Alzheimer's classification is essential to diagnosis and monitor a disease which severely impacts cognitive function. This thesis compares three representations of functional magnetic resonance imaging data, including correlation networks, causal relationships, and transformer derived patterns, and their efficacy for binary classification. Correlation networks are derived using sparse inverse covariance matrices; causal networks are derived using the Granger causality method. Standard preprocessing is run on a dataset of thirty images. The study evaluates the use of support vector machines with correlation and causal features versus two transformer architectures. The support vector machine with filtered causal features performed best with an area under the curve score of 0.8, with connections in the default mode network, hippocampus, and orbitofrontal lobes having the most predictive power. Transformers performed well given the limited data, but further study is needed on a larger dataset to assess the effectiveness of the proposed spatiotemporal architecture. These findings replicate clinical findings of vulnerable regions in early stages of Alzheimer's, and hyperconnectivity in Alzheimer's patients as brain regions try to compensate for neural degeneration.