Deep feature-based multi-class Alzheimer’s disease classification with statistical performance evaluation

Abstract

This study evaluated the performance of multiple machine learning classifiers for the classification of Alzheimer’s disease (AD) stages using deep features extracted from a pre-trained SqueezeNet model. Magnetic resonance imaging (MRI) scans were processed through SqueezeNet to generate high-dimensional feature vectors, which were then used as achieved an accuracy of 94.78% input to six classifiers: k-nearest neighbors (KNN), decision tree (DT), support vector machine (SVM), neural network (NN), naive Bayes (NB), and logistic regression (LR). Models were assessed using a 70/30% training-testing split and 5-, 10-, and 20-fold stratified cross validation. Principal component analysis (PCA) was applied to retain 99% of variance. On the original dataset consisting of 6,400 images, KNN has achieved 97.48% accuracy and 0.998 area under the curve (AUC), and when a larger dataset of 44,000 images was used it achieved an accuracy and of 94.78% and an AUC of 0.987, demonstrating the system’s robustness across scales. Statistical tests, including paired t-tests and Wilcoxon signed-rank tests, confirmed that KNN has significantly leveraged from PCA. These outcomes demonstrate that combining deep feature extraction with PCA improved the reliability and efficiency of the classifier for AD stage prediction.