Algorithm 1: Ensemble feature selection with nested cross-validation

Input: miRNA expression dataset D ∈ ℝN × F, where N is the number of samples and F is the number of miRNA features. Output:       • Minimal miRNA feature set Fminimal       • Best-performing model M*       • Mean performance metrics across all validation sets Step 1. Initialization              Fminimal ← ∅              M* ← None              $\mathrm{Let} p={\displaystyle \frac{N}{10}}$ (i.e., Leave-6-Out Cross-Validation when N = 60)              Generate p = 10 non-overlapping folds:                   ${\{(T_i,V_i)\}}_{i=1}^p$, where Ti ∈ ℝ(N − p) × F, Vi ∈ ℝp × F Step 2. Outer Cross-Validation (Leave-p-Out)              for each i ∈ {1, 2, …, p} do                   Let Ti be the outer training set and Vi be the outer validation set Step 3. Inner Cross-Validation (Stratified k-Fold) and Feature Selection                   $\mathrm{Split} T_i \mathrm{into} k \mathrm{stratified} \mathrm{folds}: {\{(t_j,v_j)\}}_{j=1}^k$                   for each j ∈ {1, 2, …, k} do                        Feature Selection on tj:                             Apply RFE, Random Forest importance, LASSO, and SelectKBest                        Model Training:                             Train classifiers (LR, RF, SVM, XGBoost, AdaBoost) on tj                        Model Evaluation:                             Evaluate on vj using Accuracy, Sensitivity, Specificity, F1 Score, and AUC                        Model Selection:                             Choose best-performing model Mj                        Update Feature Set:                             Add features to Fminimal if selected in ≥ 3 inner folds                   Select most frequent model across inner folds as:                        M* = argmaxMj (frequency of selection in inner folds) Step 4. Model Validation on Outer Fold                       Use M* and Fminimal to classify Vi                       Evaluate performance using Accuracy, Sensitivity, Specificity, F1 Score, and AUC Return: -Fminimal—final feature set -M*—best-performing model -Average performance metrics over all Vi, i = 1, 2, …, p