Preprocessing Pipeline

Prior to data entering the brain foundation model, raw sensor signals must be transformed into a standardised, denoised version of the data. This pipeline is characterized by the following stages, which are modular and replaceable:

1. Stage 1: Artifact Rejection and Denoising
2. Stage 2: Bandpass Filtering
3. Stage 3: Windowing and Normalisation

Artifact Rejection and Denoising

Raw physiological signals are corrupted by a variety of artifacts:

EEG artifacts:

• Ocular artifacts: eye blinks and saccades produce large slow potentials at frontal electrodes. Removed via ICA or regression-based eye-movement correction.
• Muscle artifacts: electromyographic noise from jaw clenching or facial movements. Addressed by high-frequency filtering and ICA.
• Motion artifacts: electrode movement during physical activity. Particularly relevant for WAUC.
• Channel dropouts: individual electrodes with poor contact produce flat or saturated signals. Detected and interpolated from neighbouring channels.

PPG artifacts:

• Motion artifacts from wrist movement dominate PPG noise. Accelerometry data from the same wristband is typically used in adaptive filtering to subtract motion-correlated components.

Bandpass Filtering

EEG signals are filtered to retain clinically and cognitively meaningful frequency bands:

Windowing and Normalisation

Windowing

Continuous recordings are segmented into overlapping windows. Window length governs the trade-off between low-frequency frequency resolution and cognitive state stationarity. Typical ranges: 2–10 seconds with 50% overlap.

Normalisation

Each window is independently normalised to remove slow drift and inter-subject amplitude differences:

• Per-channel z-score normalisation within each window.
• Robust normalisation using median absolute deviation to resist outlier contamination.
• Riemannian covariance-matrix normalisation, which is inherently scale-invariant and hardware-agnostic.