Interpreting neural decoding models using grouped model reliance.

Area of Science:

• Cognitive Neuroscience
• Machine Learning
• Neuroimaging

Background:

• Machine learning models are increasingly used for decoding psychological states from neural data.
• Interpreting these complex models remains a challenge, hindering the integration of data-driven and theory-driven approaches.

Purpose of the Study:

• To introduce and demonstrate 'grouped model reliance,' a model-agnostic method for interpreting machine learning models.
• To investigate individual differences in neural patterns associated with working memory load using this new interpretation technique.

Main Methods:

• Trained random forest and support vector machine models on single-trial electroencephalographic (EEG) data during a Sternberg working memory task.
• Applied the novel grouped model reliance method to assess model reliance on specific EEG features (e.g., frequency bands, regions of interest).

Main Results:

• Both machine learning models predominantly relied on alpha-band activity for decoding working memory load, confirming prior research.
• Grouped model reliance analysis revealed substantial inter-individual variability in the specific frequency bands and topographic distributions driving model predictions.

Conclusions:

• Grouped model reliance offers a valuable tool for interpreting machine learning models in neuroscience.
• Significant individual differences exist in the neural signatures of working memory, exceeding previous estimations and highlighting the need for personalized decoding approaches.