NeuroFusion for Robust and Explainable Multimodal Deep Learning in Fine-Grained Staging of Alzheimer's Disease Across Imaging and Clinical Biomarkers

Received: 20 August 2025 | Revised: 4 October 2025 and 17 October 2025 | Accepted: 19 October 2025 | Online: 9 February 2026

Corresponding author: Sumendra Yogarayan

Abstract

The subtle transition from healthy cognition to very mild dementia often leaves even experienced clinicians uncertain, highlighting the need for computational frameworks that integrate heterogeneous data sources while providing clinically meaningful explanations. Existing Artificial Intelligence (AI) approaches frequently reduce Alzheimer's Disease (AD) prediction to binary classification, rely on a single modality, or lack interpretability, thereby limiting their translational impact. This study introduces a multimodal Deep Learning (DL) framework designed for fine-grained, four-stage staging of AD across Non-Demented, Very Mild, Mild, and Moderate states. The framework integrates structural Magnetic Resonance Imaging (MRI) and structured clinical biomarkers within a unified architecture, employing pre-trained convolutional and transformer-based networks for imaging alongside gradient-boosted decision trees for tabular features such as demographics, cognitive scores, and laboratory measures. To promote transparency and clinical trust, the framework incorporates complementary interpretability strategies: Gradient-weighted Class Activation Mapping (Grad-CAM) to identify discriminative neuroanatomical regions and attention mechanisms to highlight influential clinical variables. Evaluation was conducted on the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, with diagnostic labels determined at the subject level using clinician consensus informed by Clinical Dementia Rating (CDR) and Mini-Mental State Examination (MMSE) scores. The dataset comprised 229 Non-Demented, 398 Very Mild, 192 Mild, and 176 Moderate cases. GhostNet achieved an F1-score of 0.888 for Non-Demented and 0.847 for Mild Dementia, whereas Very Mild remained the most challenging stage (best F1-score ≤ 0.628). On structured clinical features, Extreme Gradient Boosting (XGBoost) attained an accuracy of 95.35%. Despite reduced performance for intermediate stages due to class imbalance and overlapping phenotypes, model training remained stable with minimal overfitting. This work provides one of the first interpretable multimodal frameworks for four-stage AD staging, advancing beyond conventional binary models and demonstrating the value of integrating complementary imaging and clinical modalities. By combining robust diagnostic accuracy with transparent, clinician-facing explanations, the approach offers a scalable and trustworthy pathway toward AI-enabled dementia staging, with particular promise for deployment in resource-limited healthcare systems.