Uncertainty-Aware Longitudinal Trajectory Forecasting in Alzheimer’s Disease with a Multi-Task Transformer

Accurately predicting the progression of Alzheimer’s disease is crucial for physicians, who need not only to identify the most likely short-term diagnosis but also to understand how the condition will evolve over time and how reliable the prediction is for each patient. However, the heterogeneity of disease expression and the scarcity of longitudinal data make this task challenging. Many current deep learning methods reduce the problem to single-step classification and give little consideration to the role of uncertainty across subsequent visits. We present NeuroTrajectory-Net, a multi-task Transformer that forecasts the future Mini-Mental State Examination (MMSE) slope, the future Clinical Dementia Rating (CDR) slope, and the binary progression risk from a subject’s past visits alone. Using the OASIS-2 dataset of 150 patients (373 visits), we evaluate the model on a single patient-level (subject-disjoint) split of 90, 30, and 30 subjects for training, validation, and testing, so that no subject contributes sequences to more than one partition. On the test set, NeuroTrajectory-Net achieves a forecasting mean absolute error (MAE) of 1.82 MMSE points/year for the future MMSE slope. However, its classification head does not perform as well as single-task models (AUC = 0.51 versus 0.87–0.98). Monte Carlo Dropout yields personalized 95% prediction intervals; however, their empirical coverage on the test set is only 48.8%, so they are under-calibrated and not yet suitable for clinical decision-making. The exact target-generation rule and the handling of zero-padded visits have not yet been re-verified at the code level and are reported as reproducibility limitations; the progression classification results should therefore be regarded as preliminary. This paper introduces a multi-task framework designed to avoid target and split leakage and reports both its strengths and its limitations.