Algorithm 1. Multimodal Fusion Pipeline

Input: Cognitive features Xc, Imaging features Xi, Labels y 1. Preprocess data      - Clean dataset      - Split into folds (cross-validation) 2. Apply feature scaling      For each scaling method in {Standard, Min–Max, Robust}:           - Xc_scaled ← scale(Xc)           - Xi_scaled ← scale(Xi) 3. Unimodal modeling      - Train model Mc on Xc_scaled      - Train model Mi on Xi_scaled 4. Multimodal fusion      For each fusion strategy:           (a) Early fusion:                 X_fused ← concatenate(Xc_scaled, Xi_scaled)                 Train model M_fused on X_fused           (b) Intermediate fusion:                 hc ← encoder_c(Xc_scaled)                 hi ← encoder_i(Xi_scaled)                 h_fused ← fusion_module(hc, hi)                 Train classifier on h_fused           (c) Late fusion:                 yc ← Mc(Xc_scaled)                 yi ← Mi(Xi_scaled)                 y_fused ← combine(yc, yi) 5. Model training      - Optimize parameters      - Apply early stopping when available 6. Evaluation      - Compute AUC-ROC, Accuracy, Balanced Accuracy, F1-score, Cohen’s kappa, Average Precision      - Average results across folds Output: Performance metrics for each fusion strategy