DFINE for Nonlinear Modeling of Human iEEG Activity

This paper introduces an extension of the DFINE framework for modeling human intracranial electroencephalography (iEEG) recordings. It addresses the limitations of linear dynamical models in capturing the nonlinear structure of neural activity and the inference challenges of recurrent neural networks when dealing with missing data, a common issue in brain-computer interfaces (BCIs). The study demonstrates that DFINE outperforms linear state-space models in forecasting future neural activity and matches or exceeds the accuracy of a GRU model, while also handling missing observations more robustly. This work is significant because it provides a flexible and accurate framework for modeling iEEG dynamics, with potential applications in next-generation BCIs.

• •DFINE is a deep learning framework that integrates neural networks with linear state-space models.
• •DFINE is extended for modeling multisite human intracranial electroencephalography (iEEG) recordings.
• •DFINE outperforms linear state-space models in forecasting neural activity.
• •DFINE handles missing observations more robustly than baseline models.
• •DFINE's advantage is more pronounced in high gamma spectral bands.