Can machine learning predict how Alzheimer's will progress in individual patients?

Predicting whether an Alzheimer's patient will decline mildly or severely requires learning from sparse individual histories while generalizing across patients—a task standard regression and single-task neural networks handle poorly. Researchers developed a Bayesian meta-learner trained on multiple patient trajectories that adapts to each person's MRI volumes and disease history without retraining. On ADNI data, it matched deterministic alternatives on overall performance but substantially outperformed them on long-term forecasts with calibrated confidence intervals.