Understanding the dynamic nature of brain activity and its relationship to Alzheimer's Disease (AD) remains a critical challenge in neuroscience. AD is characterized by progressive cognitive decline, often preceded by mild cognitive impairment (MCI), a transitional stage where individuals may either recover or progress to dementia. Resting-state functional MRI (rs-fMRI) offers a powerful window into large-scale brain dynamics, but the high dimensionality and temporal variability of independent brain networks or dynamic functional network connectivity, coupled with limited data availability, pose significant challenges for modeling and interpretation.

This dissertation addresses these challenges through three complementary contributions. First, it introduces a time-attention long short-term memory model equipped with a novel interpretation framework to identify transiently realized connectivity patterns predictive of MCI progression or recovery. This method enables the discovery of potential meaningful biomarkers by isolating transient connectivity events associated with cognitive trajectories.

Second, a dynamic forecasting strategy is proposed to address data scarcity by generating synthetic brain activity. Using recursive forecasting, this approach augments independent brain network time courses by forecasting their future dynamics, thereby enhancing the robustness of downstream learning tasks. When combined with a transformer-based model, this strategy facilitates post-hoc interpretability and strengthens classification performance related to AD.

Third, the dissertation presents FINE (Frequency-aware Interpretable Neural Encoder), a neural network framework designed to learn multi-scale temporal and spectral representations of brain connectivity. By integrating convolutional, wavelet-based, transformer, and static encoding modules through an expert selector, FINE captures both dynamic and frequency-specific patterns. Its interpretable design leverages gradient-based saliency mapping to reveal biologically meaningful disruptions across subcortical, sensorimotor, and cerebellar networks associated with AD.

Together, these contributions advance interpretable deep learning for time series neuroimaging. By combining transient analysis, generative forecasting, and frequency-aware modeling, this work offers new insights into the neural dynamics underlying AD and MCI progression. These approaches support more robust predictive modeling and contribute to early identification and improved understanding of neurodegenerative processes.