Analysis of sampling techniques for imbalanced data: An n = 648 ADNI study.

Area of Science:

• Neuroimaging
• Machine Learning
• Biostatistics

Background:

• Neuroimaging datasets often exhibit class imbalance, complicating classifier development.
• Traditional classifiers may misclassify minority classes in imbalanced data, impacting diagnostic accuracy.

Purpose of the Study:

• To investigate ensemble methods combining feature selection and data sampling for imbalanced neuroimaging data.
• To identify optimal sampling techniques and feature selection algorithms for improved classification performance.

Main Methods:

• Systematic analysis of various data sampling techniques (undersampling, oversampling, combined) and feature selection algorithms.
• Evaluation using Random Forest and Support Vector Machines classifiers, assessing accuracy, AUC, sensitivity, and specificity.
• Utilized the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset for experimental validation.

Main Results:

• K-Medoids based undersampling significantly improved classifier performance compared to other sampling methods and no sampling.
• Sparse logistic regression with stability selection demonstrated competitive performance in feature selection.
• The proposed ensemble model using multiple undersampled datasets achieved stable and promising results.

Conclusions:

• Ensemble methods integrating data sampling and feature selection are effective for handling imbalanced neuroimaging data.
• K-Medoids undersampling and sparse logistic regression are key components for building robust classifiers in ADNI studies.