Owner LaBGAS Folder secondlevel/ Guides README_ENet_neuroimaging_pipeline.md, README_ENet_plotting.md

ENet_neuroimaging_pipeline implements elastic net regularised classification for small-to-moderate neuroimaging datasets, including the common case where features outnumber subjects. Combining L1 (lasso) and L2 (ridge) penalties gives automatic feature selection while controlling overfitting — L1 drives coefficients to zero, L2 keeps correlated features from being arbitrarily discarded.

Use it when the outcome is categorical and you want a sparse solution. For a dense latent-variable solution on the same kind of outcome, use PLS-DA instead.

Inputs and outputs

Inputs

  • X — [n × p] feature matrix, subjects × features
  • Y — [n × 1] outcome vector; the maximum label becomes the positive class
  • opts.covariates — [n × nCov], with opts.covariateNames

Output — a results structure.

What the pipeline does

  1. Repeated nested cross-validation — outer CV for generalisation, inner CV for hyperparameter tuning
  2. Feature preprocessing — residualisation, then scaling, applied fold-wise
  3. Hyperparameter optimisation across the alpha and lambda grids
  4. Training and evaluation on held-out folds
  5. Permutation testing for significance
  6. Bootstrap out-of-bag confidence intervals
  7. Learning curves across subsample sizes

Residualise covariates within each fold. The usage guide is explicit about this: covariates must be residualised inside every fold. Pre-residualising on the full sample leaks information from the test fold into training and produces systematically biased performance estimates. This is the single easiest way to get a result that will not replicate.

Options and defaults

Option Default
outerK 5
innerK 4
nRepeats 50
nPerm 1000
nBoot 500
learningSteps 6
alphaGrid [0.05 0.1 0.25 0.5 0.75 0.9 1]
lambdaGrid logspace(-3, 1, 25)
scale 'zscore' — also 'center', 'none'
seed 1

alpha = 1 is pure lasso; lower values mix in ridge. The grid spans that continuum so the inner CV can choose.

Reading the results structure

Performance — results.AUC (primary), results.AUC_PR, results.ACC, results.SENS, results.SPEC, results.ACC_balanced, plus fold-level variants from results.allAUC through results.allACC_balanced.

Model — results.betaStore, results.interceptStore, results.featureWeights, results.meanFeatureWeight.

Stability — results.featureStability, results.signStability, results.selectionFrequency, results.selectionTopK. The guide gives thresholds for reading these:

Value Interpretation
> 0.8 Highly stable
0.4 – 0.8 Moderately stable
< 0.4 Unstable

With a sparse model this matters more than the weight magnitudes. Which features get selected can shift substantially between folds when predictors are correlated, and a feature that enters the model in a third of folds should not be discussed as though it were a finding.

Inference — results.permutation_p, results.permutation_p_PR; bootstrap results.AUC_CI; and results.quickCV_observed, a matched null distribution baseline.

Baselines — in-sample results.AUC_global, results.AUC_PR_global; cross-validated results.AUC_global_cv, results.AUC_PR_global_cv, using the mean or median feature. Compare against the cross-validated versions.

Learning — results.learningSizes, results.learningAUC.

Plotting

README_ENet_plotting.md covers performance distributions, permutation nulls, bootstrap intervals, selection-frequency and stability maps, and learning curves.

Next: PET Source: secondlevel/ Usage guide

Updated: