High-dimensional response growth curve modeling for longitudinal neuroimaging analysis

Lu Wang, Xiang Lyu, Lexin LiComput. Stat. Data Anal 2025 · DOI

TL;DR

This paper introduces a novel method for analyzing high-dimensional longitudinal neuroimaging data, specifically addressing the challenge of modeling complex brain changes over time. Their approach combines a low-rank factor model, a re-parameterization with sparsity penalties, and an efficient computational trick to handle massive covariance matrices, enabling more robust and interpretable insights into brain development and disease progression.

What Problem Does This Paper Address?

Longitudinal neuroimaging studies, which track brain structure and function over time, are crucial for understanding development and neurological disorders. However, analyzing this data presents a significant challenge: the sheer volume of measurements (high-dimensional responses) and the complexity of their relationships over time. Traditional growth curve models struggle with the massive covariance matrices that arise, making computation infeasible and parameter estimation unreliable, especially with limited sample sizes. This paper tackles the problem of efficiently and accurately modeling these high-dimensional longitudinal outcomes, aiming to uncover meaningful patterns in brain connectivity changes associated with conditions like HIV.

Methods

The authors propose a "high-dimensional response growth curve model" with three core innovations:

  • Low-Rank Factor Model: To manage the large covariance matrix of random effects, they employ a low-rank factor model structure. This significantly reduces the number of parameters that need to be estimated, making the model more tractable.
  • Re-parameterization and Sparsity Penalty: They introduce a re-parameterization of the model coupled with an L1-type penalty. This allows for the selection of important fixed and random effect terms, effectively performing variable selection and enhancing interpretability by identifying key predictors and brain connections that change over time.
  • Computational Trick: A novel computational approach is developed that transforms the inversion of large matrices into the inversion of a stack of smaller matrices. This dramatically speeds up computation, a critical factor for high-dimensional data.

An efficient expectation-maximization (EM) algorithm is developed to estimate the model parameters, employing a two-stage approach: an initial unpenalized estimation followed by a penalized estimation stage to incorporate sparsity.

Datasets

The paper uses a longitudinal study of brain structural connectivity in association with human immunodeficiency virus (HIV) as its motivating example. This dataset includes diffusion tensor imaging (DTI) scans from 32 HIV-infected patients and 60 age-matched healthy controls over two years. Each subject has 3 to 4 time points, and brain structural connectivity is represented by a 2006-dimensional vector of fractional anisotropy measures.

The authors also conduct extensive simulations using generated data to evaluate their method's performance under various scenarios of response dimension, sample size, and noise levels.

Key Insights and Ideas

  • Bridging High-Dimensionality and Longitudinal Modeling: The core contribution is a method that effectively handles both the high dimensionality of neuroimaging data and the complexities of longitudinal changes, a significant hurdle in the field.
  • Sparsity for Interpretability: The use of sparsity penalties is crucial for identifying which brain connections and which time points are most important for understanding disease progression or development. This moves beyond simply fitting a model to identifying key biological signals.
  • Computational Efficiency is Key: The proposed computational trick is not just an optimization; it's an enabler. Without it, analyzing such large datasets would be practically impossible, opening doors for new research questions.
  • Uncovering Heterogeneous Growth Patterns: The method can distinguish between different types of growth patterns, allowing researchers to categorize brain connections based on how they change over time and how these changes differ between patient groups.

Key Results

  • Estimation Accuracy: In simulations, the proposed method achieved significantly smaller estimation errors for both the fixed effects matrix (B) and the covariance matrix (G) of random effects compared to a univariate approach.
  • Selection Accuracy: The method demonstrated high accuracy in selecting true fixed and random slopes, as well as identifying the correct rank of the underlying structure.
  • Computational Performance: For a simulation setting with r=200 and n=100, the proposed method took, on average, less than 2 minutes per data replication on a personal computer.
  • HIV Study Findings: In the HIV study application, the method identified distinct growth patterns in brain structural connectivity, highlighting differences between HIV patients and healthy controls. Specifically, it identified connections with significant differences in intercept and slope, consistent with existing literature on brain changes in HIV.

Limitations and Future Work

  • Linear Growth Curve Models: The current method primarily considers linear growth curve models. The authors suggest that for data with more time points, more flexible models like polynomial or spline-based growth curves could be explored, extending the current framework.
  • Theoretical Analysis: The authors acknowledge that a rigorous theoretical analysis of the proposed method's consistency and convergence is challenging due to the high-dimensional response and complex correlation structures. They leave this as a topic for future research, noting the lack of existing theoretical tools for such settings.

Summary updated

BibTeX
@article{wang2025high,
  author = {Lu Wang and Xiang Lyu and Lexin Li},
  title = {{High-dimensional response growth curve modeling for longitudinal neuroimaging analysis}},
  journal = {Comput. Stat. Data Anal.},
  year = {2025},
  doi = {10.1016/j.csda.2025.108239},
}
Figure