Case studies#
Analyses of real data that reproduce published findings with modewise. They need data that is too large to download on every build, so they run by hand; each report is stamped with the modewise version and commit that made it, and a test flags reports made by an older version.
To rerun one:
python examples/case-studies/fetch_tcp.py restAP restPA hammerAP # OpenNeuro ds005237, checked against the bucket's MD5s
python examples/case-studies/frequency-amplitude/analysis.py
python examples/case-studies/faces-amygdala/analysis.py
python examples/case-studies/fingerprinting/analysis.py # about 3 GB of results
Frequency-specific amplitude (Zuo et al. 2010)#
Resting BOLD fluctuates mostly below 0.1 Hz. Zuo et al. split the classic 0.01–0.08 Hz range into
slow-5 (0.01–0.027 Hz) and slow-4 (0.027–0.073 Hz) and found the balance differs by region:
slow-5 dominates cortex, most of all the default-mode network, while the basal ganglia carry
relatively more slow-4. This case study repeats the comparison on the first resting-state run of
241 subjects of the Transdiagnostic Connectome Project, with spectrum, band_amplitude and
test_paired, and checks it with an independent method (the MODWT MRA). The files are in
examples/case-studies/frequency-amplitude/.
What it shows: the regional pattern replicates, and both methods agree across regions (r = 0.99). One difference: Zuo et al. found slow-4 larger than slow-5 in the basal ganglia; here the two are about level there, with slow-5 larger everywhere else.
241 subjects of TCP (ds005237), first restAP run. Method: pipeline.toml and analysis.py.
Slow-4 vs slow-5 ALFF per region (paired sign-flip test, FWER over 434 regions)#
slow-5 > slow-4: 428 regions
slow-4 > slow-5: 0 regions
no FWER-significant difference: 6 regions
region |
t (slow4 − slow5) |
FWER p |
|---|---|---|
pGP-lh |
-1.0 |
1.000 |
NAc-core-lh |
-1.3 |
1.000 |
NAc-core-rh |
-1.4 |
1.000 |
aGP-lh |
-2.6 |
0.447 |
pGP-rh |
-3.1 |
0.166 |
aGP-rh |
-3.1 |
0.163 |
THA-DP-lh |
-3.7 |
0.028 |
lAMY-rh |
-3.8 |
0.020 |
log(slow-4 / slow-5 ALFF) by structure (mean ± SE over subjects)#
structure |
regions |
ALFF ratio |
|---|---|---|
cortex |
400 |
-0.329 ± 0.008 |
striatum |
16 |
-0.071 ± 0.008 |
thalamus |
8 |
-0.101 ± 0.011 |
medial_temporal_lobe |
8 |
-0.144 ± 0.011 |
cerebellum |
2 |
-0.387 ± 0.020 |
Striatum − cortex: +0.258 ± 0.008, positive in 99% of subjects; paired t = 33.1, permutation p = 0.0001.
Second method: MODWT MRA (sym4), log(SD of D4 / SD of D6)#
Striatum − cortex: +0.351 ± 0.009, positive in 99% of subjects; paired t = 37.2, permutation p = 0.0001. Across regions, the wavelet ratio correlates with the ALFF ratio at r = 0.99.
Cortical networks, most slow-5 first#
DefaultA -0.38 · DefaultB -0.37 · DorsAttnA -0.37 · VisCent -0.37 · ContA -0.35 · VisPeri -0.35 · DorsAttnB -0.35 · ContB -0.35 · SalVentAttnB -0.35 · TempPar -0.34 · SalVentAttnA -0.32 · ContC -0.31 · SomMotA -0.29 · DefaultC -0.29 · LimbicB -0.27 · SomMotB -0.26 · LimbicA -0.22
Faces against shapes in the amygdala (Hariri et al. 2002)#
In the emotional face-matching task, people match faces with fearful or angry expressions, or
match shapes. Faces drive the amygdala more than shapes, one of the most replicated results in
affective neuroimaging, along with the ventral visual stream that processes faces. This case
study cuts each TCP subject’s face-matching run into face and shape blocks with a 5 s
haemodynamic lag (epoch), averages each block (the mean feature), and tests faces against
shapes in every region with test_paired. Exploratory follow-ups ask whether the amygdala
response tracks anhedonia (SHAPS) or differs between patients and the general population, using a
FeatureTable built from the per-subject contrasts. The files are in
examples/case-studies/faces-amygdala/.
What it shows: all four amygdala parcels respond more to faces (t = 12.7 to 25.9, FWER p < 0.001), and the strongest responses are in extrastriate visual cortex. The exploratory tests find no relation with anhedonia and no group difference.
219 subjects of TCP (ds005237), hammerAP run. Method: pipeline.toml and analysis.py.
Not analysed: 7 runs that have parcellated data but no BIDS events or sidecar on OpenNeuro, so their blocks cannot be defined.
Faces − shapes per region (paired sign-flip test, FWER over 434 regions)#
faces > shapes: 84 regions
shapes > faces: 250 regions
region |
t (face − shape) |
FWER p |
|---|---|---|
lAMY-lh |
+12.7 |
0.0001 |
mAMY-lh |
+22.6 |
0.0001 |
lAMY-rh |
+17.3 |
0.0001 |
mAMY-rh |
+25.9 |
0.0001 |
Strongest faces > shapes:
region |
t |
FWER p |
|---|---|---|
17networks_RH_VisCent_ExStr_9 |
+41.6 |
0.0001 |
17networks_RH_VisCent_ExStr_10 |
+38.9 |
0.0001 |
17networks_LH_VisCent_ExStr_10 |
+37.3 |
0.0001 |
17networks_LH_VisCent_ExStr_9 |
+36.1 |
0.0001 |
17networks_RH_VisCent_ExStr_5 |
+35.4 |
0.0001 |
17networks_LH_VisCent_ExStr_5 |
+33.2 |
0.0001 |
17networks_RH_VisCent_ExStr_7 |
+31.6 |
0.0001 |
17networks_RH_VisCent_ExStr_4 |
+30.5 |
0.0001 |
17networks_RH_VisCent_ExStr_6 |
+28.9 |
0.0001 |
17networks_RH_DorsAttnA_TempOcc_5 |
+28.8 |
0.0001 |
Exploratory: the amygdala response (mean of four amygdala parcels, face − shape)#
Mean response: +23.0 (SD 13.6, in the TCP files’ signal units), positive in 98% of subjects.
Against SHAPS anhedonia (age and sex as covariates): t = -1.24, permutation p = 0.226, 203 subjects (16 without a score).
Patients against the general population (age and sex as covariates): t = +0.62, permutation p = 0.542, 219 subjects.
Notes#
The TCP series had the global signal regressed out. That shifts every region’s signal by the brain-wide mean, so regions with no face response tend to show shapes > faces; the count above reflects this, not a shape response.
Blocks are averages of the BOLD signal 5 s after each block’s events (a fixed lag), not a GLM with a haemodynamic response function; for a contrast of long blocks the two agree closely.
Connectome fingerprinting (Finn et al. 2015)#
Finn et al. showed that functional connectivity identifies a person: correlate the connectome
of one scan with every connectome from another session, and the best match is usually the same
person. This case study matches each TCP subject’s first restAP run against everyone’s first
restPA run (and the reverse), for the full band and for two MODWT octaves. The modes and band
signals are computed with keep = false, so only the connectivity matrices are stored. The files
are in examples/case-studies/fingerprinting/.
What it shows: about 80% of 241 subjects are identified in the full band, against a chance level of 0.4% (the best of 1000 shuffles reaches 2.1%). The 0.039–0.078 Hz octave identifies more subjects than the 0.010–0.020 Hz octave, which a 6.5-minute run samples with only a few cycles.
241 subjects of TCP (ds005237): first restAP run against first restPA run. Chance is 1/241 = 0.4%. Method: pipeline.toml and analysis.py.
band |
restAP → restPA |
restPA → restAP |
best chance-level accuracy in 1000 shuffles |
|---|---|---|---|
full band |
80.9% |
78.8% |
2.1% |
D6, about 0.010–0.020 Hz |
31.5% |
35.3% |
2.1% |
D4, about 0.039–0.078 Hz |
66.8% |
60.6% |
2.1% |
Notes#
Identification correlates Fisher-z edge vectors (434 regions, 93,961 edges); a target is identified when its best match is the same subject, as in Finn et al. (2015).
Finn et al. reported 92.9% and 94.4% between two 14-minute resting runs recorded on different days, with 268 regions and no global signal regression. Here the runs are 6.5 minutes from one session with opposite phase encoding (AP and PA), with 434 regions and global signal regression, so the numbers are not directly comparable.
A 6.5-minute run holds only about 4 to 8 cycles at 0.010–0.020 Hz, so connectivity in that band is estimated from little data; its lower accuracy says as much about estimation as about how individual the band is.