Enhancing Alzheimer's disease classification with a transformer-based model using self-supervised learning.

Area of Science:

• Neuroscience
• Artificial Intelligence
• Biomedical Informatics

Background:

• Alzheimer's disease (AD) diagnosis relies on costly and complex methods.
• Traditional machine learning (ML) models struggle with feature selection, class imbalance, and generalization for AD classification.

Purpose of the Study:

• To introduce an Enhanced TabTransformer with Self-Supervised Learning (ETT-SSL) for accurate and interpretable Alzheimer's disease classification.
• To develop a clinically practical diagnostic methodology using salivary data.

Main Methods:

• Developed an ETT-SSL framework integrating transformer architecture and self-supervised learning.
• Employed SHAP for feature selection and SMOTE for class balancing.
• Utilized salivary data for classification.

Main Results:

• ETT-SSL achieved a high accuracy of 95.8%, significantly outperforming baseline models (SVM: 72.1%, RF: 78.3%, LightGBM: 80.5%) and standard TabTransformer (85.2%).
• The framework demonstrated improved accuracy and recall, reducing false negatives in AD diagnosis.
• SHAP analysis provided model interpretability for clinical decision-making.

Conclusions:

• The proposed ETT-SSL framework offers a highly accurate, interpretable, and clinically practical approach to Alzheimer's disease diagnosis using salivary biomarkers.
• The methodology is adaptable for multimodal data integration (e.g., MRI, genomics, EHRs) to further enhance diagnostic accuracy and generalizability.