Related Experiment Video
Updated: Mar 13, 2026
08:47
Functional Imaging with Reinforcement, Eyetracking, and Physiological Monitoring
Published on: November 13, 2008
11.8K
From surface to depth: Using deep learning to predict striatal fMRI reward signaling from EEG
Nadine Herzog1,2, Luka Vähäsarja1,3, Pia Reinfeld1,4,5
1Department of Neurology, Max Planck Institute for Human Cognitive & Brain Sciences, Leipzig, Germany.
Imaging Neuroscience (Cambridge, Mass.)
|March 12, 2026
Summary
Researchers developed a deep learning model to decode brain reward signals using electroencephalography (EEG). This method reconstructs ventral striatum activity, offering a more accessible way to study reward processing for potential clinical applications.
Area of Science:
- Neuroscience
- Computational Neuroscience
- Biomedical Engineering
Background:
- Reward processing is vital for motivation and decision-making, primarily involving the fronto-striatal circuit.
- The ventral striatum (VS) is central to reward processing, but its signals are typically measured using costly fMRI.
- Limited accessibility of fMRI hinders widespread clinical use for studying reward circuitry.
Purpose of the Study:
- To adapt a deep learning (DL) model to reconstruct VS blood-oxygen-level-dependent (BOLD) activity from electroencephalography (EEG) data.
- To assess the performance of the DL model in decoding subcortical reward signals compared to linear models.
- To validate the functional and anatomical specificity of the DL-derived EEG signals.
Main Methods:
- A convolutional autoencoder DL model was trained on concurrent EEG and fMRI data from 19 healthy participants during a gambling task.
- The model reconstructed VS BOLD activity from EEG signals.
- Performance was evaluated by comparing DL-derived signals with ground truth fMRI BOLD signals using correlation coefficients and anatomical specificity tests.
Main Results:
- The DL model significantly outperformed linear models, achieving a mean correlation of r=0.323 between DL-derived and actual VS BOLD signals (vs. r=0.213 for linear models).
- The reconstructed signals were anatomically specific to the VS and other reward-related regions.
- EEG frequency band analyses indicated involvement of theta and beta bands, particularly at right centroparietal, temporal, and frontal electrodes.
Conclusions:
- Deep learning models can decode subcortical reward-related signals from surface EEG, demonstrating feasibility for accessible neuroimaging.
- The DL-derived EEG signal shows functional validity, being specific to reward areas and modulated by reward conditions.
- This approach lays the groundwork for EEG-based neurofeedback systems to modulate reward circuits, potentially aiding in treating disorders linked to reward processing deficits.
