Parallel GNN-LSTM Model Predicting Working Memory Involvement during Language and Emotion Processing

Working memory (WM) is a core cognitive system, strongly associated with the prefrontal cortex (PFC), crucial for task-relevant information storage and processing, serving as the main “relay” for higher cognition. Although WM-specific n-back tasks are examined, their measurable involvement in other cognitive processes remains less explored using computational models. Here, we developed and suggested a parallel GNN-LSTM model trained on WM task-fMRI data to predict its involvement in unseen language and emotion tasks. Integrating spatial and temporal information, the proposed GNN-LSTM model effectively learnt WM demand patterns and predicted WM demand when presented with non-WM task-fMRI data. As both the preliminary MLP and the GNN-LSTM models achieved over 90% accuracy, our approach was capable of generating a probability-based output corresponding to task demand and WM involvement, especially in language subtasks, math and story. This demonstration of cross-domain prediction, using WM signatures to create a “sensor” of cognitive demand alongside the model-based insights, offers a potential method for investigating cognitive resource allocation and informs the data-driven verification of WM theories.