This study proposes a probabilistic framework for longitudinal modeling of Alzheimer's disease progression that combines ordinal diagnosis prediction, multi-horizon trajectory generation, and decomposed uncertainty estimation. The approach utilizes a Temporal Fusion Transformer encoder and an autoregressive Mixture Density Network to generate five-year probabilistic trajectories while quantifying both aleatoric and epistemic uncertainty.
- A Temporal Fusion Transformer encoder is adapted with a CORAL ordinal output layer, asymmetric loss weighting, and converter oversampling to respect disease-stage ordering.
- An autoregressive Mixture Density Network generates five-year probabilistic trajectories for diagnosis state, CDR Sum of Boxes, MMSE orientation, and hippocampal volume.
- On the ADNI dataset, the model outperforms linear, recurrent, and transformer baselines, showing the strongest gains in MCI-versus-dementia discrimination.
- Generated trajectories achieve near-nominal 90% credible interval coverage with widening uncertainty across the forecast horizon.
- Aleatoric and epistemic uncertainties are separated using analytic mixture variance and a five-member bootstrap ensemble.
This framework provides clinicians with not only the most likely next diagnosis but also reliable forecasts of how a patient may evolve over time, addressing the limitations of single-step classification methods.