Z Normalization was applied to all the features to maintain a common scale, without distorting differences in the ranges of values. Regularization methods were used to control for variables overfitting, enhancing the interpretability, and prediction accuracy of the calculated models. We used the least absolute shrinkage and selection operator (LASSO) penalty approach shown by Tibshirani41 for feature selection. The sparsity property of LASSO which generates coefficient estimates of exactly zero, shrinks the estimation variance resulting with a more interpretable model42. Previous use of this application for genomic data43 has shown that the selective number of discriminating features can reach satisfactory classification. Regularization parameters were determined using a tenfold cross-validation (CV) procedure, with the label: continued AUD vs. remitted AUD as the response variable. The reduced set of the most discriminant features with non-zero coefficient was fed into the model to predict participants status to either continued AUD group or remitted AUD group. A supervised linear-kernel SVM that included parameter optimization was trained with a tenfold CV procedure to classify participants into the two groups. The tenfold CV procedure involved randomly
