Multiscale Ensemble Learning Approach for Early Prediction of Mild Cognitive Impairment Using Brain Imaging

Inhalt:
Abstract: Early stratification of mild cognitive impairment (MCI) patients is crucial for preventing progression to Alzheimer’s disease (AD). We introduce, a fully automated and interpretable pipeline using T1-weighted MRI (T1-MRI) that generates individualized risk reports in under 20 minutes on a standard CPU. The workflow comprises FSL BET skull stripping, ANTs MNI-152 registration, FastSurfer AAL3 segmentation, and Gaussian smoothing at 1/2/6/10 mm scales to extract ≈400 regional grey matter volumes. A two-stage feature reduction (PCA retaining ≥90 % variance → Fisher Discriminant Ratio selecting 30–50 top components) feeds into a nested CV stacking ensemble (XGBoost, LightGBM, SVM) optimized via Optuna. Study and Cohorts: Retrospective, two-cohort study comprising a derivation set (n=60) for model development with nested cross-validation and an external validation set (n=18) for out-of-sample verification. Results: In the derivation set, the ensemble achieved AUC 0.92 with ACC 0.95, SEN 0.97, and SPE 0.93; external validation yielded AUC 0.91.SHAP-based interpretability highlights the hippocampus, posterior cingulate, and temporal pole as primary predictors, aligning with Braak V–VI pathology. This method balances high accuracy, transparency, and resource parsimony, making it ideal for community screening and multi-center trials.