Riemannian tangent space mapping and elastic net regularization for cost-effective EEG markers of brain atrophy in Alzheimer's disease

Wolfgang Fruehwirt, Matthias Gerstgrasser, Pengfei Zhang, Leonard Weydemann, Markus Waser, Reinhold Schmidt, Thomas Benke, Peter Dal‐Bianco, Gerhard Ransmayr, Dieter Grossegger, Heinrich Garn, Gareth W. Peters, Stephen Roberts, Georg DorffnerarXiv 2017 · DOI · arXiv

Abstract

The diagnosis of Alzheimer's disease (AD) in routine clinical practice is most commonly based on subjective clinical interpretations. Quantitative electroencephalography (QEEG) measures have been shown to reflect neurodegenerative processes in AD and might qualify as affordable and thereby widely available markers to facilitate the objectivization of AD assessment. Here, we present a novel framework combining Riemannian tangent space mapping and elastic net regression for the development of brain atrophy markers. While most AD QEEG studies are based on small sample sizes and psychological test scores as outcome measures, here we train and test our models using data of one of the largest prospective EEG AD trials ever conducted, including MRI biomarkers of brain atrophy.

Introduction

The clinical diagnosis of Alzheimer’s disease (AD) remains a challenge, often relying on subjective assessments or high-cost neuroimaging techniques such as MRI and PET. While Quantitative EEG (QEEG) offers a low-cost, non-invasive alternative, its clinical utility has been hampered by the inherent noise of brain signals and the limitations of traditional analysis methods. Standard QEEG approaches often treat brain connectivity data—represented as covariance matrices—as if they exist in a flat, Euclidean space. However, these matrices naturally reside on a curved manifold. Ignoring this geometric structure can lead to "swelling effects" and a significant loss of diagnostic information. This paper introduces a framework that leverages Riemannian geometry to better capture the brain's complex signal dynamics, providing a reliable marker for neurodegeneration.

Methodology

The proposed pipeline transforms raw EEG signals into predictive markers through several sophisticated steps:

  1. Feature Generation: EEG segments are represented as Symmetric Positive-Definite (SPD) matrices. The study compares standard Sample Covariance (SCM) with Kendall Rank Correlation (KEN), the latter being more robust to non-linear relationships.
  2. Spatiofrequential Design: To capture a holistic view of brain activity, data from multiple frequency bands ($\delta, \theta, \alpha, \beta_1$) are concatenated into a single "spatiofrequential" matrix.
  3. Riemannian Tangent Space Mapping: To utilize standard machine learning algorithms, the curved covariance matrices are projected into a flat "tangent space." Two mapping techniques were evaluated: TANrie (using the Riemannian/Karcher mean) and TANlog (using the Log-Euclidean mean).
  4. Elastic Net Regression: An elastic net model, which combines L1 and L2 regularization, was employed to predict brain volume. This is particularly effective for high-dimensional EEG data where electrode features are often highly correlated.
  5. Validation: The models were validated using a $10 \times 10$ two-level nested cross-validation, repeated 100 times to ensure the stability and generalizability of the results.

Key Results

The study demonstrates that respecting the intrinsic geometry of EEG data leads to significantly higher predictive accuracy for brain atrophy. Tangent space mapping models consistently outperformed Euclidean reference models ($p = 0.003$ for BrainVol; $p = 0.030$ for HippVol).

Table 1: Performance comparison of Euclidean (EUC) and Tangent Space (TAN) mapping approaches across different model designs.

As shown in the results table, the best-performing model for whole-brain volume (BrainVol) utilized Covariance combined with TANlog and a spatiofrequential design, achieving an RMSE of $1.23 \times 10^{-3}$. For hippocampal volume (HippVol), the optimal model used Kendall Rank Correlation with TANlog, yielding an RMSE of $1.44 \times 10^{-7}$. Notably, the simpler TANlog approach often matched or exceeded the performance of the more computationally intensive TANrie method.

Discussion and Conclusion

The findings highlight several critical insights for AD diagnostics. First, the transition from Euclidean to Riemannian space is essential for capturing the subtle neurodegenerative signatures in EEG. Second, the success of Kendall Rank Correlation in predicting hippocampal volume suggests that memory-related brain interactions are fundamentally non-linear. Finally, the "spatiofrequential" approach confirms that the interaction across different brain frequencies is a vital component of the AD signature.

While the study utilized a relatively large clinical cohort ($N=110$), further validation on even larger datasets and longitudinal tracking of atrophy rates would be beneficial. Nevertheless, this framework represents a significant advancement toward democratizing Alzheimer's diagnostics, transforming standard EEG into a powerful tool for objective brain health assessment.

Summary updated

Abstract

The diagnosis of Alzheimer's disease (AD) in routine clinical practice is most commonly based on subjective clinical interpretations. Quantitative electroencephalography (QEEG) measures have been shown to reflect neurodegenerative processes in AD and might qualify as affordable and thereby widely available markers to facilitate the objectivization of AD assessment. Here, we present a novel framework combining Riemannian tangent space mapping and elastic net regression for the development of brain atrophy markers. While most AD QEEG studies are based on small sample sizes and psychological test scores as outcome measures, here we train and test our models using data of one of the largest prospective EEG AD trials ever conducted, including MRI biomarkers of brain atrophy.
BibTeX
@misc{fruehwirt2017riemannian,
  author = {Wolfgang Fruehwirt and Matthias Gerstgrasser and Pengfei Zhang and Leonard Weydemann and Markus Waser and Reinhold Schmidt and Thomas Benke and Peter Dal‐Bianco and Gerhard Ransmayr and Dieter Grossegger and Heinrich Garn and Gareth W. Peters and Stephen Roberts and Georg Dorffner},
  title = {{Riemannian tangent space mapping and elastic net regularization for cost-effective EEG markers of brain atrophy in Alzheimer's disease}},
  year = {2017},
  eprint = {1711.08359},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1711.08359},
}
Figure