A Leakage‑Free Group ICA–SVM Framework for Alzheimer's Disease Classification Using Static, Dynamic, and Partial Resting‑State Connectivity

This repository provides a **rigorously validated, leakage‑free machine learning pipeline** for the binary classification of Alzheimer’s disease (AD) versus cognitively normal (CN) controls using resting‑state functional MRI (rs‑fMRI). The framework combines: - **Group Independent Component Analysis (gICA)** performed strictly within each cross‑validation fold to prevent data leakage - **Multi‑measure functional connectivity** (static, dynamic mean/std, and partial correlation) - **Linear Support Vector Machine (SVM)** with Bayesian hyperparameter optimization (Optuna) - **Nested cross‑validation** for unbiased performance estimates The pipeline achieves **79.0% accuracy, 0.851 AUC, and 0.756 F1‑score** on 138 participants from the ADNI dataset, demonstrating robust discrimination while remaining fully interpretable—a critical advantage over “black‑box” deep learning approaches. --- ## Key Features - **Leakage‑free design:** gICA decomposition is computed exclusively on training subjects in each outer fold; test subjects are projected via dual regression. - **Rich connectivity features:** Static functional connectivity, dynamic connectivity (sliding‑window mean and standard deviation), and regularised partial correlation. - **Fully reproducible preprocessing:** fMRIPrep + aCompCor denoising + MNI spatial normalisation. - **Nested cross‑validation:** Outer 5‑fold CV for evaluation; inner 5‑fold CV for Optuna‑driven hyperparameter tuning (Bayesian optimisation, 50 trials). - **Interpretable model:** Linear SVM weights allow direct mapping of discriminative brain connections to resting‑state networks. - **Publication‑ready outputs:** ROC curves, confusion matrices, comparison tables, and a pipeline flowchart (300 dpi TIF). --- ## Pipeline at a Glance 1. **Data ingestion** – ADNI rs‑fMRI (138 subjects, 2 mm MNI space, TR=3 s). 2. **Preprocessing** – fMRIPrep (motion & distortion correction) → nuisance regression (aCompCor + motion parameters) → spatial smoothing (6 mm FWHM). 3. **Outer CV loop** – 5‑fold stratified split at subject level. 4. **Feature extraction (per fold):** - **Group ICA** (CanICA, 40 components) *on training set only* - **Dual regression** → subject‑specific time series - **Static FC** – full‑length Pearson correlation - **Dynamic FC** – sliding window (60 s) → mean & std across windows - **Partial correlation** – Ledoit‑Wolf regularised - Concatenation → **3 120 features** per subject 5. **Classification (per fold):** - Z‑score scaling (fit on training) - Inner 5‑fold CV → Optuna tunes `C` for linear SVM - Train final SVM → predict test fold 6. **Aggregation** – Accuracy, AUC, F1‑score (mean ± SD across outer folds).