Uncertainty-aware genomic classification of Alzheimer's disease: a transformer-based ensemble approach with Monte Carlo dropout

Abstract

Alzheimer's disease (AD) has complex genetic factors that make accurate prediction from genomic data challenging. Current methods lack confidence estimates for their predictions. Our objective is to develop and evaluate an uncertainty-aware deep learning framework for AD prediction that can identify which predictions are reliable. We developed Transformer-based, Uncertainty-aware, Ensemble Network (TrUE-Net), a deep learning framework that combines transformer and random forest models to predict AD from whole-genome sequencing data. The key innovation is using Monte Carlo Dropout to estimate prediction confidence, allowing the model to identify cases where it is uncertain. We analyzed 1050 individuals (607 AD, 443 controls) from Alzheimer's Disease Neuroimaging Initiative cohort. On 525 test samples, accuracy of TrUE-Net without uncertainty threshold was 65.1% with area under the receiver operating characteristic curve 0.664. Using uncertainty thresholds, the model classified 75.4% of samples (n = 396) as 'uncertain' and 24.6% (n = 129) as 'certain'. The uncertain group showed accuracy of 62.6% and F1 of 0.584, while the certain group showed accuracy of 72.9% and F1 of 0.821. Uncertainty quantification through Monte Carlo Dropout provides a framework for assessing prediction reliability in genomic AD classification, potentially allowing more informed interpretation of model outputs.