Abstract: Alzheimer’s disease (AD) has emerged as a significant health challenge globally, with projections reaching over 150 million affected individuals by 2050. Early diagnosis remains pivotal in managing disease progression and improving patient quality of life. Traditional diagnostic techniques rely heavily on neuropsychological assessments and qualitative MRI analysis, which suffer from subjective biases and inter-observer variability, often delaying diagnosis or leading to inaccuracies (Marcus et al., 2007; Marcus et al., 2010).
Recent breakthroughs in machine learning (ML), especially ensemble models combined with explainability techniques like SHAP (SHapley Additive exPlanations), have penned a new era in medical diagnostics where models can be both accurate and transparent (Lundberg & Lee, 2017). Our approach leverages Random Forest classifiers trained on the OASIS dataset—comprising heterogenous, multimodal data including MRIs, clinical scores, and demographics. The model’s decision process is elucidated through SHAP, allowing clinicians to understand the relative importance of features such as regional brain atrophy, age, and cognitive scores, thus aligning model outputs with biological plausibility and increasing clinical trust.
Furthermore, spatial localization through Grad-CAM overlays provides anatomical context to model decisions, highlighting brain regions like hippocampus and temporal lobes that are traditionally associated with AD (Selvaraju et al., 2017). This combined approach exemplifies a transparent, high-performing framework compatible with clinical workflows, offering a benchmark for future multi-modal, explainable AI models for neurodegenerative diseases, and emphasizes the road toward trustworthy AI-driven diagnostics that reconcile accuracy with interpretability (Mahavar et al., 2025).

Keywords: Alzheimer’s disease, MRI, Random Forest, SHAP, explainable AI, OASIS, ensemble learning, Grad-CAM.

Downloads: | DOI: 10.17148/IJIREEICE.2025.131039