Multi-task longitudinal forecasting with missing values on Alzheimer's disease.

Area of Science:

• Computational neuroscience
• Biostatistics
• Machine learning

Background:

• Standard machine learning models struggle with joint learning and time-dependent, heterogeneous data common in dementia assessment.
• Existing methods often fail to adequately handle missing values in longitudinal datasets.

Purpose of the Study:

• To reformulate the SSHIBA Bayesian multi-view latent variable model for joint learning of dementia-related tasks.
• To address limitations in handling time-dependent, heterogeneous data with missing values.

Main Methods:

• Developed a novel Bayesian Variational inference framework for simultaneous missing value imputation and multi-view information integration.
• Utilized a semi-supervised formulation to leverage temporal data structure and handle missing data effectively.
• Created a common latent space to combine information from different time-points and data views.

Main Results:

• The proposed model demonstrated superior performance in imputing missing values compared to baseline methods.
• Achieved improved prediction performance for simultaneous diagnosis, ventricle volume, and ADAS score prediction.
• The semi-supervised formulation significantly enhanced imputation strategies.

Conclusions:

• The SSHIBA framework effectively imputes missing values in longitudinal dementia data.
• The model successfully performs joint prediction of diagnosis, ventricle volume, and ADAS score.
• Outperforms baseline methods in both imputation and multi-task prediction accuracy.