From surface to depth: Using deep learning to predict striatal fMRI reward signaling from EEG

Reward processing is critical for motivation, learning, and decision making. It involves a network centered on the fronto-striatal circuit, with the ventral striatum (VS) playing a pivotal role. While functional magnetic resonance imaging (fMRI) has been instrumental in mapping subcortical VS reward signals, its cost and limited accessibility hinder broader clinical applications. In this study, we adapted a convolutional autoencoder deep learning (DL) model to reconstruct VS blood-oxygen-level-dependent (BOLD) activity from task-based electroencephalography (EEG) data, both recorded during a two-choice gambling task known to elicit reward-related activation in the VS. The model was trained on consecutive EEG-fMRI data from 19 healthy participants, allowing it to identify patterns that generalize across individuals. Results show that the DL model significantly outperforms linear baseline models in predicting VS activity: across leave-one-out folds, the mean correlation between the DL-derived and the ground truth VS BOLD signal was $\overline{r}=0.323$ , compared with $\overline{r}=0.213$ for the linear model. Further validation confirmed that the DL-derived signal is anatomically specific to the VS and other reward-related areas, and that it is modulated by reward conditions, indicating its functional validity. EEG feature analyses revealed theta to beta frequency band involvement, particularly in right centroparietal and temporal, as well as frontal electrodes. Although the model's generalization performance was modest, these findings demonstrate the feasibility of decoding subcortical reward-related signals from surface EEG using interpretable deep learning models. This work contributes to the foundation for EEG-based neurofeedback systems aimed at modulating subcortical reward circuits, with potential clinical applications for disorders characterized by impairments in these circuits. Future improvements in model generalization may be achieved by training on larger and more diverse datasets.