Comprehensive evaluation of AT(N) imaging biomarkers for predicting cognition

Tom Earnest, Braden Yang, Deydeep Kothapalli, Aristeidis Sotiras, the Alzheimer’s Disease Neuroimaging InitiativemedRxiv 2024 · DOI

Abstract

We demonstrated that different imaging biomarker definitions can lead to variability in downstream predictive tasks. Researchers should consider how their biomarker operationalizations may help or hinder the assessment of disease severity.

Introduction

In Alzheimer’s Disease (AD) research, the AT(N) framework—representing Amyloid (A), Tau (T), and Neurodegeneration (N)—serves as the biological gold standard for disease classification. However, the field currently lacks a unified standard for defining these biomarkers. Researchers often choose between binary cutoffs (e.g., "amyloid positive"), continuous scales, or complex staging systems. This study systematically evaluates 42 different biomarker definitions to determine how these choices impact the prediction of cognitive performance and what information is lost when complex imaging data is simplified.

Methods

The research team utilized data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI), analyzing a baseline sample of 490 individuals (288 cognitively unimpaired, 163 with mild impairment, and 39 with dementia). The study compared 42 distinct biomarker definitions across three modalities: Amyloid-PET ([18F]-Florbetapir), Tau-PET ([18F]-Flortaucipir), and structural MRI.

Modeling and Validation

The evaluation involved:

  • Variable Types: Testing binary cutoffs, non-binary categorical measures (e.g., quartiles or Braak staging), and continuous measures.
  • Modeling Approaches: Comparing standard Linear Regression against Support Vector Machines (SVMs). The SVMs were designed to extract predictive power from 82 distinct brain regions of interest (ROIs) rather than relying on a single summary average.
  • Validation: Results were validated using 10-fold cross-validation repeated 10 times, with statistical significance assessed via bias-corrected Nadeau-Bengio t-tests.

Results and Key Insights

The Dominance of Tau

Tau biomarkers were consistently identified as the strongest single predictors of cognitive performance. While amyloid is a prerequisite for an AD diagnosis, its presence is a relatively poor predictor of a patient's current cognitive state. In contrast, the accumulation and spread of tau "tangles" correlate closely with functional impairment. The Tau SVM emerged as the top-performing single-pathology model with an RMSE of 0.419.

Comparison of predictive performance across various biomarker definitions for Amyloid, Tau, and Neurodegeneration.

The Limitations of Binarization

A critical finding was that "binarizing" tau and neurodegeneration data—reducing a spectrum of damage to a simple positive/negative label—significantly degrades predictive accuracy. Interestingly, amyloid appears to be an exception; binary cutoffs for amyloid performed nearly as well as continuous measures, suggesting it may function as a binary "trigger" for the disease process rather than a linear driver of symptoms. Through cross-validation, the authors identified optimal SUVR cutoffs for binarization: 1.26 for Amyloid and 1.44 for Tau.

Multivariate vs. Univariate Modeling

The study found that machine learning models (SVMs) looking at spatial patterns across the brain significantly outperformed models using a single summary region. This indicates that the spatial distribution of pathology carries vital information that simple averages miss.

Comparison showing that multivariate SVM models outperform univariate summary measures in predicting cognitive scores.

Imaging vs. Fluid Biomarkers

In a head-to-head comparison involving 246 individuals with available Cerebrospinal Fluid (CSF) data, imaging biomarkers significantly outperformed CSF-based predictions. This is likely because CSF reflects a global pathological load, whereas PET and MRI provide localized information about brain regions directly tied to specific cognitive functions.

Conclusion

The study highlights that combining all three AT(N) categories provides the highest overall accuracy, achieving an RMSE reduction of up to 21.1% over baseline models. While these findings provide a clearer roadmap for biomarker selection, the authors acknowledge limitations, including the lack of demographic diversity in the ADNI cohort. Future research should extend these evaluations to more diverse populations and emerging blood-based biomarkers to ensure these predictive models are robust across all patient groups.

Reference: Earnest, T., et al. (2024). Comprehensive evaluation of AT(N) imaging biomarkers for predicting cognition. medRxiv preprint. https://doi.org/10.1101/2024.11.25.24317943

Summary updated

Abstract

We demonstrated that different imaging biomarker definitions can lead to variability in downstream predictive tasks. Researchers should consider how their biomarker operationalizations may help or hinder the assessment of disease severity.
BibTeX
@misc{earnest2024comprehensive,
  author = {Tom Earnest and Braden Yang and Deydeep Kothapalli and Aristeidis Sotiras and the Alzheimer’s Disease Neuroimaging Initiative},
  title = {{Comprehensive evaluation of AT(N) imaging biomarkers for predicting cognition}},
  year = {2024},
  doi = {10.1101/2024.11.25.24317943},
  howpublished = {medRxiv},
}
Figure