Uncertainty-Aware Longitudinal Forecasting of Alzheimer’s Disease Progression Using Deep Learning

Clinical Alzheimer's disease (AD) progression is inherently longitudinal and heterogeneous, yet most existing deep learning models treat it as a static, single-step classification problem. This study addresses the need for a more nuanced approach by asking: Can we generate multi-year, probabilistic trajectories of disease progression that respect the ordinal nature of AD stages, and can we reliably quantify the model's uncertainty to signal when a prediction should be treated with caution?

The authors propose a unified framework that combines three key methodological innovations:

1. Ordinal Temporal Fusion Transformer (TFT): An architecture adapted for multi-horizon forecasting that uses a CORAL ordinal output layer to respect the natural progression from cognitively normal to mild cognitive impairment (MCI) and dementia.
2. Probabilistic Trajectory Generation: Using a Mixture Density Network (MDN) conditioned on the TFT's patient-context representation, the model generates five-year probabilistic forecasts for multiple clinical markers, including diagnosis state, CDR Sum of Boxes, MMSE scores, and hippocampal volume.
3. Decomposed Uncertainty Estimation: The model separates aleatoric uncertainty (inherent unpredictability in a patient's disease course) from epistemic uncertainty (model-specific ignorance) using a combination of analytic mixture variance and a five-member deep ensemble.

The model demonstrates superior performance in predicting next-visit diagnoses compared to standard linear, recurrent, and transformer baselines, with particularly strong gains in distinguishing MCI from dementia. The generated trajectories show high calibration, with 90% credible intervals that widen appropriately over the forecast horizon. Crucially, the epistemic uncertainty component successfully identifies 'out-of-distribution' scenarios; it increases for rare progression patterns and when the model is evaluated on the external OASIS-3 cohort, where prediction errors are higher. This provides a clinically actionable signal for when the model's output is less reliable.