Tool 01 — Cohort Stratifier

Discover autism subtypes.

Every subject gets a 256-dim Cortex embedding. Projected to 2D via PCA. Cluster to reveal neural subgroups.

Upload connectivity data

CSV, Excel, or .npz with per-subject connectivity rows (4,950 cols). Cortex embeds each into 256-dim latent space, PCA-projects to 2D. Color defaults to the classifier's P(ASD) ≷ 0.5.

Accepted shapes (any of):
• (N, 4950) — N subjects, pre-extracted features
• (100, 100) — single subject connectivity matrix
• (T, 100) — single subject ROI time series (T ≥ 30) — we compute connectivity for you
Sample CSV — 30 subjects (1.4 MB)Hansen 2023 (HCP-group) .npy — 100×100 matrix
Showing demo data. Upload your own .npz to get real Cortex predictions.
P(ASD) > 0.5 (40)P(ASD) ≤ 0.5 (40)

Click a point

Each dot is a subject positioned by their 256-dim Cortex embedding (PCA to 2D). Closer dots = more similar neural connectivity signatures.

What does this mean?

In plain English

You uploaded 80 brains. Cortex looked at each one's connectivity pattern — how strongly different brain regions talk to each other — and assigned every brain a position in a map where brains with similar patterns sit close together.

Of those 80, Cortex's classifier labels 40 as autism-like and 40 as typical-like (50% / 50%).

What to look for

Your cohort contains at least one cluster with strong autism/control separation. That's a useful signal: Cortex found a neural pattern that correlates with diagnosis. Next step: examine what else those cluster members share (age, sex, symptom severity) to characterize the subtype.

Important caveats

Cortex's validation accuracy is 58% — above chance (50%) but well below clinical diagnostic standards. Treat these groupings as hypothesis-generating, not confirmatory. The 256-dimensional embedding carries much richer information than the binary autism/typical label, but the 2D projection you see discards most of it. This is a research tool for exploratory analysis.

How to prepare your data

  1. Preprocess rs-fMRI through CPAC or fmriprep. Bandpass 0.01–0.1 Hz, detrend, z-score.
  2. Extract time-series with Schaefer 100-parcel atlas.
  3. Compute ROI-to-ROI correlation, Fisher-z transform, take upper triangle → 4,950-dim vector per subject.
  4. Stack into matrix X of shape (N_subjects, 4950), save as np.savez("cohort.npz", X=X).
  5. Upload above. Cortex returns embeddings, predictions, and reconstruction errors.