Rethinking modeling Alzheimer's disease progression from a multi-task learning perspective with deep recurrent neural...

Area of Science:

• Neuroscience
• Medical Imaging
• Machine Learning

Background:

• Alzheimer's disease (AD) is a progressive neurodegenerative disorder.
• Predicting AD progression using longitudinal data is crucial but challenging due to missing features and small sample sizes.
• Existing models struggle with the complexity of multi-timepoint AD progression data.

Purpose of the Study:

• To develop an accurate Alzheimer's disease progression prediction model.
• To address challenges of missing data and small sample sizes in longitudinal AD studies.
• To leverage correlations between different prediction tasks and time points.

Main Methods:

• Proposed a multi-task learning framework to impute missing values and predict future progression.
• Hypothesized and analyzed correlations among clinical diagnosis, cognitive scores, ventricular volume, imputation, and prediction tasks across multiple time points.
• Developed an end-to-end deep multi-task learning method with balanced multi-task dynamic weight optimization.
• Validated the model on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset.

Main Results:

• The multi-task learning model demonstrated improved performance in predicting AD progression.
• Achieved significant improvements in metrics such as mAUC, BCA, and MAE for ADAS-Cog13 and Ventricles.
• Showcased benefits and flexibility, particularly for long-term prediction (M60 time point).

Conclusions:

• The proposed deep multi-task learning framework effectively predicts Alzheimer's disease progression.
• The model's ability to adaptively impute missing values enhances prediction accuracy.
• This approach offers a flexible and robust solution for analyzing longitudinal AD data.