Identifying discriminative EEG features of Unsuccessful and Successful stopping during the Stop Signal Task
The stop-signal task is often used to study inhibitory control. When combined with electrophysiological recordings, the N2 and P3 event-related potentials (ERPs) are regularly observed. Numerous studies link both amplitude and latency differences of the N2 and P3 to failed versus successful stopping. A slower N2-P3 complex when stopping fails has repeatedly been reported across many studies and found to correlate moderately with behavioral stopping speed. However, most studies rely on averaging across trials, thereby limiting the examination of trial-by-trial dynamics. In the present study, we employed different machine learning-approaches to classify successful from failed stop trials based on time-frequency single-trial EEG data. We also tested whether attenuating the slowing effect would alter classification performance. To preserve interpretability, we first identified five group-level EEG components time-locked to stopping and then used the time-frequency representation as features in different models. Our findings suggest that regularized logistic regression can reliably classify successful from failed stopping with an AUC = 0.72. Correcting for ERP latency differences did not markedly reduce overall classification (i.e., AUC = 0.71), but the model had to compensate by leveraging subtler, broadly distributed time-frequency features. Our feature importance measure indicated that a component closely resembling the N2-P3 complex contributed largely to the classification performance, producing a sparse model. Once the slowing effect was attenuated in the data, the model still retained predictive performance but had to rely on 15 times as many time-frequency features across the five components. Thus, it is likely that multiple overlapping processes unfold during stopping that influence response inhibition in addition to the N2-P3 complex. While the N2-P3 complex is consistently evoked during stopping and carry large discriminative ability, considering additional auxiliary processes might further our understanding into mechanisms underlying response inhibition.