An interpretable Bayesian framework for Alzheimer's disease prediction with uncertainty quantification

Alzheimer's disease is a neurodegenerative disorder with a variable rate of progression. Predictive tools that can leverage the available modalities of data in any setting to predict the progression of the disease will benefit clinicians and patients alike. However, most of the tools lack the ability to quantify the uncertainty and do not provide the reasoning behind the predictions. In this work, we propose a novel Bayesian Encoder-Decoder GRU (BEND-GRU) framework to predict a patient's Alzheimer's Disease Assessment Scale (ADAS-13) and Clinical Dementia Rating - Sum of Boxes (CDR-SB) cognitive scores for the second and third years using the baseline and first-year data from the ADNI dataset. We quantified the uncertainty in the predictions, incorporating both the model variability and data noise, and the predictions were interpreted using the Integrated Gradients method. The BEND-GRU model accurately predicted Year 2 and Year 3 scores for both ADAS-13 (Year 2: MAE 2.98, R² 0.83; Year 3: MAE 3.60, R² 0.80) and CDR-SB (Year 2: MAE 0.69, R² 0.76; Year 3: MAE 0.95, R² 0.72). The ablation study shows the framework can be deployed in limited resource settings, as it yielded competitive results utilizing modest modalities such as demographics and cognitive scores. MMSE score and delayed recall total score were among the key predictors. The framework can predict the future progression of AD, providing clinicians with confidence intervals around each prediction along with explanatory reasoning.