Alzheimer’s disease diagnosis and risk prediction using advanced machine learning models

Abstract

Alzheimer's Disease (AD), which impairs memory, cognition, and behavior, is a progressive Neurodegenerative Disorder (ND). Accounting for 60-80% of dementia cases, AD advances from mild memory loss to complete loss of interaction with the environment. As there is no cure for AD, early detection is vital to mitigate disease progression. Currently, Machine Learning (ML) and Deep Learning (DL) have exhibited promise in neuroimaging-based AD prediction by employing approaches, namely Lasso Regression (LR), Convolutional Neural Networks (CNN), Support Vector Machine (SVM), and Deep Neural Networks (DNN). Nevertheless, these models face limitations, comprising overfitting, high error rates, and limited small dataset performance. To address these challenges, a two-part framework is introduced in this study. The AD Neuroimaging Initiative (ADNI) MRI dataset is used by the first implementation to enhance prediction and classification accuracy through the GELU Swish Radial Basis Function Network (GS-RBFN). MRI images undergo preprocessing steps, such as normalization, skull stripping, and spatial smoothing, followed by essential brain tissue segmentation with Brownian Log Scaling Archimedes Optimization-centric Watershed Segmentation (BLSAOWS) and Feature Selection (FS) using the Base Switch Rule Infimum and Supremum-based Rock Hyrax Swarm Optimization (BSRISRHSO) approach. AD classification is further performed by the GS-RBFN classifier. GS-RBFN (Gaussian—Swish Radial Basis Function Network) and TT Self-Weighted Deep-AD3-Net—for early detection and staging of AD. The proposed models employ novel optimization algorithms (BLSAOWS and BSRISRHSO) for improved segmentation and feature selection from MRI datasets. The performance analysis of GS-RBFN model learns the features efficiently, it achieves superior accuracy (98.45%), precision (98.44%), F-measure (98.44%), Sensitivity (98.44%), Recall (98.45%), and specificity (98.45%). In the second implementation, AD risk scoring and stage prediction are focused on using The AD Prediction Of Longitudinal Evolution (TADPOLE) dataset. This methodology involves ranking critical variables, Risk Score (RS) calculation, and brain shrinkage analysis. Essential variables are selected by the Recursive Hypothesis-Creation Algorithm (RHCA), with the True True Self-Weighting Mechanism (TT-SWM) calculating RS. By using the Queue-Boltzmann-ConstantSphere (QBCS) technique, brain shrinkage in the hippocampus is measured, whereas Gray-Level Co-occurrence Matrix (GLCM) features facilitate staging through the Deep-AD3-Net classifier. Experimental evaluation using the TADPOLE dataset achieved an accuracy of 98.45%, sensitivity of 97.82%, and AUC of 0.981, outperforming recent state-of-the-art models. The frameworks demonstrate robustness and clinical potential for early intervention and treatment planning. The study advances existing AD diagnostic approaches by improving interpretability, computational efficiency, and predictive reliability. Keywords: Alzheimer’s disease, Data augmentation, Brownian Log Scaling Archimedes Optimization-based Watershed Segmentation (BLSAOWS), Base Switch Rule Infmum and Supremum-based Rock Hyrax Swarm Optimization (BSRISRHSO), GELU and SWISH-based Radial Basis Function Network (GSRBFN), Recursive Hypothesis-Creation Algorithm (RHCA), Phylogenetic Method (PM), one Gray-Level Co-occurrence Matrix (GLCM), Gray-Level Co-occurrence Matrix (GLCM Genetic Algorithm (GA), Deep-AD3-Net classifier.