Capstone — Spectral fingerprint

Brief

For all subjects in the ds-eegbci baselines (R01 eyes-open, R02 eyes-closed) — or a ds-lemon raw-release subset, or ds-dortmund if LEMON is dropped (§13 items 5 and 17) — compute per-subject PSDs with documented Welch settings, extract individual alpha frequency and the aperiodic exponent, report group distributions and eyes-open versus eyes-closed differences, and justify your filter choices in one paragraph.

Do not use the LEMON preprocessed release for this: it is band-passed 1–45 Hz and unsuitable for aperiodic fits.

Inputs

• ds-eegbci baselines for the documented subset (see the subset rule below), or the alternatives above.
• Lessons L1.3 (Welch), L1.4 (windows), L1.5 (filters), L1.6 (line noise) and L1.7 (spectral parameterization).
• The list of documented defective ds-eegbci subjects (S088, S089, S092, S100; S038 and S104 are also often dropped).

Subset rule and runtime

Capstone notebooks run on a documented subset within the §11 limit of ten minutes on Colab’s free tier, and scale to the full cohort locally:

• The documented subset for C1 is 10–20 ds-eegbci subjects, listed by ID in the notebook’s first cells, chosen to exclude the documented defective subjects unless the notebook is demonstrating that they are harmless for this analysis.
• A FULL_COHORT switch at the top of the notebook (default False) extends the run to all subjects for local execution; the per-subject function is identical, so the figure panel and the distributions simply grow.
• Per-subject PSDs and specparam fits for the full cohort may be precomputed by data/scripts/ and shipped as assets (the w-aperiodic-explorer asset is one such product) so that the full-cohort figures can be reproduced without the full download.
• Every Welch and specparam parameter is stated in one place in the notebook, the notebook downloads only the runs it needs, uses _shared/helpers.py, contains no absolute paths, and prints in its final cell the numbers the rubric asks for (per-subject IAF and exponent, group medians, the EO–EC difference, and the flagged subject with its reason).

Deliverables

1. Notebook nb-c1-spectral-fingerprint with every Welch parameter (segment length, overlap, window, frequency range) and every specparam setting (fit range, aperiodic mode, peak limits) stated in one place.
2. A figure panel: per-subject PSDs (log power, labelled units), the distribution of IAF and of the aperiodic exponent across subjects, and the eyes-open versus eyes-closed difference.
3. A methods paragraph justifying the filter choices (and stating whether any filtering was applied before the PSD at all, and why the line noise was or was not removed).

Rubric

• Correct units (µV²/Hz; log axes labelled as such).
• Welch and specparam parameters stated and justified.
• Peak vs slope distinguished: the eyes-closed effect is described as a peak change, a slope change, or both, with evidence from the fits.
• At least one subject flagged as anomalous with a reason.
• The documented defective ds-eegbci subjects handled explicitly — excluded with reason, or shown to be harmless for this analysis.

Example report structure

1. Data and subset (which runs, which subjects, why; the FULL_COHORT state of the run being reported).
2. Spectral estimation (parameters; the resolution–variance trade-off you accepted; the hardware fingerprint of the recording).
3. Parameterization (fit range, settings, goodness-of-fit summary).
4. Results (figure panel; group distributions; EO vs EC).
5. Anomalies (who, what, how you decided).
6. Methods paragraph.

Estimated time: TODO(confirm) — the minutes value in the frontmatter is a site-design estimate, not from the spec.