Abstracts

Rationale: In recent years, it has become increasingly clear that “ultra-long term” (i.e., multi-month) electroencephalogram (EEG) recordings can improve epilepsy diagnosis and can trigger interventions (e.g., neurostimulation) to prevent seizures or mitigate their effects.  Low-channel-count wearable or subcutaneous EEG devices can meet this need. However, analyzing such data comes with some novel challenges. In particular, the conventional spatial filtering techniques used to remove EEG artifacts (e.g., blink potentials, EMG) with conventional 20+ channel EEG montages will not work well with the limited number of channels typical of these devices. Frequency band-based filters are also ineffective because these artifacts are not periodic. Here we explore an alternative approach, temporal filtering via deep neural networks (DNNs) for single channel EEG artifact correction. DNNs are capable of learning very complicated filters from large amounts of training data and should be able to learn to model and remove the waveforms typical of EEG artifacts. As a proof-of-concept, we have attempted to correct for blink potentials due to their prevalence and short duration. Methods: We collected EEG data from eight neuronormal undergraduates at 30 electrodes while they performed a target detection task. Conventional spatial filters were used to estimate blink artifacts in these data and an equal number of blink-contaminated and blink-free one second clips of data were sampled from the data. Subsequently, a DNN was trained to remove blink artifacts from data at a channel in the center of the forehead using the spatial filter output as a teaching signal. The DNN’s performance was then compared with that of the spatial filter using held out data Results: On average, the power spectrum density (PSD) of the blink contaminated EEG data show a dramatic 8.1 dBuV²/Hz increase relative to the spatially filtered EEG data from 0-12 Hz. The PSD of the DNN filtered data very closely matches that of the spatially filtered data (Mean absolute deviation of 0.45 dBuV²/Hz and 0.27 dBuV²/Hz in the 0-50 Hz band for blink-contaminated and blink-free trials respectively). Inspection of raw waveforms shows that the DNN is generally quite good at capturing the blink artifact. However, because of the large magnitude of the artifact, small filtering errors can still cause significant EEG mismatch and the DNN filter captures approximately 43% of the EEG data variance. Conclusions: Our results demonstrate the feasibility of DNN temporal filters for EEG artifact removal, which could potentially improve the utility of data acquired from limited channel wearable or subcutaneous EEG devices. However, additional work is needed to generalize these results to a greater variety of artifacts, to a greater variety of electrode locations, and on data containing epileptiform activity to confirm the usefulness of this technique. Funding: We gratefully acknowledge the support of the NVIDIA Corporation’s donation of the Titan X Pascal GPU used for this research and funding from the Canadian Institutes of Health Research.