Risk calibrated selective tabular transformers for explainable Alzheimer’s disease staging

Early and reliable staging of Alzheimer’s disease (AD) using tabular clinical biomarkers is crucial but challenging due to heterogeneous features and model overconfidence. This research proposes a Risk-Calibrated Tabular Transformer (RC-TabFormer) that tokenises each feature, applies self-attention, employs temperature scaling for calibration, integrates a selective prediction mechanism that abstains on low-confidence cases, and fuses outputs with a LightGBM classifier. The research outcomes on Alzheimer’s Disease Neuroimaging Initiative (ADNI) datasets with subject-level cross-validation, RC-TabFormer achieves 95.2 ± 0.5% accuracy, 96.7 ± 0.6% F1, 98.4 ± 0.4% AUC, 1.2 ± 0.2% Expected Calibration Error (ECE), and 91.2 ± 0.5% MCC for binary AD vs non-AD classification. On three-class Staging Alzheimer’s Disease, Mild Cognitive Impairment, Cognitively Normal (AD/CN/MCI) it attains 89.8 ± 0.5% accuracy, 89.7 ± 0.5% macro-F1, 95.1 ± 0.4% macro-AUC, 1.7 ± 0.3% ECE, and 84.3 ± 0.5% Matthews correlation coefficient (MCC), surpassing strong baselines by up to 5 points while markedly improving calibration. Coverage–risk curves show that selective prediction reduces errors at lower coverage, and explainability analyses using Integrated Gradients and SHAP identify key biomarkers such as Clinical Dementia Rating Scale Sum of Boxes (CDSRB) and tau (τ) ratios. RC-TabFormer provides a unified risk-aware framework that integrates Transformer-based tabular representation learning, LightGBM fusion, probability calibration, selective abstention, and feature-level explanation. The results demonstrate its potential for further external and prospective evaluation as a clinical decision-support framework.