Fig. 3. Effect of sparser temporal sampling of autoregressive lags on model accuracy.

For each channel in the data, we considered two categories of models: ARX models with dense temporal sampling (i.e., iEEG lags are spaced 1 ms apart) and ARX models with sparse temporal sampling (i.e., iEEG lags are spaced τ > 1 ms apart). For the latter, the exact sampling time τ was chosen separately for each channel based on delay embedding theory using the method described in ref. ⁷⁰. We trained sparse and dense models of various lags (L ∈ 1, 50, 100, 150, 200, 250, 300) and plotted the normalized MSE across subject-channels as a function of L. Although the sparse ARX model with L lags has access to a much larger duration of iEEG history (i.e., τLms) in comparison to a dense model (i.e., Lms), the former has lower accuracy. Thus, temporally short-range models with higher sampling rates are more predictive.