Bayesian multitask learning regression for heterogeneous patient cohorts.

Area of Science:

• Machine Learning
• Bayesian Inference
• Biomedical Data Analysis

Background:

• Multitask learning (MTL) aims to enhance individual task performance by leveraging shared information across related tasks.
• A critical challenge in MTL is determining the optimal structure of task relatedness, which is often unknown a priori.
• Interpretability and uncertainty quantification are crucial for reliable application of machine learning in the biomedical domain.

Purpose of the Study:

• To develop a Bayesian multitask learning model capable of inferring task relationship structures directly from data.
• To offer two model variations: one with a diffuse Wishart prior assuming equal task relatedness, and another using a Bayesian graphical LASSO prior for sparse relatedness.
• To ensure model interpretability and uncertainty quantification through linear mappings and conjugate priors, respectively.

Main Methods:

• Proposed a Bayesian multitask learning framework with flexible priors on the task precision matrix.
• Implemented a diffuse Wishart prior for an uninformative approach to task relatedness.
• Utilized a Bayesian graphical LASSO prior to encourage sparsity in the inferred task relatedness structure.
• Employed linear mappings for interpretability and conjugate priors for full posterior inference.

Main Results:

• The proposed model successfully recovered underlying task relationships and identified jointly relevant features using synthetic data.
• Demonstrated superior predictive performance compared to Single Task Learning (STL) models across all tested biomedical applications.
• Outperformed existing multitask learning methods in the majority of scenarios evaluated.
• Validated the model's utility in Alzheimer's disease progression, Parkinson's disease assessment, and cervical cancer screening compliance.

Conclusions:

• The developed Bayesian multitask learning model effectively infers task structures, leading to improved predictive accuracy in biomedical applications.
• The model provides interpretable insights into task relationships and quantifies uncertainty, making it suitable for sensitive domains.
• This approach offers a robust alternative to existing MTL methods, particularly when task relatedness is unknown or complex.