Multiclass classification of Alzheimer's disease prodromal stages using sequential feature embeddings and regularized...

Area of Science:

• Neuroscience
• Medical Imaging Analysis
• Machine Learning for Healthcare

Background:

• Early detection of neurodegenerative diseases like Alzheimer's disease (AD) is critical for timely intervention.
• Mild cognitive impairment (MCI) presents heterogeneous features, challenging accurate classification across cognitive normal (CN), MCI, and AD stages.
• Existing classification methods often rely on pairwise binary comparisons, complicating direct multiclass evaluation and interpretation.

Purpose of the Study:

• To develop a robust framework for direct multiclass classification of neurodegenerative disease stages (CN vs. MCI vs. AD).
• To overcome the limitations of binary classification approaches in handling the heterogeneity of MCI data.
• To establish a new state-of-the-art (SOTA) performance benchmark for multiclass classification in neurodegeneration detection.

Main Methods:

• A novel framework integrating unsupervised ensemble manifold regularized sparse low-rank approximation with a regularized multikernel support vector machine (SVM).
• Extraction of joint feature embeddings from MRI and PET neuroimaging data.
• Combination of imaging features with clinical data (Apoe4, Adas11, MPACC digits, Intracranial volume) for classification using a regularized multikernel SVM.

Main Results:

• Achieved SOTA performance in CN vs. MCI vs. AD multiclass classification with a mean accuracy of 84.87±6.09 and an F1 score of 84.83±6.12.
• Demonstrated strong generalization to binary classification tasks, achieving SOTA results across most categories.
• The CN vs. MCI binary classification showed a minimal performance decrease of 0.2% compared to existing best scores.

Conclusions:

• The proposed framework effectively addresses the challenge of classifying heterogeneous neurodegenerative disease stages.
• This integrated approach offers a more accurate and interpretable method for multiclass classification compared to sequential binary tasks.
• The findings pave the way for improved diagnostic tools in early neurodegeneration detection and intervention.