The behavioral test you pick decides the brain anatomy you can see
Newman-Norlund, Ahmadi, Warren, Salman, Yang, Willis, Guan, Bonilha & Fridriksson ·
University of South Carolina
Poster PDF ·
All four demos ·
rnorlund@mailbox.sc.edu
Nobody told this model about aphasia types. It only ever saw subscores and lesion maps. Click through and watch where it puts the damage.
Pick a step above.
It shows: a lesion map predicted from a behavioral profile, under four different ways of summarising the same patients. Set any profile you like and see the anatomy that implies.
It does not show the poster's headline result. The poster asks how alike the anatomical fingerprints of different scores are within one lens, using partial Spearman against five NMF atoms with lesion volume controlled. This demo is per-region linear regression across 70 JHU regions with no such control — a different target and a different statistic. Quantify this demo and you get a different ordering. The poster's numbers are:
| lens | mean fingerprint r |
|---|---|
| WAB severity subscores | +0.50 |
| WAB error types | +0.09 |
| PNT error types | +0.18 |
It over-draws the healthy end. Max every subscore and the model puts 15 regions above 10% damage. The 59 C-STAR patients above AQ 90 average 7.6, and the 25 highest average a median of 3 — seven of them have no region above 10% at all. A linear model with this much residual variance shrinks toward the cohort mean, so a perfect profile is drawn as far more lesion than a real high scorer carries. The readout at bottom-left compares the two live.
Limits worth knowing. Per-region R² runs 0.06–0.30: this is a group-level conditional mean, not a diagnostic for one person. Every patient in C-STAR has had a stroke, so the model has never seen a healthy brain and shrinks toward the cohort floor rather than to zero — superior longitudinal fasciculus and posterior insula are damaged in almost everyone and top every map. Demographics are covariates fit on this cohort, not causes.
Switch view mode below to see the WAB vs PNT contrast. In "PNT unique" mode, only PNT sliders move the map, WAB sliders have no effect. That dissociation is what the abstract claims.
Total predicted lesion is this profile's predicted lesion map. Use it for
the presets: Broca peaks on IFG pars opercularis and insula, Wernicke on posterior STG and
MTG, Global everywhere.
Behaviour effect subtracts each region's intercept, showing only the
departure from the cohort average. It isolates what the lens contributes, but note that it
goes blank for anyone near or better than average — Mild anomic sits above average in
no region, and Conduction in none either way. For those, read it on the Spared tab.
Absolute fixes the scale at 0–100% lesion proportion, so moving a slider changes how much colour there is. Stretch re-normalises to whatever is on screen, so the map always spans the full colormap and only the relative pattern shows — useful for comparing shapes, misleading about magnitude.
These are linear models. Dragging an error proportion to 1.0 — far past anything in C-STAR, where Mixed (PNT) tops out at 0.10 — makes them predict lesion loads above 100%. Clamping keeps every slider inside its observed range. Predictions are capped at 100% either way.
Regions predicted below this stay grey. With it at 0 the map never goes dark: a perfect WAB profile still predicts 2–5% in 36 of 70 regions, and γ<1 brightens exactly those weak values. At 10% a perfect profile leaves 15 regions coloured, the cohort median 35, the worst profile 47 — so "better scores, less lesion" becomes visible.
γ<1 brightens weak signals; γ>1 emphasizes the strongest peaks. After stretch, values span the full colormap range.
All three models include these as covariates so the dissociation is adjusted for age, sex, race, and education.