🧠 neurology

ATN Classification and Machine-Learned Plasma Biomarker Phenotypes Reveal Distinct Alzheimer's Pathology in a Population-Based Cohort

In a large population-based cohort, this study demonstrates that while theory-driven ATN classification and data-driven machine learning phenotypes show only modest concordance—largely driven by GFAP rather than shared amyloid, tau, and neurodegeneration biomarkers—both frameworks effectively predict longitudinal cognitive decline, suggesting that integrating these complementary approaches offers a more comprehensive characterization of Alzheimer's pathology.

Original authors: Chea, E. F.

Published 2026-01-15

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Chea, E. F.

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

The Big Picture: Two Ways to Sort the Puzzle Pieces

Imagine Alzheimer's disease research is like trying to sort a massive box of mixed-up puzzle pieces. Scientists want to know which pieces belong together to form the picture of the disease.

This paper compares two different ways of sorting these pieces using blood tests (plasma biomarkers) from over 4,400 older adults in the United States.

The "Rulebook" Method (ATN): This is like a strict librarian who has a rulebook. If a piece is red, it goes in Box A. If it's blue, it goes in Box B. The rules are clear-cut (binary): you either have the "Amyloid" piece, the "Tau" piece, or the "Neurodegeneration" piece, or you don't.
The "Pattern Finder" Method (Machine Learning): This is like an artist looking at the box without a rulebook. They let the pieces naturally clump together based on how similar they look. They don't force them into boxes; they just see which pieces naturally stick together.

The researchers wanted to see: Do these two methods sort the pieces the same way? And do the groups they create predict who will have memory problems later?

The Ingredients: What Was Measured?

The scientists looked at four specific "ingredients" in the blood:

Amyloid (A): A sticky protein associated with Alzheimer's.
Tau (T): Another protein that tangles up in the brain.
Neurodegeneration (N): Signs that brain cells are dying.
GFAP: A marker of inflammation (swelling/irritation) in the brain.

The Twist: The "Rulebook" (ATN) only uses the first three ingredients. It ignores the fourth one (GFAP) because the rulebook wasn't designed for it. The "Pattern Finder" (Machine Learning) uses all four ingredients to make its groups.

The Results: How Did They Compare?

1. The Sorting Was Surprisingly Different

When the researchers compared the "Rulebook" groups with the "Pattern Finder" groups, they found they didn't match very well.

The Analogy: Imagine the Rulebook sorts people into "Tall," "Medium," and "Short." The Pattern Finder sorts them into "Muscular," "Lean," and "Heavy."
The Finding: A person who is "Tall" in the Rulebook might be "Lean" in the Pattern Finder. The two methods only agreed about 11-13% of the time. They are looking at the data through different lenses.

2. The Secret Ingredient: GFAP

The researchers discovered why the two methods agreed even a little bit. It was almost entirely because of that fourth ingredient, GFAP.

The Analogy: GFAP is like a bridge. It connects the "Rulebook" world to the "Pattern Finder" world.
The Proof: When the researchers removed GFAP from the Pattern Finder's list and forced it to use only the same three ingredients as the Rulebook, the agreement dropped to almost zero (3%).
The Takeaway: The two methods are actually looking at very different things. The small amount of agreement we saw was mostly because GFAP happens to fit into both systems. Without it, the two methods would be completely independent.

3. Finding Hidden Groups

The "Pattern Finder" (Machine Learning) found some groups the "Rulebook" missed:

The "Severe" Group: A tiny group (1.2%) with very high levels of all bad markers. The Rulebook caught most of them, but the Pattern Finder isolated them perfectly.
The "Mystery" Group: A very small group (0.3%) that had signs of brain damage but no signs of the classic Alzheimer's proteins (Amyloid or Tau). The Rulebook would have labeled them "Normal," but the Pattern Finder saw they were actually sick with something else (perhaps vascular issues).
The "Big Middle" Group: The Pattern Finder found one giant group (78% of everyone) that the Rulebook had chopped up into many tiny, separate boxes. The Pattern Finder realized that for most people, the disease is a smooth gradient, not distinct steps.

4. Predicting the Future

Both methods tried to guess who would lose memory over the next four years.

The Result: Both were successful, but only slightly. They could explain about 2% of why people's memory changed.
The Analogy: It's like trying to predict the weather. Knowing the temperature (biomarkers) helps, but you also need to know the wind, humidity, and pressure (genetics, lifestyle, other diseases). The blood tests are just one piece of a very complex puzzle.

The Conclusion: Why This Matters

The paper concludes that neither method is "wrong," but they are different.

The Rulebook (ATN) is great for clear communication. It gives doctors and researchers a common language (Positive/Negative) to talk about the disease.
The Pattern Finder (Machine Learning) is great for discovery. It finds the "gray areas" and the rare, weird cases that the strict rules miss.

The Final Metaphor:
Think of the Rulebook as a black-and-white photo. It's sharp, clear, and easy to understand, but it misses the subtle shades of gray.
Think of the Pattern Finder as a high-definition color photo. It shows all the details, the gradients, and the hidden textures, but it's harder to summarize in a single sentence.

The study suggests that to truly understand Alzheimer's in the general population, we shouldn't just pick one. We need to use the Rulebook to get the basics right and the Pattern Finder to see the full, complex picture.

Important Note: The paper emphasizes that these findings are based on blood tests in a general population. They are not yet a tool for doctors to diagnose individual patients in a clinic, but rather a way to understand how the disease works in the real world.

Technical Summary: ATN Classification and Machine-Learned Plasma Biomarker Phenotypes in a Population-Based Cohort

Problem Statement
Alzheimer's disease (AD) research currently relies on two divergent approaches for characterizing pathology: the theory-driven ATN (Amyloid/Tau/Neurodegeneration) framework and data-driven unsupervised machine learning phenotyping. The ATN framework, established by the National Institute on Aging and Alzheimer's Association, classifies individuals using binary cutoffs for biomarkers (Aβ, Tau, Neurodegeneration). While clinically useful, this approach dichotomizes continuous biomarker data, potentially obscuring subthreshold biological variation. Conversely, unsupervised machine learning methods (e.g., clustering) identify latent phenotypes without a priori constraints but lack direct mapping to established pathological constructs. The concordance between these approaches in population-representative samples, as opposed to enriched clinical cohorts, remains incompletely characterized. This study addresses the gap in understanding how theory-driven binary classifications align with data-driven continuous phenotyping in a nationally representative sample of older adults.

Methodology
The study analyzed data from 4,465 participants (mean age 69.7 years) from the Health and Retirement Study (HRS) 2016 Venous Blood Study, a nationally representative longitudinal cohort. Four plasma biomarkers were measured: Aβ42/40 ratio, p-tau181, Neurofilament Light Chain (NfL), and Glial Fibrillary Acidic Protein (GFAP).

ATN Classification: Participants were classified into 14 observed ATN profiles using literature-based binary cutoffs (Aβ42/40 < 0.067 for A+; p-tau181 > 2.2 pg/mL for T+; NfL > 20 pg/mL for N+). GFAP was excluded from the primary ATN classification due to its lack of a specific domain mapping.
Data-Driven Clustering: Unsupervised machine learning was applied to the four biomarkers (including GFAP). The study employed:
- K-means clustering: To identify optimal clusters ( $k$ ) using the elbow method, silhouette analysis, and NbClust.
- Gaussian Mixture Modeling (GMM): As a probabilistic alternative.
- Variational Autoencoder (VAE): A deep learning approach for non-linear dimensionality reduction to a 2D latent space.
Agreement and Validation: Concordance between ATN profiles and clustering results was quantified using the Adjusted Rand Index (ARI) and Normalized Mutual Information (NMI). Stability was assessed via bootstrap resampling and Jaccard similarity coefficients.
Sensitivity Analyses: The study conducted extensive sensitivity tests, including:
- Excluding GFAP from clustering to isolate the contribution of the three core ATN biomarkers.
- Comparing $k=3$ vs. $k=4$ solutions to evaluate the resolution of rare subgroups.
- Testing alternative distance metrics (Manhattan vs. Euclidean).
- Varying ATN cutoff thresholds.
Longitudinal Analysis: Both classification systems were evaluated for their ability to predict 4-year cognitive decline (2016–2020) using multivariable linear regression.

Key Results

ATN Distribution: The ATN framework yielded 14 profiles. The most prevalent were A+/T-/N- (27.4%) and A-/T-/N- (22.6%). Profiles consistent with full AD pathology (A+/T+/N+) were associated with the lowest cognitive scores and highest dementia prevalence.
Clustering Outcomes: K-means clustering identified four optimal clusters ( $k=4$ $k = 4$ ):
- Cluster 1 (1.2%): Severe AD-like pathology with high NfL, GFAP, and p-tau181.
- Cluster 2 (19.9%): Mixed astrocytic activation and neurodegeneration.
- Cluster 3 (78.6%): A large, heterogeneous group with low-to-moderate pathology spanning the full spectrum of ATN profiles.
- Cluster 4 (0.3%): A small, stable subgroup characterized by non-AD neurodegeneration (low Aβ42/40 ratio, preserved cognition).
Concordance: Agreement between ATN profiles and machine-learned clusters was modest (ARI = 0.119, NMI = 0.113).
- Role of GFAP: A critical finding was that GFAP accounted for the majority of this concordance. When GFAP was excluded from clustering (using only Aβ42/40, p-tau181, and NfL), agreement with ATN dropped drastically to ARI = 0.03. This indicates that the three core ATN biomarkers alone produce groupings nearly orthogonal to ATN binary categories, and the observed alignment is largely driven by GFAP's bridging role across pathological domains.
Dimensionality Reduction: The VAE revealed a localized non-linear structure (a "crescent" trajectory for severe AD pathology) that PCA did not capture, though PCA explained more global variance (67.1%).
Predictive Utility: Both ATN profiles and clusters predicted 4-year cognitive decline with modest effect sizes (ATN $R^2$ = 0.024; Clusters $R^2$ = 0.019). Neither approach significantly outperformed the other, and both effects were consistent with the multifactorial nature of cognitive aging in population samples.

Significance and Claims
The paper claims that theory-driven ATN classification and data-driven phenotyping capture "partially overlapping but largely distinct information." The modest concordance (approx. 11–13% shared information) is not an analytical failure but a reflection of fundamental methodological differences: ATN imposes binary boundaries on continuous biological gradients, while clustering preserves these gradients.

The study emphasizes that:

Binary vs. Continuous: The primary source of discordance is the dichotomization of continuous biomarker data, not the specific placement of cutoff thresholds.
GFAP's Unique Role: GFAP acts as a "bridge" biomarker that aligns clustering with ATN; without it, the two approaches are nearly independent.
Resolution of Rare Phenotypes: Data-driven clustering successfully identified rare, stable biological subgroups (e.g., Cluster 4, non-AD neurodegeneration) that were obscured or merged in the binary ATN framework.
Complementary Utility: Neither framework is superior; rather, they offer complementary lenses. ATN provides clinically interpretable, biologically grounded stratification suitable for regulatory and trial contexts, while unsupervised clustering offers granular phenotyping capable of revealing transitional states and heterogeneous subtypes in population research.

The authors conclude that integrating both frameworks may support a more comprehensive characterization of AD-related pathology in epidemiological research, moving beyond rigid categorical definitions to capture the continuous and heterogeneous nature of biological aging.