Sleep EEG is frequently contaminated by sweat artifacts, which arise from changes in skin conductance caused by thermoregulatory sweating. These artifacts are most prominent during slow-wave sleep, particulrly during and following the second sleep cycle, when sweating increases and electrode-skin impedance drifts slowly over time.

The resulting signal manifests as large-amplitude, low-frequency fluctuations that overlap spectrally with physiologically meaningful delta activity and ocular signals.

Conventional high-pass filtering used to suppress sweat artifacts therefore also attenuates true delta power and EOG components due to this frequency overlap, leading to distortion of sleep-relevant features.

Throughout the development of the Hypnodyne Autoscorer, we systematically indexed and characterized dozens of EEG artifact types and developed proprietary vision-based elastic models and statistical algorithms to capture their diverse morphologies.

Building on this foundation, we leveraged the latest machine learning technologies to train a deep neural network with a large temporal receptive field - spanning multiple epochs and optimized for the ZMax montage - featuring a hierarchical, multi-resolution representation with progressive context aggregation and feature refinement.

This architecture integrates both local and global temporal cues through successive downsampling and reconstruction stages, enabling precise separation of EEG features that occupy overlapping frequency bands.

Much like a human scorer, who often needs to review neighboring epochs and adjacent signal segments to determine whether a low-frequency oscillation originates from sweat or neural activity, the model leverages contextual information from surrounding regions of the signal to make the same distinction.

As a result, the model effectively removes the majority of sweat-related band power from the EEG signal while preserving genuine neural activity.