Regularized multi-task learning with individual-feature-based task correlations for Alzheimer’s cognitive score prediction

Abstract

Background and Objective:

Predicting multiple cognitive scores from brain features for Alzheimer’s disease (AD) patients can aid in early intervention treatments and enhance disease management. Regularized multi-task sparse learning has become an important approach since it can predict multiple cognitive scores and identify biomarkers in one process. However, existing methods often assign the same correlation coefficient for a pair of tasks on all features, even though their relationships at different features are usually different. Introducing inaccurate task correlations into multi-task learning can hinder the improvement of models’ prediction performance. This study overcomes the above limitation by introducing a novel multi-task learning framework that captures task correlations at a fine-grained, feature-specific level.

Methods:

We propose a novel individual-feature-based task correlation matrices guided multi-task learning (IFTMTL) method. The method constructs a non-smooth convex objective function that jointly learns regression models for multiple cognitive scores. This objective function integrates task and feature correlations to enhance predictive performance. Specifically, the fine-grained inter-task correlations are modeled at the feature level using a set of task correlation matrices, while feature correlations are captured via the Pearson coefficient. An iterative optimization algorithm is developed to jointly update the task correlation structures and model parameters.

Results:

The proposed IFTMTL significantly outperforms the 11 competitive methods in the normalized mean squared error (nMSE) and correlation coefficient (CC) metrics. Specifically, IFTMTL improves the nMSE metric of the Multi-Task Feature Learning method by 4.09%, and the CC metric by 1.68%. Furthermore, IFTMTL identifies the most severely affected brain regions in AD, including the left hippocampus, left middle temporal, and right entorhinal.

Conclusion:

IFTMTL improves AD prediction performance by effectively incorporating unequal task correlations at the feature level, enabling better knowledge transfer across tasks. The method outperforms existing approaches in predicting cognitive scores and identifying significant biomarkers, such as the hippocampus and middle temporal regions, which are crucial for AD prediction and clinical analysis.