Hierarchical multiclassification of Alzheimer’s disease based on DIABLO fusion of multimodal neuroimaging and clinical information

Abstract Current approaches relying on structural magnetic resonance imaging (sMRI) and functional magnetic resonance imaging (fMRI) often employ simple data concatenation or decision-level fusion, which overlooks deeper intermodal relationships. Moreover, most classification models do not reflect clinically established diagnostic pathways. To address these limitations, we introduce a feature-level fusion framework based on multiblock partial least squares discriminant analysis (DIABLO), combined with hierarchical multi-class modeling, to leverage multimodal synergies for improved classification of clinically-diagnosed AD. This study integrates sMRI, fMRI, and clinical information from the ADNI dataset (872 participants: 443 NC, 300 MCI, 129 AD). Feature fusion was performed using kernel principal component analysis (KPCA) and DIABLO, respectively. A hierarchical multi-classification model was constructed and validated under three diagnostic scenarios (NC → MCI → AD, MCI → NC → AD, AD → NC → MCI). We compared seven classifiers (SVM, RF, GBDT, XGBoost, AdaBoost, Bagging, Stacking) based on KPCA fusion. Ablation experiments confirmed the synergistic value of multimodal integration: three-modal DIABLO achieved an AUC of 0.934 for NC vs. MCI, substantially outperforming single-modal fMRI (AUC 0.797). The DIABLO-based model using sMRI, fMRI, and clinical information achieved AUCs of 0.992 for AD vs. non-AD and 0.934 for NC vs. MCI, outperforming KPCA-based fusion. This study demonstrates that fusing sMRI, fMRI, and clinical information provides complementary biological features, improving multi-classification accuracy across the AD spectrum. DIABLO appears to be a powerful and interpretable multimodal fusion framework, showing great potential for auxiliary AD diagnosis and warranting further validation in larger and more diverse cohorts.