Reinforcement learning processes as forecasters of depression remission

Area of Science:

• Neuroscience
• Computational Psychiatry
• Machine Learning

Background:

• Reinforcement learning (RL) is linked to depression symptoms and illness progression.
• Understanding neurobiological markers can inform depression treatment strategies.

Purpose of the Study:

• To investigate if neural prediction error (nPE) and neural expected value (nEV) from a probabilistic learning task can predict depression remission.
• To determine if treatment or symptom severity moderates these predictions.

Main Methods:

• Support vector machines (SVM) were used to analyze blood-oxygen-level dependent (BOLD) responses.
• Participants (n=55) with depression were assessed at baseline and follow-up.
• Treatment groups included cognitive behavioral therapy (CBT) and naturalistic illness course.

Main Results:

• Both nPE and nEV classifiers significantly predicted remission.
• The nEV classifier demonstrated superior predictive accuracy compared to the nPE classifier.
• Anhedonia showed a significant interaction with nPE-predicted remission, independent of nEV.

Conclusions:

• nEV and nPE serve as relevant biobehavioral signals for depression outcomes, irrespective of treatment.
• nEV is a stronger predictor of remission than nPE.
• RL variables may enhance individualized medicine approaches for depression.