Eigen-Adjusted Functional Principal Component Analysis

Ci-Ren Jiang, Eardi Lila, John A. D. Aston, Jane-Ling Wang

TL;DR

This paper introduces "eigen-adjusted FPCA," a novel method for analyzing functional data that incorporates covariate information into the covariance structure more efficiently than existing approaches. This new method is particularly useful for understanding complex relationships in data like brain imaging, offering a flexible and computationally feasible way to perform dimension reduction.

What Problem Does This Paper Address?

Functional Principal Component Analysis (FPCA) is a powerful tool for reducing the dimensionality of functional data (data that varies over a continuous domain, like time or space). However, when this data also depends on external factors called "covariates" (e.g., patient demographics, spatial locations), existing FPCA methods face challenges. Some methods only adjust the "mean function" (the average shape of the data), which is insufficient for capturing complex variations. Others adjust both the mean and covariance functions, but these can be computationally expensive and make subsequent analyses difficult. This paper addresses the need for an FPCA method that efficiently incorporates covariate information into the covariance structure, leading to more flexible and interpretable results, especially for high-dimensional data like functional Magnetic Resonance Imaging (fMRI).

Methods

The core innovation is the "eigen-adjusted FPCA" model. Instead of letting the entire covariance function vary with covariates, this method focuses on how the eigenvalues of the covariance function change with covariates, while keeping the eigenfunctions (which represent the modes of variation) independent of covariates. This provides a common reference frame for comparing different covariate effects.

Key methodological components include:

  • Covariate-dependent eigenvalues: The model assumes that the eigenvalues, which capture the strength of different functional components, are functions of the covariates.
  • Covariate-independent eigenfunctions: The eigenfunctions, representing the shapes of these components, are assumed to be the same across all covariate values.
  • Local linear smoothers: Used to estimate the mean function and the pooled covariance function, which is a simplification of the full covariate-dependent covariance.
  • Weighted Least Squares (WLS): A novel approach proposed to estimate the covariate-specific eigenvalues, designed to work with both dense and sparse data.
  • Eigen-decomposition: Applied to the pooled covariance to obtain the common eigenfunctions.

Datasets

The paper uses the following datasets:

  • Simulated data: Used extensively for evaluating the performance of the proposed methods under various scenarios.
  • Human Connectome Project (HCP) resting-state fMRI data: This dataset of 40 unrelated healthy subjects is used to apply the eigen-adjusted FPCA model to investigate functional connectivity in the brain.

Key Insights and Ideas

  • Focusing on eigenvalues is key: By allowing only the eigenvalues to vary with covariates, the model achieves a balance between flexibility and computational efficiency. This is a significant departure from previous methods that either ignored covariate effects on covariance or attempted to model the entire covariate-dependent covariance function.
  • Common eigenfunctions as a reference frame: Having covariate-independent eigenfunctions is crucial for comparing the effects of different covariates on the functional components. This allows for meaningful statistical analysis of the covariate-specific eigenvalues.
  • WLS for eigenvalue estimation: The proposed WLS approach is shown to be effective for estimating covariate-specific eigenvalues, outperforming a PC-based approach in simulations, especially for sparse data.
  • Application to brain connectivity: The method successfully identifies distinct brain regions (parcellations) based on their functional connectivity patterns, demonstrating its utility in real-world complex data analysis.

Key Results

  • Simulation studies: The proposed WLS estimator for eigenvalues consistently outperformed a PC-based approach in terms of Integrated Squared Error (ISE) across different sampling schemes (complete and sparse data).
  • fMRI application: The eigen-adjusted FPCA successfully identified brain regions with distinct functional connectivity patterns. For example, it was able to separate the motor cortex from the rest of the cerebral cortex and identify the visual cortex as a distinct cluster. The results were comparable to established parcellation methods, but derived in a more data-driven manner.
  • Computational efficiency: The eigen-adjusted FPCA is shown to be computationally more efficient than fully-adjusted FPCA, making it suitable for large datasets like fMRI.

Limitations and Future Work

  • Approximation when eigenfunctions vary: The paper acknowledges that if the eigenfunctions themselves directly vary with covariates, the eigen-adjusted FPCA model provides an approximation. However, simulations suggest this approximation is still considerably better than simpler approaches.
  • Potential for negative eigenvalue estimates: In finite samples, the WLS estimator might produce negative estimates for positive eigenvalues. The authors suggest potential remedies like truncation or non-negative least-squares approaches, noting that this is a finite sample issue.
  • Future directions: The paper suggests that as more research questions focus on second-order structures in functional data (like functional connectivity), techniques like eigen-adjusted FPCA will become increasingly important and utilized. Further investigation into the behavior of the model when eigenfunctions vary with covariates could also be a fruitful area.

Summary updated

BibTeX
@article{jiang2022eigen,
  author = {Ci-Ren Jiang and Eardi Lila and John A. D. Aston and Jane-Ling Wang},
  title = {{Eigen-Adjusted Functional Principal Component Analysis}},
  journal = {J. Comput. Graph. Stat.},
  year = {2022},
  doi = {10.48550/arxiv.2204.05622},
}

Loading related content...

Figure