Comparison Between Performance of ANN-based Models and AR in Predicting EEG-triggered-TMS Time series

Closed-loop brain stimulation as an emerged concept in brain stimulation techniques includes a crucial prediction block to predict the upcoming activity of the brain. In this paper, we propose to evaluate different artificial neural network (ANN) structure's ability in predicting neural signal in the presence of non-invasive stimulation. For this purpose, we implemented a brain stimulation experiment using transcranial magnetic stimulation (TMS) technique with a concurrent recording of electroencephalography (EEG) signal from the brain activity. Using this EEG-triggered-TMS signal, the prediction performance of ANN-based networks including multilayer perceptron (MLP), recurrent neural network (RNN), and a long short-term memory (LSTM) neural network with the previously used in literature, linear autoregressive (AR) model was compared. Results in both one-step-ahead prediction and multi-step-ahead prediction demonstrated that the AR model outperforms ANN-based structures and the MLP has relatively better performance among other ANNs. This comparison as our preliminary step toward using ANNs in predicting neural dynamics suggests that in this level of analysis, linear models can sufficiently capture the neural dynamics and ANN does not make an improvement.