"Uncertainty-Aware Genomic Classification of Alzheimer's Disease: A Transformer-Based Ensemble Approach with Monte Carlo Dropout"

Authors: Taeho Jo*, Eun Hye Lee

This study presents TrUE-Net, an innovative deep learning framework that addresses a critical challenge in genomic-based Alzheimer's disease prediction: the lack of confidence estimates in model predictions. By combining transformer and random forest models with Monte Carlo Dropout for uncertainty quantification, the framework analyzed whole-genome sequencing data from 1,050 individuals (607 AD, 443 controls) from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort. The model demonstrated the ability to distinguish between "certain" and "uncertain" predictions, achieving 72.9% accuracy (F1=0.821) on high-confidence predictions while maintaining transparency about cases where predictions are less reliable.

Key Innovation: TrUE-Net introduces uncertainty quantification to genomic AD classification, enabling clinicians and researchers to identify which predictions are most trustworthy—a crucial advancement for translating AI predictions into clinical decision-making where understanding model confidence is essential for patient care.