A Comparative Study of PCA and LDA for Dimensionality Reduction in a 4-Way Classification Framework

Abstract Alzheimer’s disease (AD), recognized as the secondmost impactful neurological disorder and currently incurable, stands as the leading cause of dementia. An imperative research focus is efficiently diagnosing the stages of patients, distinguishing early or late Mild Cognitive Impairment and AD from those with normal cognitive function. Advancements in anatomical and diffusion-weighted imaging, coupled with machine learning techniques, have significantly progressed in this predictive domain. However, in real-world trials, datasets often contain numerous features, and the curse of dimensionality can introduce challenges such as increased computational complexity, overfitting, and diminished model interpretability. To address these issues, the present study explores the efficacy of Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) as dimensionality reduction techniques. LDA, a supervised technique emphasizing class separability, surpasses PCA, particularly in selecting features that significantly contribute to discriminating between classes. The 3D-LDA features obtained were subsequently assessed across various machine learning algorithms, leading to the establishment of a 4-way classification framework that utilized the K-Nearest Neighbors model. The outcome of this evaluation yielded an impressive accuracy rate of 87% in predicting the four different classes.