Abstract
Alzheimer's disease (AD) is a progressive neurodegenerative disease and the current diagnosis tools, which are used, including clinical examination and neuroimaging, are costly, time consuming and complicated to generalize across demographics. Traditional machine learning (ML) methods such as Support Vector Machines (SVM), LightGBM, and Random Forest (RF) are known to be having weakness in terms of feature selection, class imbalance, and generalization. This paper presents a new Enhanced TabTransformer with Self-Supervised Learning (ETT-SSL) framework of AD classification. The structure combines the transformer-based architecture, self-supervised learning (SSL) for feature representation, SHAP-based feature selection, and the use of SMOTE to balance classes. Experimental results also indicate that ETT-SSL reaches a high accuracy of 95.8%, which is a much better result than baseline models (SVM: 72.1, RF: 78.3, LightGBM: 80.5) and even the standard TabTransformer (85.2%). Additionally, ETT-SSL offers greater accuracy and improved recall, overcoming the issue of false negatives in AD diagnosis. SHAP analysis can be used to improve transparency in clinical decision-making by making use of the model interpretability. The proposed framework recommends a salivary, interpretability and clinically practical methodology that can be used to expand to multimodal data source (e.g. MRI scans, genomic markers and electronic health records) to improve accuracy and generalisability.