Evaluating Homoscedastic Uncertainty Weighting and Generative Tau Imputation in a Multimodal, Multitask Deep Learning Framework for Alzheimer's Disease.

Abstract

Alzheimer’s disease is a progressive neurodegenerative disorder and the leading cause of dementia. Multimodal longitudinal deep learning models are increasingly used to forecast disease stage and clinical decline, yet it remains unclear whether popular methodological refinements meaningfully improve performance in practice. We analyzed data from 2,271 participants in the Alzheimer’s Disease Neuroimaging Initiative. A multimodal CNN–BiLSTM model was trained jointly to predict diagnostic stage and four clinical outcomes: Alzheimer’s Disease Assessment Scale, Mini-Mental State Examination, Functional Activities Questionnaire, and Clinical Dementia Rating–Sum of Boxes. We evaluated three commonly used strategies: homoscedastic uncertainty weighting for multitask learning, saliency-guided snapshot analysis, and conditional variational autoencoder–mixture density network imputation of missing longitudinal tau trajectories. Preprocessing was implemented in a split-safe manner, with all preprocessing objects fit on training data only and then applied unchanged to validation and test data. Uncertainty weighting showed a modest descriptive classification benefit, particularly in the no-tau setting, but did not provide a consistent regression advantage across endpoints. Learned task weights were stable across splits and assigned lower relative precision to classification than to the regression heads. Saliency and occlusion analyses indicated that cognitive and magnetic resonance imaging inputs carried the largest aggregated attribution, whereas the tau branch contributed negligibly when included. Snapshot models retained some classification signal but underperformed the intact model and did not preserve regression performance. The tau imputation model reconstructed sparse tau trajectories reasonably well, but adding imputed tau did not improve downstream prediction. In this CNN–BiLSTM pipeline, added methodological complexity yielded mixed and limited practical benefit, underscoring the importance of empirically testing refinements rather than assuming they improve multimodal Alzheimer’s disease forecasting.