Adaptive probabilistic regression for real-time motor excitability state prediction from human EEG

Transcranial magnetic stimulation (TMS) is a promising tool for neuromodulatory interventions in research and clinical settings, yet its effects are highly variable. Real-time EEG-TMS seeks to mitigate this variability by analyzing electroencephalography (EEG) signals to determine the optimal timing for stimulation. However, existing EEG-informed TMS methodologies are predominantly open-loop, relying on predefined brain states and disregarding immediate and long-term stimulation effects [1, 2, 3].

Here, we propose an adaptive closed-loop TMS approach that allows dynamically adjusting stimulation parameters based on real-time EEG activity. This feedback loop aims to align TMS with periods of maximum brain responsiveness and the desired changes in the stimulated network.

We developed S4EEGNet, an end-to-end deep probabilistic convolutional neural network (CNN) to predict motor excitability states from ongoing EEG signals. Our model combines depthwise and separable convolutional layers [4] with structured state space sequence layers (S4) [5], leveraging spatial, temporal, and spectral dependencies of multi-channel EEG signals.

To validate our approach for real-time online application, we streamed previously collected raw EEG-TMS data in silico. Our findings indicate that the combination of transfer learning and continual fine-tuning yields high prediction performance both within and across TMS sessions. Transfer learning stabilizes and accelerates individual model training, while continual fine-tuning enables adaptation to changes in predictive brain activity patterns over time.

Our approach holds promise to increase TMS neuromodulatory effects in brain state-dependent EEG-TMS interventions. Additionally, it offers potential for broader neuromodulation applications and closed-loop EEG paradigms, such as brain-computer interfaces.