Generative Forecasting of Brain Activity Enhances Alzheimer's Classification and Interpretation.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Imaging

Background:

• Understanding cognition requires analyzing intrinsic brain activity from resting-state functional magnetic resonance imaging (rs-fMRI).
• Deep learning models show potential for analyzing complex rs-fMRI data but are limited by dataset size, especially for neurodegenerative diseases like Alzheimer's Disease (AD).

Purpose of the Study:

• To explore multivariate time series forecasting of rs-fMRI independent component networks for data augmentation.
• To evaluate the utility of LSTM-based and Transformer-based (BrainLM) models in AD classification.
• To demonstrate how generative forecasting improves classification performance and identify AD-specific brain network sensitivities.

Main Methods:

• Utilized resting-state functional magnetic resonance imaging (rs-fMRI) data.
• Applied multivariate time series forecasting using a Long Short-Term Memory (LSTM) network and a novel Transformer-based model (BrainLM).
• Employed generative forecasting as a data augmentation strategy for Alzheimer's Disease classification.

Main Results:

• Generative forecasting using both LSTM and BrainLM models enhanced classification performance for Alzheimer's Disease.
• The Transformer-based BrainLM model showed promise in capturing complex spatiotemporal patterns in rs-fMRI data.
• Post-hoc interpretation of BrainLM identified specific brain network sensitivities associated with Alzheimer's Disease.

Conclusions:

• Generative forecasting of rs-fMRI data is a viable strategy for augmenting datasets and improving deep learning model performance in neurodegenerative disease classification.
• The BrainLM model offers a powerful tool for both data augmentation and interpretability in neuroscience research.
• This approach holds potential for advancing the early detection and understanding of Alzheimer's Disease.