Title: Utilizing Bayesian Meta-Learning to Model the Trajectory of Alzheimer’s Disease

Accurately forecasting whether a patient with Alzheimer’s disease will undergo mild or severe progression is critical for developing personalized treatment plans. Clinicians generally aim to predict the distribution of a discrete disease score based on an individual’s current MRI volume and their historical disease trajectory. However, this task poses significant challenges for traditional statistical regression models and single-task neural networks. Fitting separate models for each patient is impractical, as individuals usually have limited observations, while models that overlook individual-level correlations tend to generalize poorly.

Meta-learning offers a robust alternative, enabling the dynamic prediction of distributions without the need for retraining and effectively capturing nonlinear relationships between outcomes and covariates. Addressing this opportunity, we introduce a Bayesian meta-learner designed to be trained on data from multiple individuals while customizing the predictive disease score distribution to reflect each person’s unique historical data.

This approach allows for predictions on unseen individuals without requiring retraining and exhibits linear scaling relative to the number of historical observations. Furthermore, it is mathematically guaranteed to produce less overconfident predictions for long-term disease scores compared to deterministic counterparts. When tested on real-world data from the Alzheimer’s Disease Neuroimaging Initiative (ADNI) database, our model demonstrated performance comparable to both single-task models and deterministic meta-learners. Notably, it achieved substantial improvements in accuracy when predicting long-term disease progression.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC