Thesis acceptance

modewise generalises the analysis pipeline of the author’s master’s thesis. This document checks how far modewise reproduces that pipeline. Where modewise deliberately differs, it gives the sourced reason.

The test is tests/acceptance/test_thesis.py. It is opt-in and runs on local data: the thesis’s own outputs, which are not public in this form, and a ds005237 checkout (see the test’s docstring). The cohort is the 14 subjects for which thesis outputs exist.

What was rebuilt

Thesis stage	modewise	Where
Parcellation (Schaefer 400, 17 networks, plus Tian scale II; 432 channels, z-scored)	the thesis’s parcellated series as input, named by the same atlas label files	[channels] with schaefer and tian tables
Subject selection (SHAPS and TEPS anhedonia rules)	examples/tcp/make_participants.py	[participants]
MVMD (K = 10, α = 1000, exponential init, tolerance 1e-8, τ = 1e-3)	mvmd with the same settings	examples/tcp/thesis.toml
HHT (Hilbert spectrum of the modes)	hilbert → hilbert_spectrum, 0.01–0.25 Hz, 224 log bins	same
CWT (Morlet, 224 frequencies, 0.01–0.25 Hz)	cwt on the same range and count	same
Face and shape blocks	epoch	same
FC per block (Pearson, per mode)	functional_connectivity per mode and block	same
ROI selection (PFCv, PFCm, amygdala)	select, after whole-brain MVMD (the thesis default)	same
Group FC statistics (planned in the thesis code, never finished)	test_paired, test_groups	docs/statistics.md
DenseNet features and kNN / random forest	not in modewise: FeatureTable and stack hand over to ML code	docs/features.md

Where modewise reproduces the thesis exactly

Check	Result
Whole-run MVMD on resting-state runs, all 14 subjects	identical modes: correlation 1.0 for every mode of every subject
Centre-frequency labels of those modes	within 0.00036 Hz. The modes are identical; the labels differ by one update step (which iteration’s estimate is reported), far below the 0.0026 Hz frequency resolution of a 488-sample run.
Per-block MVMD on the thesis’s own blocks	identical modes and centre frequencies
Pearson FC of the thesis’s own blocks	identical (largest difference 1e-15)
Fisher z of those blocks	identical once the thesis’s ±0.9999 clip is applied
Block count and onsets (hammerAP)	18 and 18, same onsets
Subject grouping of the cohort	identical (7 anhedonic, 7 control)

Where modewise differs, and why

Topic	Thesis	modewise	Reason
Decomposition scope	the stored task outputs decompose each block separately; the thesis code was later changed to decompose the run and slice	decompose the full run, then cut	docs/epochs.md: a 23-sample block cannot resolve slow-5 or slow-4. The thesis code’s own later fix agrees.
Haemodynamic lag	none	explicit (lag_s = 5.0 in the example; required)	docs/epochs.md
Block length	from onset to the next event, about 17.9 s (23–24 samples)	the merged events’ own duration, 16.8–17.6 s	events.tsv durations; the thesis blocks include the following cue
Fisher z	r clipped to ±0.9999 first	no clip; diagonal undefined	docs/connectivity.md: no source supports the clip
Instantaneous frequency	backward phase difference	centred difference	docs/time-frequency.md: no half-sample lag
Envelope features	log1p, then divided by each channel’s maximum	raw amplitude	feature scaling for a network, not part of the method (docs/time-frequency.md)
CWT	scirs2 Morlet with scale = ω0/(2πf·dt), no Fourier factor; scalograms divided by their maximum	Torrence & Compo with the Fourier factor (λ ≈ 1.033), checked against pycwt; unscaled power	docs/time-frequency.md. Patterns agree only loosely (median channel correlation ≈ 0.36), because the frequency labels shift by about 3% and the scaling differs.
TEPS anhedonia rule	the missing-value codes −300, −900 and 99 averaged as answers	same in thesis_group; corrected_group treats them as unanswered	the NDA definitions file marks them as codes. Cohort-wide, the correction moves 8 subjects between groups; none of the 14 here.

Stored thesis outputs that could not be reproduced

• Task HHT envelope and instantaneous frequency do not follow from the stored block modes by the thesis’s own formula. Checked:

  • the scipy-convention analytic signal (the thesis used scirs2’s copy of it)

  • the backward difference

  • log1p with per-channel max-divide

  • mode reordering

  • a full-run decomposition sliced to the block

  Correlations stay at 0.4–0.8. These groups lack the decomposition_scope attribute, which the thesis code itself treats as stale, so they were most likely written by a different code version from the stored modes.

• Resting-state HHT follows from the stored whole-run modes only approximately: envelope r ≥ 0.94 and frequency r ≥ 0.83 across subjects.

Neither affects modewise: its Hilbert transform is checked independently against scipy and the emd package (docs/time-frequency.md).

Results on the cohort

From whole-run MVMD, ROI selection, blocks with a 5 s lag, and Pearson FC per mode (3780 edge features, 2646 of them resolvable in a block):

Test	Result
Face vs shape, paired, 14 subjects; all 16384 sign flips enumerated (exact)	no edge survives FWER (smallest p = 0.73) or FDR (q < 0.05: none)
Anhedonic vs control, face blocks, 7 vs 7, 5000 permutations	no edge survives FWER (smallest p = 0.71) or FDR (none)

With 14 subjects these tests have little power, so the null results say nothing about the thesis’s classification findings, which used other features (DenseNet on spectra). They show that the full chain, from parcellated series to corrected group statistics, runs from one pipeline file and one dataset spec.

Not reproduced

• DenseNet feature extraction and kNN / random forest classification: these sit outside modewise by design. The FeatureTable and Results.stack handovers provide their inputs.

• Parcellation from volumes: the thesis’s parcellated series is the input. modewise can parcellate volumes (docs/volumes.md, matched against nilearn), but rebuilding from the fMRIPrep runs needs gigabytes of downloads and is not part of this test.