Predictive biomarkers

Train cross-validated models that predict a participant's group or a continuous phenotype from brain connectivity, using NBS-Predict-style network biomarkers with leakage-free feature selection and permutation testing.

Research question

Can a subnetwork of connectivity edges predict an out-of-sample outcome — a diagnostic group or a continuous clinical score — rather than only differ between groups on average?

Who this is for

Psychiatry and neurology groups developing connectome-based diagnostic or prognostic markers
Clinical trial teams seeking predictive endpoints (treatment response, symptom severity)
Cognitive neuroscience and neurotech labs relating networks to continuous behavioral or cognitive measures

How this differs from cohort comparison

Cohort comparison (NBS, cohort metrics) is descriptive: it identifies edges or metrics that differ between groups on average. Predictive biomarkers is predictive: it estimates how well those edges generalize to held-out subjects, which is the question that matters for a usable biomarker.

Data requirements

Data	Required	Purpose
Connectivity matrix per subject	Yes	Edge features for the model
Group labels on datasets	For classification	Two-group target (e.g. patient vs. control)
Continuous target value per subject	For regression	Phenotype to predict (e.g. cognitive score)

Minimum cohort: at least four labeled matrices; for classification, at least two per group; for regression, at least two distinct target values. Larger cohorts give more stable cross-validation estimates.

Workflow

Upload matrices for all subjects → Assign group labels (or enter targets) → Cohort analysis → Prediction tab → Run prediction → Interpret

Step 1 — Upload cohort datasets

Create a study and upload connectivity matrices for all participants (precomputed connectomes, fMRI functional connectivity, or EEG-derived PLV matrices).

Step 2 — Define the target

Open Cohort analysis (/areas/neurology/studies/{id}/cohort) and the Prediction tab:

Classification — assign a group label to each matrix dataset in the assignment table (reused from cohort comparison; labels persist on the dataset).
Regression — enter a continuous target value per matrix dataset in the target table.

Step 3 — Choose the model and run

Pick an estimator and run the prediction:

Classification — logistic regression (default), random forest, or elastic net.
Regression — ridge (default), random forest, or elastic net.

Inside every cross-validation fold, an NBS-style suprathreshold edge selection runs on the training fold only, so the held-out fold never informs feature selection (no leakage). A permutation test shuffles the target to establish a null performance distribution.

Step 4 — Read the results

Classification — ROC AUC, balanced accuracy, sensitivity/specificity, and an out-of-fold confusion matrix.
Regression — R², MAE, Pearson r, and an observed-vs-predicted scatter.
Both — a permutation p-value, an edge-stability map (how often each edge was selected across folds), and a subject-level prediction table.

Step 5 — Interpret

Use Interpret on the completed prediction run. The grounded narrative cites only model metrics, the most stable edges, and the permutation result — and is explicit that this is biomarker discovery, not a validated diagnostic.

Expected outputs

Out-of-sample performance (AUC or R²) with a permutation p-value
Edge-stability map identifying robust biomarker-candidate connections
Subject-level predictions (probabilities or predicted values) and residuals
AI narrative citing model performance and stable edges
Run record with parameters and pipeline version for reproducibility

Example insight

A frontolimbic subnetwork classifies patients from controls at AUC 0.82 (permutation p = 0.01), with the most stable edges concentrated in default-mode hubs — a candidate connectivity biomarker that warrants external replication.

Typical analyses

Analysis	Target	Question
Diagnostic biomarker	Patient vs. control	Do connectivity edges classify diagnosis out-of-sample?
Treatment response	Responder vs. non-responder	Can baseline networks predict who responds?
Symptom severity	Continuous clinical score	Does a subnetwork track severity (regression R²)?
Cognitive prediction	Continuous cognitive score	Do connectivity edges predict cognitive performance?