Deep Learning Forecasts Alzheimer's Progression with Uncertainty Awareness

Predicting the progression of Alzheimer's disease (AD) is vital for clinical management, but existing deep learning models often oversimplify the problem by treating diagnoses as discrete categories and failing to quantify forecast reliability. A new probabilistic framework addresses these limitations by offering a more comprehensive approach to longitudinal AD forecasting. This framework integrates ordinal diagnosis prediction, the generation of multi-horizon trajectories, and a detailed decomposition of uncertainty. It adapts a Temporal Fusion Transformer encoder with specific layers and loss functions to respect disease stage ordering and improve sensitivity, particularly for transitions from Mild Cognitive Impairment (MCI) to dementia. Conditioned on the learned patient context, an autoregressive Mixture Density Network generates five-year probabilistic trajectories for diagnosis, cognitive scores, and hippocampal volume. The model demonstrates superior performance over traditional and deep learning baselines, especially in discriminating between MCI and dementia. Crucially, it provides near-nominal 90% credible interval coverage, with uncertainty increasing over time, and effectively separates aleatoric (inherent) from epistemic (model-based) uncertainty, offering more reliable and interpretable forecasts.