ABSTRACT: The focus of this supplement request is to leverage and reinforce our ongoing biomarker identification work with methods specifically focusing on Alzheimer's disease (AD) and related disorders (ADRD). Deep learning methods that we are developing in the parent grant can produce an optimal performance based on learning end-to-end directly from the data. Our goal is to leverage models trained to classify AD from the full brain fMRI dynamics for capturing novel dynamic biomarkers of AD via trained model introspection. However, it is notoriously difficult to train models to predict directly from full brain dynamics without prior dimensionality reduction. To overcome this difficulty, we will develop self-supervised approaches that would take advantage of unrelated datasets and provide a performance boost that would allow obtaining reliable classification improvements even on small data. This improved classification directly transfers into more reliable introspection of why the model classifies subjects to AD. We plan to improve the robustness of these predictive/introspective methods and study these full-brain fMRI dynamic measures in younger adults who have CSF risk markers assessed for AD in order to evaluate the potential for leveraging such models as biomarkers of AD.