A novel approach to enhance feature selection using linearity assessment with ordinary least squares regression for Alzheimer’s Disease stage classification

• Published in: Multimedia tools and applications Vol. 83; no. 38; pp. 86059 - 86078
• Main Authors: Mabrouk, Besma, Bouattour, Nadia, Mabrouki, Noura, Sellami, Lamia, Ben Hamida, Ahmed, for the Alzheimer’s Disease Neuroimaging Initiative
• Format: Journal Article
• Language: English
• Published: New York Springer US 01.11.2024; Springer Nature B.V
• Subjects: Alzheimer's disease; Brain; Classification; Cognitive ability; Computer Communication Networks; Computer Science; Correlation coefficients; Data Structures and Information Theory; Disease control; Impairment; Least squares method; Linearity; Magnetic resonance; Medical imaging; Multimedia Information Systems; Regression models; Special Purpose and Application-Based Systems; Tensors; Track 8: Medical Imaging; Alzheimer’s disease; Brain connectivity maps; Ordinary Least Squares (OLS); Linearity assessment; Logistic regression model; White matter fiber disruption; Diffusion Weighted Imaging (DWI)
• ISSN: 1573-7721; 1380-7501; 1573-7721
• DOI: 10.1007/s11042-024-20254-3

Diagnosing Alzheimer’s disease (AD) in its prodromal stage is a significantly crucial area of research. Approximately 50% of individuals within the well-known Mild Cognitive Impairment (MCI) cohort are estimated to progress to AD, and the factors influencing conversion remain unknown. Gaining insights into the disease evolution can enhance support strategies and potentially slow down the pathology. Utilizing the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database, our objective is to construct a framework for distinguishing between Normal Controls (NC) and different stages of Alzheimer’s Disease (AD), encompassing Earlier Mild Cognitive Impairment (EMCI), Later Mild Cognitive Impairment (LMCI), and AD patients. In pursuit of this objective, we preprocessed Diffusion Tensor and Magnetic Resonance brain images from 237 subjects, generating corresponding brain connectivity maps. Notably, we introduce an innovative linearity assessment method that utilizes the Ordinary Least Squares (OLS) linear regression model to identify and select relevant features for classification. This approach effectively identifies features with strong linear relationships to the target variable. Our method’s superiority is demonstrated through a comparative analysis with the traditional SelectKBest approach. By integrating this feature selection strategy with a Logistic Regression model, our study achieves both efficient and highly accurate classification outcomes, highlighting the effectiveness of the proposed method. In a four-class classification scenario, the model attained an accuracy of 66 % ± 0.06 . In binary classification, the results were equally impressive, with an area under the curve of 0.68 ± 0.10 % for CN vs. EMCI discrimination, 99 ± 0.02 % for distinguishing LMCI from adjacent classes CN and EMCI, and 0.79 % ± 0.08 for discriminating AD from healthy subjects. Additionally, the calculation of Pearson’s correlation coefficient has been employed to identify cortical regions affected by changes, explore the nature of fiber disconnection propagation from one stage to another, and establish the traceability of the interference origin between stages. The summarized results reveal an apparent flow of white matter disruption from the right to the left hemisphere.