Transformer-Based Representation Learning with Explainable AI for Robust Alzheimer’s Stage Prediction

Abstract

Alzheimer’s disease is the most prevalent form of dementia among the elderly and poses a significant global healthcare burden, with millions of new cases reported each year. It is characterized by progressive neurodegeneration, beginning with mild cognitive impairment and advancing to severe memory loss, behavioral changes, and reduced awareness. A pretrained Swin Transformer model is utilized to extract high-dimensional and discriminative features that capture subtle neuroanatomical variations across stages such as Mild, Moderate, Very Mild, and Normal conditions. The extracted features are evaluated using baseline classifiers including Generalized Linear Model (GLM), Generalized Learning Vector Quantization (GLVQ), and Perceptron. The proposed Optimal Rule Fit Classifier (ORFC) demonstrates superior performance by effectively modeling complex and non-linear relationships. To enhance usability, a Tkinter-based graphical interface supports dataset handling, preprocessing, training, and evaluation, while a Flask-based server enables real-time remote predictions. Additionally, Explainable Artificial Intelligence (XAI) techniques provide interpretable insights, such as brain-region atrophy and ventricular enlargement. Experimental results show that the proposed system achieves 99.43% accuracy, highlighting its effectiveness for early diagnosis and clinical decision support.