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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Medical Informatics

Background:

• Alzheimer's disease (AD) progression is variable, necessitating accurate predictive tools for clinical management.
• Current predictive models often lack uncertainty quantification and interpretability, limiting their clinical utility.
• Leveraging diverse data modalities is crucial for improving AD progression prediction.

Purpose of the Study:

• To develop a novel Bayesian Encoder-Decoder GRU (BEND-GRU) framework for predicting Alzheimer's Disease Assessment Scale (ADAS-13) and Clinical Dementia Rating - Sum of Boxes (CDR-SB) scores.
• To quantify prediction uncertainty by incorporating model variability and data noise.
• To provide interpretable predictions using the Integrated Gradients method.

Main Methods:

• Utilized baseline and Year 1 data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.
• Developed a Bayesian Encoder-Decoder GRU (BEND-GRU) model for score prediction.
• Employed Integrated Gradients for model interpretability and uncertainty quantification.

Main Results:

• The BEND-GRU model achieved high accuracy in predicting Year 2 and Year 3 ADAS-13 and CDR-SB scores.
• Year 2 ADAS-13 prediction: MAE 2.98, R² 0.83; Year 3 ADAS-13 prediction: MAE 3.60, R² 0.80.
• Year 2 CDR-SB prediction: MAE 0.69, R² 0.76; Year 3 CDR-SB prediction: MAE 0.95, R² 0.72.
• Ablation studies confirmed the framework's effectiveness even with limited data modalities (e.g., demographics, cognitive scores).

Conclusions:

• The BEND-GRU framework offers accurate and interpretable predictions of Alzheimer's disease progression.
• The model quantifies uncertainty, providing clinicians with confidence intervals for predictions.
• Key predictors identified include Mini-Mental State Examination (MMSE) score and delayed recall total score, highlighting the framework's utility in resource-limited settings.