Model-based feature construction for multivariate decoding.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning in Neuroscience

Background:

• Conventional neuroscience decoding methods struggle with high-dimensional neural data and lack interpretable results.
• Predicting discrete brain states from neural activity presents challenges in feature selection and interpretation of accuracy estimates.

Purpose of the Study:

• To propose a novel model-based decoding approach addressing challenges in feature selection and interpretability in neuroscience.
• To provide a principled dimensionality reduction and mechanistically meaningful interpretation of decoding results.

Main Methods:

• Inverting a dynamic causal model (DCM) of neurophysiological data on a trial-by-trial basis.
• Training and testing a discriminative classifier on a reduced feature space from model parameter estimates.
• Reconstructing the separating hyperplane for enhanced interpretability.

Main Results:

• The proposed method achieves significant above-chance decoding performance.
• The approach allows for a neurobiologically meaningful interpretation of decoding results.
• Demonstrated application using DCM of electrophysiological recordings in rodents.

Conclusions:

• The model-based decoding approach offers a principled solution for dimensionality reduction and enhances the interpretability of neural decoding.
• This method can be applied to various modeling approaches and brain data, supporting trial or subject label decoding.
• It supplements existing model-based decoding techniques and aids in structural model selection.