Generative forecasting of brain activity enhances Alzheimer's classification and interpretation

TL;DR

This study focuses on multivariate time series forecasting of independent component networks derived from rs-fMRI as a form of data augmentation, using both a conventional LSTM-based model and the novel Transformer-based BrainLM model to assess their utility in AD classification.

Abstract

Understanding the relationship between cognition and intrinsic brain activity through purely data-driven approaches remains a significant challenge in neuroscience. Resting-state functional magnetic resonance imaging (rs-fMRI) offers a non-invasive method to monitor regional neural activity, providing a rich and complex spatiotemporal data structure. Deep learning has shown promise in capturing these intricate representations. However, the limited availability of large datasets, especially for disease-specific groups such as Alzheimer's Disease (AD), constrains the generalizability of deep learning models. In this study, we focus on multivariate time series forecasting of independent component networks derived from rs-fMRI as a form of data augmentation, using both a conventional LSTM-based model and the novel Transformer-based BrainLM model. We assess their utility in AD classification, demonstrating how generative forecasting enhances classification performance. Post-hoc interpretation of BrainLM reveals class-specific brain network sensitivities associated with AD.

Figures (4)

• Figure 1: The Generative Forecasting Pipeline for Classification and Interpretation. The rs-fMRI data is preprocessed and decomposed using GICA into 53 ICs with corresponding time courses. Extended scan time courses are truncated to create the baseline dataset, while regular scans with shorter lengths are replicated to align with the extended scans. The processed time series data are then segmented using a sliding window approach. Two generative forecasting models, stateless LSTM and BrainLM, are trained for forecasting. Data augmentation is performed using either the LSTM-based or BrainLM methods with replicated or truncated data. Finally, the TA-LSTM models are trained on the AD classification task to assess the effectiveness of the augmented data. A post-hoc perturbation-based interpretation analysis is conducted to examine class-level brain network sensitivity.
• Figure 2: BrainLM model architecture
• Figure 3: TA-LSTM AD Classification Result.
• Figure 4: Perturbation-based CN-AD Brain Network Sensitivity Interpretation. Sensitivity levels: Red: AD > CN; Blue: CN > AD. (MOG: Middle Occipital Gyrus; IOG: Inferior Occipital Gyrus; ACC: Anterior Cingulate Cortex; IFG: Inferior Frontal Gyrus)