Abstract
Deep neural networks (DNNs) excel at extracting insights from complex data across various fields, however, their application in cognitive neuroscience remains limited, largely due to the lack of approaches with interpretability. Here, we employ two different and complementary DNN models, a one-dimensional convolutional neural network (1D-CNN) and a bidirectional long short-term memory network (BiLSTM), to classify cognitive task states from fMRI data, focusing on the cognitive underpinnings of the classification. The 1D-CNN achieved an overall accuracy of 81% (Macro AUC = 0.96), while the BiLSTM reached 78% (Macro AUC = 0.95). Despite the architectural differences, both models demonstrated a robust relationship between prediction accuracy and individual cognitive performance (p < 0.05 for 1D-CNN, and p < 0.001 for BiLSTM), with lower classification accuracy observed in individuals with poorer task performance. Feature importance analysis highlighted the dominance of visual networks, suggesting that task-driven state differences are primarily encoded in visual processing. Attention and control networks also showed relatively high importance. We observed individual trait-based effects and subtle model-specific differences: 1D-CNN yielded slightly better overall performance, while BiLSTM showed better sensitivity for individual behavior. This study highlights the application of interpretable DNNs in revealing cognitive mechanisms associated with task performance and individual variability.