Abstract
Alzheimer’s Disease is a major leading cause of dementia globally and is also responsible for seventh leading cause of death. While, existing deep learning approaches can indicate whether the case should be classified as AD or non-AD, they lack explainability and interpretability.
The current existing practices mainly involve Magnetic Resonance Imaging (MRI) scans, clinical assessments and genetic data which were obtained from Alzheimer’s Disease Neuroimaging Initiative (ADNI) and while these single data sources might be reliable, AD is a disease which shows subtle early symptoms and is progressive in nature.
In this study, we propose an integrated multimodal architecture that integrates genetic risk markers, clinical evaluations, and (Magnetic Resonance Imaging) MRI scans into a unified diagnostic paradigm. The model employs a Bidirectional Long Short-Term Memory (Bi-LSTM) network to identify longitudinal trends in the clinical and genetic data, and a Convolutional Neural Network (CNN) to extract spatial characteristics from MRI scans. We use Explainable AI (XAI) techniques like SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), Gradient-weighted Class Activation Mapping (Grad-CAM), and Partial Dependence Plot (PDP/ICE) plots that collectively show how specific attributes and brain regions affect the model's predictions to guarantee transparency and clinical trust. When combined with strong explainability, this multimodal fusion offers a more thorough and comprehensible method of diagnosing Alzheimer's disease.
According to our evaluation results, the suggested multimodal framework yields a mean accuracy of 83.5%, which is in line with performance reported in previous multimodal Alzheimer's disease studies that incorporate genetic indicators, clinical scores, and MRI, under a strict evaluation method like Leave-One-Subject-Out (LOSO), where each subject is examined separately, this level of performance is expected. This makes the task more difficult but also more clinically realistic.
The integration of several Explainable AI (XAI) techniques, such as SHAP, LIME, Grad-CAM, and PDP/ICE plots, not only increased overall accuracy but also offered a thorough understanding of how specific features affected model predictions at patient and global levels. The most significant clinical and genetic predictors were consistently found to be CDRSB, MMSE, RAVLT_Learning, and APOE4 in the XAI analysis, whereas Grad-CAM indicated disease-relevant brain regions such the hippocampus and medial temporal lobe, which closely matched known AD pathology. While the underlying model predictions stayed constant, parameter tweaks inside the XAI approaches showed that changes mostly influenced the explanation’s smoothness and visual clarity. This stability highlights the framework's potential value for practical diagnostic support by showing that it not only provides great predictive performance but also solid and clinically interpretable reasoning.