Results in Engineering · Apr 20, 2026
Hybrid MRI–tabular deep learning for predicting Alzheimer’s amyloid positivity with external validation and an explainable clinical framework
By Rezaee K, Zadeh HG, Khosravi M
As anti-amyloid therapies become integrated into routine care, neurology clinics increasingly need low-burden, MRI-first tools to identify amyloid positivity (Aβ+), given that PET or cerebrospinal fluid (CSF) analysis remains costly or invasive. Deep learning offers a natural solution, but prior MRI-based studies often suffer from limited calibration, small single-site cohorts, ad-hoc tuning, and poor external generalization; few reports present decision-relevant operating points. We introduce a hybrid, multi-output deep learning framework that simultaneously predicts continuous amyloid burden and a calibrated Aβ positivity (Aβ+) probability (using Binary Cross-Entropy and focal loss with post-hoc temperature scaling). The framework processes T1-weighted MRI and optionally fuses this data with low-cost covariates. Training employs site-aware nested cross-validation, ComBat harmonization, and multi-objective hyperparameter optimization (HPO) targeting both discrimination and calibration. We assess clinical utility using pre-specified, fixed thresholds (Youden-J and fixed-sensitivity), confusion matrices, and decision-analytic readouts. On the ADNI dataset, the classifier achieves AUROC≈0.90 (AUPRC≈0.92) with Accuracy≈0.82 at frozen thresholds. The regressor yields MAE≈8.6 CL, RMSE≈13.6 CL, and R2≈0.54. On OASIS−3 (an unseen external test set), we maintain Accuracy≈0.84 with Sensitivity≈0.83 and Specificity≈0.85 under the same thresholds; calibration improves markedly after temperature scaling (ECE≈0.025). Our innovations—including the multi-output design, principled calibration, harmonization, and frozen-policy external evaluation—close key gaps in risk interpretability and transportability without increasing acquisition burden. To enable clinical translation, we report frozen thresholds and calibration/decision curves on external data, outlining multicenter validation. Clinically, this method supports risk-stratified referral for PET/CSF with auditable behavior (global SHAP on covariates, cross-site parity checks) and stable operating points.