Reproducible, tested, and honest about its limits
A research-grade pipeline — not a peer-reviewed study or a clinical tool.
Pipeline
MNE preprocessing on 31 shared channels (1–40 Hz band-pass, per-dataset notch, 250 Hz resample, average reference, 2 s epochs, amplitude rejection) → interpretable spectral and complexity features → one region-level vector per participant → participant-grouped cross-validation with no epoch leakage → a single frozen external test → participant-level bootstrap intervals.
Limitations
Small external cohort (n=30, wide intervals); ~4:1 class imbalance and an age/sex confound in the development cohort; quality-control exclusions are Parkinson's-concentrated; and dataset shift dominates the feature space. Preprocessing and feature choices are documented defaults, not exhaustively validated, and no EEG-domain expert has reviewed them.
Reproducibility
Fixed seeds, a locked uv environment, config-driven parameters, decision records in docs/decisions/, and a green CI pipeline. Every number on this site traces to executed code — see the source and full methods on GitHub.