Machine learning provides novel neurophysiological features that predict performance to inhibit automated responses

Area of Science:

• Cognitive Neuroscience
• Neurophysiology
• Machine Learning in Neuroscience

Background:

• Event-related potentials (ERPs) are traditionally used to study cognitive processes.
• The direct link between classical ERPs and behavioral variations remains debated.
• A need exists for data-driven methods to identify robust neurophysiological predictors of behavior.

Purpose of the Study:

• To identify the best neurophysiological predictors of inter-individual performance variations using a data-driven strategy.
• To test the predictive power of established ERP components versus novel features.
• To investigate the role of different brain activity bands (theta, alpha) in response inhibition.

Main Methods:

• Utilized a dataset of 240 healthy young individuals performing a Go/Nogo task.
• Employed machine learning and source localization to analyze neurophysiological data.
• Extracted data-driven features and compared their predictive accuracy against traditional ERPs (Nogo-N2, Nogo-P3).

Main Results:

• Classical Nogo-N2 and Nogo-P3 components showed near-chance prediction levels.
• A novel feature linked to motor cortex activity predicted group membership with ~68% accuracy.
• Theta and alpha band features predicted behavioral performance up to ~78%, with theta band being most significant.

Conclusions:

• Data-driven neurophysiological features, particularly in the theta band, offer superior prediction of behavioral performance compared to classical ERPs.
• Processes occurring between Nogo-N2 and Nogo-P3, especially those involving motor cortex activity, are critical for response inhibition.
• This study highlights the potential of machine learning and source localization for discovering novel neurophysiological correlates of cognition.