Monotonic Gaussian Process for spatio-temporal disease progression modeling in brain imaging data

Clément Abi Nader, Nicholas Ayache, Philippe Robert, Marco Lorenzi, Alzheimer's Disease Neuroimaging InitiativeNeuroImage 2020 · DOI

TL;DR

This paper introduces a novel probabilistic model, Monotonic Gaussian Process Analysis (MGPA), for disentangling spatio-temporal disease progression from high-dimensional brain imaging data. By incorporating clinically-inspired priors like monotonicity and time-reparameterization, MGPA offers a more interpretable and accurate way to model diseases like Alzheimer's, outperforming standard source separation techniques.

What Problem Does This Paper Address?

Alzheimer's disease (AD) and other neurodegenerative disorders involve complex, long-term changes in the brain that are difficult to track and model. Current methods often struggle with the high dimensionality of brain imaging data, the temporal mismatch between disease onset and clinical symptoms, and the unknown timing of disease progression across individuals. Accurately modeling these spatio-temporal dynamics is crucial for understanding disease mechanisms, staging patients, and evaluating treatment efficacy. Existing approaches like Independent Component Analysis (ICA) can identify spatial patterns but often fail to capture the temporal evolution or handle the inherent variability in disease onset and progression.

Methods

The core of the proposed method, MGPA, is a Bayesian generative model that uses matrix factorization to decompose spatio-temporal data into monotonic temporal sources and corresponding spatial activation maps. Key methodological contributions include:

  • Spatio-temporal Matrix Factorization: Decomposing imaging data into temporal sources (representing disease progression over time) and spatial sources (representing brain regions affected).
  • Monotonic Gaussian Processes (GPs): Modeling temporal sources with GPs that are constrained to be monotonic. This reflects the biological assumption that disease progression generally follows a consistent direction (e.g., increasing severity or decreasing function).
  • Time-Reparameterization: Using a time-warping function to align individual patient data onto a common temporal axis, accounting for the unknown onset and speed of disease progression in each subject.
  • Multi-Scale Spatial Sources: Modeling spatial sources as sparse codes convolved at multiple resolutions to capture changes at different scales within the brain.
  • Bayesian Framework with Variational Inference: Employing a Bayesian approach to naturally incorporate prior knowledge and uncertainty, optimized using stochastic variational inference for scalability.
  • Sparsity Induction: Using Variational Dropout to encourage sparse spatial maps, helping to isolate relevant brain structures.

Datasets

The study utilizes data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) database. This includes:

  • MRI scans
  • FDG-PET scans
  • AV45-PET scans
  • ADAS13 scores (a measure of cognitive decline)

The paper also uses synthetic data to benchmark the method against standard source separation techniques like ICA, NMF, and PCA.

Key Insights and Ideas

  • Monotonicity is Key: Enforcing monotonicity on temporal progression models significantly improves their biological plausibility and interpretability, especially for diseases like AD where progression is generally unidirectional.
  • Time-Reparameterization is Crucial: The ability to infer individual time-shifts is essential for accurately modeling disease progression from cross-sectional data, as it accounts for the unknown onset and speed of the disease.
  • Multi-Scale Analysis Matters: Brain changes occur at various spatial scales. MGPA's ability to model spatial sources at multiple resolutions allows it to capture fine-grained and broader patterns of disease impact.
  • Bayesian Priors Enhance Interpretability: Incorporating clinically-informed priors (like monotonicity) into a Bayesian framework leads to more meaningful and interpretable spatio-temporal components compared to purely data-driven methods.

Key Results

  • Synthetic Data Performance: MGPA outperformed ICA, NMF, and PCA in reconstructing temporal and spatial sources on synthetic data, particularly when time-axis inference was required. It achieved lower Mean-Squared Error (MSE) for temporal sources and higher Structural Similarity Index Measure (SSIM) for spatial reconstruction.
  • ADNI Data Application: When applied to ADNI data, MGPA successfully disentangled differential temporal progression patterns across different brain regions and imaging modalities (MRI, FDG-PET, AV45-PET).
  • Clinical Plausibility: The estimated time-shifts correlated significantly with clinical measures of cognitive decline (ADAS11, FAQ, MMSE) and showed distinct distributions across different clinical stages (NL, MCI, AD), demonstrating the model's clinical relevance.
  • Model Consistency: The model's predictions for gray matter density, glucose metabolism, and amyloid load showed good agreement with raw concentration measures across different brain areas.

Limitations and Future Work

  • Identifiability: The high number of parameters in spatio-temporal models can lead to identifiability issues, which the authors address through sparsity constraints and convolutions but acknowledge as an ongoing challenge.
  • Monotonicity Constraint: While beneficial, the strict monotonicity constraint might be too restrictive for capturing the full complexity of some disease progressions.
  • Initialization Sensitivity: The optimization process was found to be sensitive to the initialization of spatial sources, a common issue in complex non-convex problems.
  • Longitudinal Data Integration: The current study primarily uses cross-sectional data. The authors suggest integrating longitudinal data for each individual to enable more specific time-reparameterization.
  • Heteroscedastic Noise: The current homoscedastic noise model might not be optimal for brain images where background areas have little relevant information.
  • Tau Protein Data: The model could benefit from integrating data measuring Tau protein concentration, a key biomarker in AD.
  • Mixed Pathologies: Future work could explore disentangling contributions from mixed pathologies (e.g., AD with Lewy bodies) in patients.
  • Validation on Diverse Cohorts: Further validation on independent, harmonized datasets is planned to demonstrate generalization.
  • Source Code Release: The authors plan to release the source code to facilitate wider use of the model.

Summary updated

BibTeX
@article{nader2020monotonic,
  author = {Clément Abi Nader and Nicholas Ayache and Philippe Robert and Marco Lorenzi and Alzheimer's Disease Neuroimaging Initiative},
  title = {{Monotonic Gaussian Process for spatio-temporal disease progression modeling in brain imaging data}},
  journal = {NeuroImage},
  year = {2020},
  doi = {10.1016/j.neuroimage.2019.116266},
}
Figure