Three models, one protocol, no results yet.

Reference point for every other model. Spatial filtering followed by a linear classifier, on the full montage and on each reduced configuration.

Inputs

Band-pass filtered epochs. Channel set varies by condition. Window length to be fixed at pre-registration.

Architecture

Spatial filtering (CSP-family) into a linear discriminant classifier.

Training data

Public EEG/EMG datasets, not yet selected.

Evaluation

Subject-dependent and subject-independent splits reported separately.

Ablations

Channel count sweep; EMG suppressed versus retained.

Limitations

Linear methods set a floor, not a ceiling. A weak baseline would make every later model look better than it is, so this one is tuned honestly.

Learns spatial and temporal filters directly from the epoch rather than relying on hand-designed features.

Inputs

Raw or minimally filtered multichannel epochs.

Architecture

Compact convolutional network in the EEGNet/ShallowNet family. Exact configuration to be fixed before training.

Training data

Same datasets and splits as the classical baseline.

Evaluation

Identical protocol to the baseline so the comparison means something.

Ablations

Channel count sweep; with and without EMG-dominant components.

Limitations

Deep models on small EEG datasets overfit readily and are sensitive to preprocessing choices. Subject-independent results are the ones that will be believed.

The model the study exists to test. Takes separated EEG-dominant and EMG-dominant streams as distinct inputs and learns to combine them, instead of treating one as noise in the other.

Inputs

Two streams produced by the separation stage, aligned in time.

Architecture

Per-stream encoders with a learned fusion stage. Encoder choice inherited from whichever single-stream model performs best.

Training data

Same datasets and splits as the other two models.

Evaluation

Compared against EEG-only, EMG-only and naive-combined at every channel configuration.

Ablations

Fusion removed; one stream zeroed at inference; separation quality degraded deliberately to test sensitivity.

Limitations

If separation leaks, the fusion model can appear to benefit from information the streams were supposed to isolate. Leakage checks are part of the protocol rather than an afterthought.