Improving closed-loop TMS timing using the Wavenet model

Recently, it has become evident that the excitability state of the neuronal network during external stimulation, such as TMS, has a major impact on the stimulation efficacy. By triggering TMS based on real-time EEG, we can target phases of ongoing oscillations that are associated with high excitability states. However, precise timing can be challenging due to rapidly changing brain states and the time consumed in algorithmic decision making. This can be solved by forward predicting the EEG time series. Until now, only the autoregressive model (AR) has been used for this purpose. Here, we propose the deep learning model Wavenet to improve the timing accuracy. The Wavenet model consists of dilated causal convolution layers, as well as residual and skip connections between them. Through a softmax layer, the model outputs a conditional probability distribution over 256 possible values for the next timestep. We trained the model on band-pass-filtered (8–12 Hz) eyes-open resting-state EEG data from the C3 channel. In total, 180 minutes of data from 60 healthy subjects were used. We then tested the model's capacity of detecting peaks occurring 50 ms ahead in the EEG signal, and compared these results with those of an order 30 Yule-Walker AR model. This was done by predicting EEG time series 100 ms into the future with both models and determining the phase in the middle of the prediction. Figure 1b shows the distribution of the predicted phases in such instances where a peak occurred in the measured signal. Wavenet's phase estimation accuracy was significantly better (means at high and low peaks 0±23 and 176±31) compared with AR (7±58 and 187±59). These findings imply that a deep learning model for TMS timing can significantly improve the accuracy of stimulation targeting and thus the controllability and effectiveness of TMS. Keywords: closed-loop stimulation, TMS-EEG, deep learning, signal forecasting

Abstract licence: CC BY-NC

A person counts once a day per paper. We keep no cookie and no address: a reader is a hash that changes every day. Any month under five reads shows no number.