Feature level-based group lasso method for amnestic mild cognitive impairment diagnosis

Area of Science:

• Neuroimaging
• Machine Learning
• Biomedical Data Analysis

Background:

• Brain morphological changes are observed in amnestic mild cognitive impairment (aMCI).
• Existing classification methods often fail to fully utilize these brain morphological measures.

Purpose of the Study:

• To improve the classification accuracy of aMCI by enhancing multitask learning frameworks.
• To fully consider the relevance among related tasks and supplementary information from unrelated tasks.

Main Methods:

• A feature level-based group lasso (FL-GL) method was proposed.
• A correlation matrix guided feature selection within a group lasso framework.
• Support vector machine (SVM) classifiers were trained on selected features and combined using weighted voting.

Main Results:

• The FL-GL method demonstrated improved classification accuracies of 6.12% and 4.92% on two independent datasets (Xuan Wu Hospital and ADNI).
• Leave-one-out cross-validation confirmed the method's effectiveness.

Conclusions:

• The feature level-based group sparsity term is crucial for the method's performance.
• Considering feature-level correlations enhances multitask learning for improved aMCI and normal control classification.