Decoding the Neural Signature of Reward
Post by Leanna Kalinowski
The takeaway
Researchers have established a whole-brain machine-learning model that can predict the brain’s response to different levels of reward.
What's the science?
Our actions towards positive and negative outcomes are often dependent on the brain’s ability to process rewarding and punishing stimuli. Dysregulations in the brain’s reward processing system are therefore a hallmark sign of many neuropsychiatric disorders, including substance use disorders. Previous researchers have developed mathematical models that can predict how the brain responds to rewarding and punishing stimuli; however, these models often rely on the activity from single brain regions, making it difficult to generalize their findings. This week in NeuroImage, Speer and colleagues ran a series of reward tasks and developed a whole-brain machine learning model to predict the brain’s response to reward.
How did they do it?
To establish the machine learning model (referred to as the Brain Reward Signature or BRS), the researchers administered the Monetary-Incentive-Delay task to 40 participants. Each trial began with a cue phase, where participants are shown a cue that signals the monetary reward or punishment associated with the trial (potential reward of 5€, potential loss of 5€, or no monetary outcome). Following a brief delay, they were asked to press a button when a target square appeared for a limited amount of time. Depending on the trial cue and their accuracy in pressing the button, they were then informed as to whether they lost or gained money. Participants underwent 108 trials of this task, with brain activity being simultaneously measured using functional magnetic resonance imaging (fMRI).
To test the accuracy of their BRS model, the researchers then applied it to results from a publicly available dataset that used a slightly different version of the Monetary-Incentive-Delay task. In this task, everything but the cues were the same as before; instead of being shown one of three cues, the participants were shown one of five cues (potential reward of 5€, potential reward of 1€, potential loss of 5€, potential loss of 1€, or no monetary outcome).
To test whether their BRS model is specific to the neural signature of reward and not other emotional responses (i.e., disgust), the researchers then applied it to a newly developed Disgust-Delay Task. In this task, participants were asked to press a button when a target rectangle appeared and then were given feedback on whether they hit or missed the target. Then, if they hit the target, they were shown a neutral image; if not, they were shown a disgusting image. Participants completed 72 trials of this test, again with brain activity being measured using fMRI.
What did they find?
The researchers first found that their BRS model could predict the neural signature of monetary rewards versus losses in the Monetary-Incentive-Delay task with high accuracy. Brain regions critical to the BRS were often brain regions previously associated with monetary or reward-related tasks in previous studies (as identified by the NeuroSynth database). When testing the accuracy of their model on a dataset that used an expanded version of the task, the researchers found that it not only could predict monetary rewards versus losses but also could predict the magnitude of rewards and losses. When testing whether their model was specific to the neural signature of reward, the researchers found that their model could predict unsuccessful versus successful trials from the Disgust-Delay task (i.e., a non-monetary reward), but it could not predict neural differences in viewing a neutral versus disgusting image (i.e., a non-reward emotional response).
What's the impact?
This BRS model can successfully and accurately predict the magnitude of rewards and losses across different samples and tasks. Further, this model is specific to reward and not generalizable to other emotional responses (e.g., disgust). As this model is trained on full-brain responses, it is much more generalizable and reproducible than previous models that were trained on specific brain regions.