Fig. 10. Generalizability of trained models across stimulation frequencies and sessions.

a We plotted the normalized test MSE as a function of the number of stimulation frequencies used as part of the training set. We found that including more frequencies results in a more generalizable model. b However, we did not find it to be beneficial when we extend the dataset using recordings from multiple sessions. Specifically, when we considered two sessions and the three possible models that can be trained from them (using only session 1 or only session 2 or both in the training set), we found that unseen data from session 1 was best predicted by the model trained on data from Session 1 only (*p < 0.05).