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

Area of Science:

• Neuroimaging
• Computational Biology
• Biostatistics

Background:

• Analyzing high-dimensional brain images is complex.
• Understanding spatio-temporal disease progression is crucial for neurodegenerative disorders.
• Current methods may not fully capture dynamic changes in brain structure.

Purpose of the Study:

• To introduce a novel probabilistic generative model for disentangling spatio-temporal disease trajectories from brain imaging data.
• To accurately model realistic disease progression using advanced statistical techniques.
• To identify disease-specific temporal patterns and affected brain regions.

Main Methods:

• Spatio-temporal matrix factorization with anatomically informed priors.
• Modeling temporal sources using monotonic, time-reparameterized Gaussian Processes.
• Modeling spatial sources as multi-scale sparse codes to handle non-stationarity.

Main Results:

• The model successfully disentangled spatio-temporal disease trajectories in synthetic data, outperforming standard methods.
• Applied to large-scale clinical imaging data, the model identified differential temporal progression patterns.
• Revealed disease-specific time scales correlating with clinical diagnoses and mapping key neurodegeneration regions.

Conclusions:

• The proposed model offers a powerful approach for analyzing complex brain imaging data.
• It enables the disentanglement of disease-specific progression patterns and temporal dynamics.
• This method has the potential to advance our understanding of neurodegenerative diseases and aid in clinical diagnosis.