Evaluation of multiple ATN biomarker definitions for prediction of cognitive impairment

Tom Earnest, Braden Yang, Deydeep Kothapalli, Aristeidis SotirasAlzheimer 2024 · DOI

TL;DR

This study investigates how different ways of defining Alzheimer's disease biomarkers (amyloid, tau, neurodegeneration - ATN) impact their ability to predict cognitive impairment. The key finding is that while adding ATN biomarkers generally improves prediction, using simple binary classifications might be sufficient, and more complex definitions don't necessarily offer a significant advantage for cross-sectional prediction.

What Problem Does This Paper Address?

The amyloid/tau/neurodegeneration (ATN) framework is a widely used system for classifying Alzheimer's disease (AD) based on biological markers. However, researchers often operationalize these biomarkers in various ways: some use simple "yes/no" (binary) classifications, while others employ continuous measures or more complex categorical staging systems. This variability makes it difficult to compare findings across studies and raises the question of whether more complex definitions are truly necessary for predicting cognitive impairment. This paper aims to directly evaluate how these different definitions of ATN biomarkers affect their predictive power for cognitive status.

Methods

The researchers used a dataset of 473 individuals from the Alzheimer's Disease Neuroimaging Initiative (ADNI) who had PET scans for amyloid-beta and tau, along with volumetric MRI data. They explored a broad spectrum of ATN biomarker definitions, including:

  • Binary: Simple presence/absence classifications.
  • Non-binary categorical: Staging models or quantiles of continuous values.
  • Continuous: Direct measurement values.

They trained linear regression models to predict the Preclinical Alzheimer Cognitive Composite (PACC) score, a measure of cognitive function. Additionally, they employed Support Vector Machines (SVMs) to create multivariate ATN measures using regional PET uptake and gray matter volumes. The study design involved rigorous cross-validation to ensure robust out-of-sample performance estimates, using Root Mean Square Error (RMSE) as the primary metric. They compared models with ATN predictors against a baseline model using only demographic covariates (age, sex, APOE status) and also compared different ATN operationalizations against each other.

Datasets

The study utilized data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). This is a well-established, publicly available dataset that has been instrumental in AD research. No new benchmarks were introduced; rather, the study leverages existing ADNI data to explore different analytical approaches.

Key Insights and Ideas

  • ATN biomarkers are predictive: The addition of any ATN biomarker definition to a model significantly improved the prediction of cognitive ability compared to using only demographic information. This reinforces the importance of these biological markers in understanding cognitive decline.
  • Tau and combined ATN show strongest benefits: The most substantial improvements in prediction accuracy were observed when tau biomarkers were included, or when all three ATN components (amyloid, tau, and neurodegeneration) were used together.
  • SVMs offer an advantage: Multivariate ATN measures derived using SVMs also outperformed the baseline models, suggesting that integrating information from multiple regions or biomarker types can enhance predictive power.
  • Simplicity might suffice for cross-sectional prediction: Surprisingly, when comparing different ATN definitions against each other, using simple binary classifications for ATN biomarkers did not lead to significantly worse prediction accuracy than using more complex categorical, continuous, or SVM-based measures for cross-sectional cognitive impairment. This suggests that for predicting current cognitive status, simpler definitions might be adequate.
  • Definition matters for feature importance: Feature importance analysis revealed that certain ATN definitions consistently performed better than others, highlighting that the specific way a biomarker is operationalized can influence its perceived importance in predictive models.

Key Results

  • Adding any ATN measure to a model with standard covariates (age, sex, APOE) significantly improved PACC prediction accuracy (Figure 1A).
  • Models including tau or all three ATN assessments showed the largest benefits.
  • SVM-based ATN models also outperformed baseline models.
  • However, when compared to models using only binary ATN measures, the addition of categorical, continuous, or SVM-based ATN measures did not significantly improve prediction accuracy (Figure 1B).

Limitations and Future Work

The authors acknowledge that their findings are based on cross-sectional data. They emphasize that further research is needed to determine if these conclusions hold true for predicting longitudinal cognitive decline or for specific subpopulations, such as those with preclinical AD. The study also suggests that exploring the impact of different ATN operationalizations on predicting other outcomes of interest would be a valuable next step.

Summary updated

Abstract

Abstract Background Application of the amyloid/tau/neurodegeneration (ATN) framework is varied, with some research relying on binary assessment of biomarkers and some using continuous or multidimensional measures. There are few investigations which directly evaluate how differing operationalizations of ATN affect prediction of cognitive impairment. Method We selected 473 individuals from ADNI who received PET imaging for amyloid‐beta (AV‐45) and tau (flortaucipir), as well as volumetric MRI. We evaluated a wide range of methods for deriving ATN biomarker assessments, spanning continuous, binary, and non‐binary categorical (i.e., staging models or quantiles of continuous values) measures (Table 1). We trained linear regression models to predict Preclinical Alzheimer Cognitive Composite (PACC) scores, combining different variations of ATN measures as predictors. We additionally used support vector machines (SVMs) to derive multivariate ATN measures; these models used regional PET uptakes or gray matter volumes (68 FreeSurfer regions) to predict PACC. A first set of experiments tested how addition of ATN improved prediction over a baseline model with just covariates (age, sex, APOE status), while a second set of experiments tested how models including non‐binary or SVM‐based ATN measures compared to ones using only binary measures. We used a repeated, nested, and cross‐validated design to evaluate the out‐of‐sample model performance as quantified by the root mean square error. Error estimates were compared using bias‐corrected t‐tests. Result Compared to models with standard covariates, addition of any ATN measure resulted in an increase in accuracy for predicting cognitive ability (Figure 1A). Benefits were largest for models which included tau or all three ATN assessments. SVM‐based models also outperformed baseline models. However, when compared to models using all‐binary ATN measures, addition of categorical, continuous, or SVM‐based ATN measures did not significantly improve prediction accuracy in any combination (Figure 1B). Feature importance analysis indicated that some ATN definitions consistently outperformed others (Figure 2). Conclusion Binary measures of ATN may be sufficient for prediction of cross‐sectional cognitive impairment. Further work is required to evaluate if these results hold in specific subpopulations (e.g., preclinical AD) or for predicting other measures of interest (e.g., longitudinal cognitive decline).
BibTeX
@inproceedings{earnest2024evaluation,
  author = {Tom Earnest and Braden Yang and Deydeep Kothapalli and Aristeidis Sotiras},
  title = {{Evaluation of multiple ATN biomarker definitions for prediction of cognitive impairment}},
  booktitle = {Alzheimer},
  year = {2024},
  doi = {10.1002/alz.092507},
}

Loading related content...

Figure