brain age gap estimated from sMRI data of +5.9 years compared to controls. ⁹⁰ The brain age gap was further validated in four independent cohorts, including ADNI, ⁹¹ and was associated with AD biomarkers, including elevated Aβ, advanced AT stages, the APOE ε4 allele, and increased plasma NfL levels. While this study did not report prognostic capabilities of the brain age gap, its close relationship with a variety of AD characteristics makes it a prime candidate for future studies in this arena. Indeed, a model developed using neuropsychological tests to estimate the brain age gap distinguished between stable MCI and progressive MCI, ⁹² and sMRI may be an even more powerful tool. For its widespread use, however, robust normative samples are required. Generative models applied to iSTAGING data were used to produce GenMIND, a collection of 18,000 synthetic normative samples across the lifespan. ⁹³ These synthetic data effectively augment real‐world applications like brain age gap estimation.

ADNI sMRI data have also been used in studies aimed at detecting distinct patterns of atrophy associated with vascular or neurodegenerative pathology, such as vascular dementia (VD), LBD, or TDP‐43, present in a large majority (∼70%) of individuals diagnosed with AD. A novel index, Deep Signature of Pathology Atrophy Recognition (DeepSPARE), identified neuroimaging features associated with AD, VD, and LBD ⁹⁴ and was associated with distinguishing features, such as cognitive scores and Aβ and tau deposition characteristic of AD, WMH volume and CAA characteristic of VD, and CSF α‐synuclein and Lewy body stages characteristic of LBD.

5.2.3. CSF and exploratory blood biomarkers

Established CSF and, more recently, plasma AD biomarkers detect AT(N) characteristics. Clinical‐stage AT(N) plasma assays, including p‐tau217, Aβ42/Aβ40, NfL, and GFAP, are reviewed in Section 4, while this section focuses on CSF and exploratory blood‐based biomarkers not yet in late‐stage clinical validation.

One challenge in using CSF AD biomarkers to detect treatment effects in clinical trials is the lack of standardization of different analyte assays. To enable quantitative comparison across different biomolecules, studies, and laboratories, one study proposed the CentiMarker metric, developed in part using ADNI biomarker data. ⁹⁵ CentiMarker is a standardized scale, anchored at no abnormality (0) and nearly maximum abnormality (100). It allows measurement of a treatment effect in terms of biological effect, such as a return to biomarker normality, and the direct comparison across different AD biomarkers.

Disease progression is complex, and studies increasingly support processes such as metabolic disturbances, neuroinflammation, and synaptic dysfunction as being integrally associated with AD pathological proteins (Veitch et al., manuscript in preparation). Proteomics approaches have identified panels of proteins that reflect physiological processes underlying disease progression and that have prognostic or diagnostic capabilities. These additional biomarkers may provide additional information beyond AT(N) biomarkers. Eleven proteins beyond Aβ and p‐Tau (α‐synuclein, ApoE, CLU, GFAP, GRN, NfL, NRGN, SNAP‐25, TREM2, VILIP‐1, YKL‐40) were predictive of AD. ⁹⁶ Of these, NfL was the most predictive, and CSF proteins outperformed plasma proteins. While some of these remain to be fully characterized, they offer prognostic potential.

Microglial activation in response to Aβ accumulation triggers a cascade of neuroinflammatory processes, resulting in tau phosphorylation. Osteopontin, a protein produced by pro‐inflammatory microglia, may serve as a novel biomarker of AD. CSF osteopontin levels were significantly elevated in A+N+ and A+T+ non‐demented participants from Pre‐symptomatic Evaluation of Experimental or Novel Treatments for AD (PREVENT‐AD) and ADNI. ⁹⁷ Elevated osteopontin was associated with an accelerated conversion to dementia. ⁹⁶ Similarly, levels of CSF 14‐3‐3ζ, a protein that interacts with p‐tau in neurofibrillary tangles, were elevated in ADNI AD participants, particularly in T+ individuals, compared to CU individuals. ⁹⁸ These elevated levels distinguished A+T+ from A+T– individuals (AUC 0.89) and were closely linked to CSF biomarkers of synaptic dysfunction and neuroinflammation, cognitive decline, and neuroimaging markers of disease progression.

A panel of 48 key CSF proteins predicted clinical diagnosis, neurodegeneration, and cognitive impairment comparably to established AD CSF biomarkers (Aβ42, p‐tau181, t‐tau) and further enhanced their predictive ability. ⁹⁹ CSF Aβ42, p‐tau181, and t‐tau reflect early AD pathology but do not detect moderate to late tau deposition, which is required for accurate disease staging and has greater correlations with cognitive decline. Two 16‐protein panels were identified in ADNI participants that outperformed established AD biomarkers in predicting biomarker transitions and disease progression over ten years. ¹⁰⁰ The first, distinguishing between A–T– and A+T– individuals, contained proteins involved predominantly in neuroinflammation and mitochondrial dysfunction, whereas the second, distinguishing between A+T– and A+T+ individuals, contained proteins implicated in tau phosphorylation and protein clearance, supporting the biological validity of the panels. A plasma protein panel including angiotensin predicted neuropsychiatric symptoms independently of AD pathology, suggesting broader applications of proteomics approaches. ¹⁰¹

However, discovery of predictive proteins is often limited by the number of proteins studied (typically several thousand) or by sample size. To overcome these constraints, a large study measured over 7000 CSF proteins in over 2200 participants from four cohorts, including ADNI, grouped by A/T status. ¹⁰² These analyses developed a 10‐protein panel comprising proteins with diverse mechanisms related to AD pathogenesis, including neuronal cell death and neuroinflammation, which rivalled CSF Aβ42 and p‐tau181 in diagnostic performance (Figure 8). It accurately distinguished between biomarker positive (A+T+) and biomarker negative (A–T–) (AUC > 0.98), between AD and CU (AUC 0.89), and between A+ and A– (AUC 0.93) participants, but did not detect other dementia disorders. In patients undergoing amyloid‐ or tau‐targeted therapies in whom CSF Aβ42 and p‐tau181 are unreliable, this panel may serve as an alternative biomarker. Notably, the same 10 proteins also showed good predictive power in plasma, albeit with slightly lower AUCs. This overcomes a key disadvantage of CSF proteomic panels, namely the need for invasive lumbar puncture. Testing the efficacy of these CSF protein panels in plasma is, therefore, a high priority for the development of proteomics biomarkers.

FIGURE 8.

Performance of the 10‐protein AD prediction model. (A) Identification of a 10‐protein panel used to construct the AD prediction model. Classification performance of the resulting model in distinguishing (B) A−T− from A+T+ individuals; (C) clinical AD from CO individuals; and (D) DLB, FTD, PD, and other non‐AD individuals relative to controls. (E) Time‐to‐event analysis for development of AD, shown by Kaplan–Meier curves (with 95% confidence intervals), comparing individuals predicted to be proteomic signature‐positive (green) versus proteomic signature‐negative (red). Reproduced under open access from. ¹⁰² AD, Alzheimer's disease; A−T−, amyloid‐negative/tau‐negative; A+T+, amyloid‐positive/tau‐positive; CO, cognitively normal controls; corr, Pearson correlation; DAA, differential abundance analysis; DLB, dementia with Lewy bodies; FTD, frontotemporal dementia; PD, Parkinson's disease.

Predictive plasma proteomics panels have also been reported without initial discovery in CSF. ¹⁰³ , ¹⁰⁴ These also include proteins involved in processes such as immunity, neuroinflammation, neuronal damage, and cellular migration. In CU participants from the Framingham Heart Study (FHS) with replication in ADNI, plasma protein risk scores predicted the risk of developing MCI or AD and were associated with cognitive decline, plasma AD biomarkers, and brain atrophy. ¹⁰³ A study in a Singaporean memory cohort with replication in ADNI confirmed GFAP, NEFL, AREG, and PPY as predictors of cognitive decline over 4 years. AREG is a novel protein thought to be involved in neuroinflammation. ¹⁰⁴ The consistent identification of proteins from these biological pathways across diverse cohorts and the similar efficacy of a panel in CSF and plasma ¹⁰² suggest that plasma proteomics panels may find widespread use as non‐invasive biomarkers providing complementary information to plasma AD biomarkers.

Sporadic and genetically defined AD appear to have distinct proteomic signatures.  Proteomic analysis of brain tissue, CSF, and plasma helped identify biomarkers that differentiate between sporadic and genetically defined AD, including ADAD and a subtype defined by TREM2 risk variants. ¹⁰⁵ A panel of 8 brain, 40 CSF, and 9 plasma proteins was altered in sporadic AD (Figure S3) and in ADAD, but with greater effect sizes. Proteins in this panel reflect AD‐related biological pathways, as well as innate immune response and Parkinson's disease pathways. An additional proteomic signature differentiated between individuals with sporadic AD, and those with TREM2 risk variant carriers.

Blood is also a rich source of other biological molecules beyond proteins that may have potential as AD biomarkers, providing complementary information beyond Aβ and tau biomarkers. Lipids and other metabolites are altered during the progression of AD. Reductions in levels of ether phosphatidylcholine species, possibly reflecting early metabolic alterations in disease progression such as deterioration in peroxisome function, predicted a two‐fold higher likelihood of conversion to dementia in ADNI MCI participants who were initially stable. ¹⁰⁶ A dementia risk score constructed from changes in these lipid species discriminated between MCI and CU in an external validation cohort (Aspirin in Reducing Events in the Elderly [ASPREE] clinical trial) and was associated with Aβ, tau and hypometabolism, outperforming a model comprised of clinical variables and APOE ε4 status.

MicroRNAs (miRNAs) are small, non‐coding RNAs that regulate gene expression. Polymorphisms in the genes encoding for hsa‐miR‐29c‐3p and hsa‐miR‐146a‐5p were associated with CSF Aβ42, sTREM2, and β‐site APP cleaving enzyme 1 (BACE1) activity, and their expression patterns correlated with cognitive decline and changes in established AD biomarkers. ¹⁰⁷ Nine baseline plasma miRNAs were significantly associated with Aβ status (CSF Aβ42 or Aβ PET), two with tau status (CSF p‐tau or tau‐PET), and eight with neurodegeneration (MRI atrophy or CSF t‐tau). ¹⁰⁸ The addition of miRNAs to demographic data increased classification performance for AT(N) positivity by up to 9%. Pathway enrichment analysis in the studies identified pathways relevant to AD, including apoptosis and inflammation, ¹⁰⁷ and estrogen receptor and insulin growth factor 1 signaling. ¹⁰⁸

The epigenetic mechanism of DNA methylation is integrally involved in AD. 44 cytosine preceding guanine nucleotides (CpGs) and 44 differentially methylated regions were associated with time to dementia onset in CU participants from FHS and ADNI. ¹⁰⁹ A DNA methylation risk score derived from these changes predicted future cognitive decline in an independent cohort. The DNA methylation changes were concentrated in regions associated with neuroinflammatory response and metabolic dysfunction. Genetic factors beyond the APOE ε4 allele may provide additional predictive ability. A deep learning model of SNPs in chromosome 19 that considered epistatic interactions identified 35 SNPs that predicted decline, of which the most powerful were rs561311966 (APOC1) and rs2229918 (ERCC1/CD3EAP). ¹¹⁰

Panels of protein, lipids, miRNAs, and epigenetic changes are all associated with relevant biological pathways that complement information provided by AT(N) plasma biomarkers. While it appears likely that blood biomarkers will eventually supersede CSF biomarkers, the question of which biological molecule will be most successful has yet to be determined. Practical considerations such as cost, ease, and consistency of measurement, and stability of biomolecules will need to be considered.

5.2.4. Comparison of AT(N) biomarkers in diagnosis

The 2024 Alzheimer's Association revised criteria for diagnosis and staging of Alzheimer's disease use biomarkers, along with clinical symptoms, to biologically stage AD in research settings and are increasingly applied in clinical diagnosis. ¹¹¹ Plasma, CSF and neuroimaging AT(N) biomarkers, however, are not entirely equivalent as they measure slightly different pathological forms (soluble vs. insoluble Aβ and tau) or indicators of neurodegeneration (CSF and plasma t‐tau, hypometabolism, atrophy). ¹¹² A systematic comparison of the diagnostic ability of these AT(N) biomarkers in ADNI participants reported that CSF Aβ42/Aβ40 (AUC 0.84), tau PET (AUC 0.92), and FDG‐PET (AUC 0.91) best distinguished ADNI CU from dementia participants. ¹¹³ Both FDG‐PET (AUC 0.82) and tau PET (AUC 0.79) were effective in distinguishing MCI from dementia, though all biomarkers showed limited diagnostic accuracy (AUCs < 0.7) in distinguishing MCI from CU participants (Figure S4).

5.2.5. Multimodal approaches

Combinations of multiple modalities may more accurately predict progression than single modalities. A Multimodal Hazard Score (MHS) combining age, genetic (PHS), MRI, and cognitive data from ADNI, predicted MCI to dementia progression with a hazard ratio of 27 (80th vs. 20th percentile). ¹¹⁴ This reduced theoretical clinical trial sample sizes by 67%. ¹¹⁴ Similarly, individual level risk models comprising stage‐dependent combinations of clinical, cognitive, MRI, and CSF data predicted clinical conversion. ¹¹⁵ The combination of age, cognitive tests, and CSF biomarkers predicted conversion from CU to MCI over 5 years (AUC 0.81), whereas cognitive tests, depression, hippocampal volume, whole brain volume, CSF biomarkers, and PHS predicted MCI conversion to dementia over the same period (AUC 0.92). A combination of sMRI, Aβ PET, CSF, cognition, and genetics data accurately identified current and future AD cases and predicted 90.8% of MCI to dementia conversions and 88.5% of MCI to CU reversions. ¹¹⁶ Finally, a nomogram comprising hippocampal volume, APOE ε4 status, cognitive and functional scores, and CSF sTREM2 levels, predicted cognitive decline in non‐demented T/N+ ADNI participants, independently of Aβ status. ¹¹⁷

These studies report greater diagnostic and prognostic powers using combined modalities, but it remains to be seen whether multimodal models will see clinical adoption. The trade‐off between the time, expense, and patient burden of additional screening must be weighed against the added prognostic power when considering these approaches for further development, particularly in light of the FDA approval of the Lumipulse assay as a biomarker of AD. Instead, there may be a place for multiple modalities applied sequentially in the clinic, with inexpensive initial screening using plasma biomarkers or neuropsychological tests followed by further scans as needed, in the “funnel” manner described by Gladstein 2025 for clinical trial enrollment. ⁴²

Multiple modalities also predicted Aβ positivity. A model integrating MRI data with demographic data, APOE ε4 genotype, and neuropsychological scores achieved moderate accuracy ¹¹⁸ and a second achieved higher accuracy by combining these with plasma biomarkers in a Chinese dataset and in ADNI MCI participants (AUC = 0.97). ¹¹⁹ Readily available clinical information may therefore augment plasma biomarkers in predicting Aβ status.

6. CONCLUSIONS

The period between 2023 and mid‐2025 saw extraordinary progress in the field of AD clinical care, with the FDA approving anti‐amyloid therapies and a plasma AD biomarker. By providing high‐quality, longitudinal, multimodal data, including sMRI, fMRI, PET imaging, CSF biomarkers, genetic data, and detailed neuropsychological assessments, ADNI has enabled numerous studies central to these advances. Notably, the rapid development of ML techniques has accelerated progress by enabling the extraction of meaningful data from single modalities and the selection of powerful combinations of different modalities. The biological significance of many of the results reported here is supported by ADNI studies of disease progression (Veitch et al., manuscript in preparation).

Overall, ADNI studies have: (1) helped validate plasma Aβ42/Aβ40 as a biomarker of Aβ pathology and plasma p‐tau217 as an AD biomarker associated with regional Aβ and tau, and describe plasma NT1‐tau as an AD‐specific biomarker of neurodegeneration; (2) explored digital approaches to clinical trial recruitment, refined subject selection through risk stratification, and reduced trial costs and sample sizes using machine learning and predictive modeling; (3) improved existing cognitive outcome measures using psychometric approaches and described surrogate outcome measures for clinical trials; (4) explored the meaning of a “perceived clinical difference” in disease modifying therapies; (5) supported the assessment of disease‐modifying therapies by providing longitudinal imaging, fluid biomarker, and cognitive data used to simulate treatment outcomes, predict real‐world effectiveness, and evaluate societal and economic impacts of anti‐amyloid therapies such as lecanemab and donanemab; (6) assessed the long term effects of ChEI use and treatment of modifiable risk factors; (7) helped improve sensitivity of cognitive measures, such as CDR‐SB and ADAS‐Cog, and explored use of the discrepancy between self‐ and partner‐reported cognition for early disease stages diagnostic strategies; (8) developed ML‐driven tools that combine cost‐effective and minimally invasive inputs—such as brief cognitive assessments and MRI—to guide diagnosis and treatment; (9) described advances in diagnosis and prediction using imaging, especially in preclinical and prodromal stages and improved staging accuracy, identified distinct subgroups of disease progression; (10) characterized novel CSF biomarkers and proteomics panels and expanded proteomics approaches to less invasive plasma; (11) explored novel blood biomarker classes such as microRNAs, lipidomics, and epigenetic changes; (12) improved generalizability and harmonization of data across cohorts, helping to validate findings in more representative samples, standardize biomarker thresholds, and align cognitive and imaging measures across international studies.

These contributions clearly demonstrate ADNI's continued importance as a foundational resource in translational AD research. As ADNI‐4 brings digital tools and participant engagement into focus, its impact is likely to grow, supporting the next generation of diagnostic, therapeutic, and clinical trial innovations.

CONFLICT OF INTEREST STATEMENT

Shaveta Kanoria is employed by both NCIRE and Houston Methodist Academic Institute. Only NCIRE provided financial support for the work presented in this manuscript. Dallas P. Veitch and Melanie J. Miller have no conflicts to declare. Paul S. Aisen reports grants from NIH, Lilly, Eisai, the Alzheimer's Association. He consults with Merck, Roche, Genentech, Abbvie, Biogen, ImmunoBrain Checkpoint, AltPep, and Neurimmune. He serves on an advisory board for Roche. Laurel A. Beckett reports institutional support from NIH/NIA grant U19 AG024904. She serves on DSMBs for clinical trials at UC Davis and UCSF, and on external advisory boards for Alzheimer's Disease Centers at UCSD, University of Pittsburgh, Washington University in St. Louis, and UCSF. Robert C. Green receives compensation for advising the following companies: Allelica, Atria, Fabric, and Genomic Life; and is co‐founder of Genome Medical and Nurture Genomics. Danielle J. Harvey provides consultation to NervGen Pharma Corp and serves on the PLOS One Statistical Advisory Board. Clifford R. Jack Jr. reports institutional support from NIH and no other disclosures. William Jagust reports institutional support from the Alzheimer's Association, Roche/Genentech, and NIH/NIA. He serves on an advisory board for Lilly and holds stock in Molecular Medicine and Optoceutics. Edward B. Lee reports institutional support from NIH grants and the Delaware Community Foundation. He provides consultation to WaveBreak Therapeutics and Lilly, and honoraria from academic institutions and grant review panels. He has received travel support from multiple foundations and holds a patent (VCP Activators) unrelated to this manuscript. He serves on the Executive Council of the American Association of Neuropathologists. Kwangsik Nho reports institutional support from NIH. Rachel Nosheny reports institutional support from NIH and grants from the California Department of Public Health and Genentech, Inc. She serves on the advisory board of the International Neurodegenerative Disease Research Center and holds a leadership role in ISTAART. All other disclosures are reported as none. Ozioma C. Okonkwo reports institutional support from NIH and serves as Treasurer of the International Neuropsychological Society. Richard J. Perrin reports institutional support from NIH and foundation grants. Ronald C. Petersen consults with Roche, Genentech, Eli Lilly, Eisai, Novo Nordisk, and Novartis, and honoraria from Medscape. Monica Rivera Mindt reports institutional support from NIH and foundation grants. She received honoraria for multiple speaking engagements and consulting fees from Harvard University. Andrew J. Saykin reports institutional support from NIH and serves on several advisory boards and DSMBs. Leslie M. Shaw reports institutional support from NIH/NIA, DOD, and FNIH. He has received consulting fees and honoraria from Biogen and Roche, and in‐kind equipment support from Fujirebio and Roche. Arthur W. Toga reports institutional support from NIH and the Alzheimer's Association. He received honoraria from the Korean Human Brain Mapping Congress and serves on multiple advisory boards and leadership committees. Duygu Tosun reports institutional support from NIH. Susan M. Landau reports institutional support from NIH. She also consults with Banner Health and honoraria from Eisai, ATRI/IMPACT‐AD, and J&J. She received travel support from Alzheimer's Association and other organizations, and serves on advisory boards and editorial committees. Michael W. Weiner received institutional support for his research from the following funding sources: National Institutes of Health (NIH)/NINDS/National Institute on Aging (NIA), Department of Defense (DOD), California Department of Public Health (CDPH), University of Michigan, Siemens, Biogen, Hillblom Foundation, Alzheimer's Association, Johnson & Johnson, Kevin and Connie Shanahan, GE, VUmc, Australian Catholic University, The Stroke Foundation, and the Veterans Administration. He is employed by Northern California Institute for Research and Education (NCIRE), and University of California. He has served on Advisory Boards for Acumen Pharmaceutical, Alzheon, Inc., Amsterdam UMC; MIRIADE, Cerecin, Merck Sharp & Dohme Corp., NC Registry for Brain Health, ProMIS Neurosciences, Inc., and REGEnLIFE. He also serves on the USC ACTC grant which receives funding from Eisai. He serves on the Editorial Board for the Journal for Prevention of Alzheimer's Disease (JPAD) and served on the Editorial Board for Alzheimer's & Dementia from 2005 to 2025. He has provided consulting to Acadia Pharmaceuticals, Acumen Pharmaceuticals, Alzeca, Alzheon, Inc., Anven, ALZpath, Boxer Capital, LLC, Cerecin, Inc., Clario, Dementia Society of Japan, Dolby Family Ventures, Eisai, GLG Consulting, Guidepoint, Health and Wellness Partners, Indiana University, IXICO, LCN Consulting, MEDA Corp., Merck Sharp & Dohme Corp., Duke U.; NC Registry for Brain Health, NovoNordisk, Owkin France, ProMIS Neurosciences, Prova Education, Quantum Leap Health, REGEnLIFE, Sai MedPartners, T3D Therapeutics, U. Penn, University of Southern California (USC), and WebMD. He has acted as a speaker/lecturer for BrightFocus Foundation, China Association for Alzheimer's Disease (CAAD) and Taipei Medical University, as well as a speaker/lecturer with academic travel funding provided by: AD/PD Congress, Amsterdam UMC, Banner Health, Cleveland Clinic, CTAD Congress, Foundation of Learning; Gates Ventures, Health Society (Japan), Kenes International, U. Madison Wisconsin, U. Penn, U. Toulouse, Japan Society for Dementia Research, Korean Dementia Society, Merck Sharp & Dohme Corp., National Center for Geriatrics and Gerontology (NCGG; Japan), University of Madison Wisconsin, University of Southern California (USC, and Stead Impact Ventures). He holds stock options with Alzeca, Alzheon, Inc., ALZPath, Inc., and Anven. Author disclosures are available in the supporting information.

CONSENT STATEMENT

All human subjects provided informed consent.

Supporting information

Supporting Information

Supporting Information

Supporting Information

Kanoria S, Veitch DP, Miller MJ, et al. Clinical impact of the Alzheimer's Disease Neuroimaging Initiative: A review of studies using ADNI data (2023 to June 2025). Alzheimer's Dement. 2026;22:e71353. 10.1002/alz.71353