Analyzing the Effects of Caffeine on Brain Networks Using Spatiotemporal Graph Neural Network

Abstract

Caffeine is recognized for improving cognitive processes including attention and alertness. Yet, its impact on brain connections is still inadequately investigated. This this thesis examines the impact of coffee on brain connection through the use of spatiotemporal graph neural networks (STGNNs) to resting-state functional magnetic resonance imaging (fMRI) data derived from the MyConnectome dataset. We simulate functional connectivity and dynamic patterns across 116 areas of interest by developing spatial and temporal graph representations of brain activity. Three STGNN models—LSTM-GAT, LSTM-GCN, and RNN-GAT. The LSTM-GAT model attained superior performance, exhibiting an accuracy of 80%, precision of 0.70, recall of 0.75, F1-score of 0.72, and AUC of 0.80, utilizing attention processes and long-term temporal modeling to elucidate caffeine-induced connection alterations. LSTM-GCN exhibited a commendable performance (accuracy: 79.4%, AUC: 0.79), but RNN-GAT shown inferior efficacy (accuracy: 77%, AUC: 0.78). The MLP baseline had the lowest accuracy (65%, AUC: 0.66), highlighting the superiority of graph-based methodologies. These findings illustrate the effectiveness of STGNNs in interpreting intricate brain connection patterns and establish a basis for forthcoming multi-subject investigations to improve generalizability.