Multiple kernel learning with random effects for predicting longitudinal outcomes and data integration.

Area of Science:

• Biostatistics
• Machine Learning
• Epidemiology

Background:

• Predicting disease risk and progression is crucial in clinical research.
• Longitudinal data, collected over time, presents unique challenges for statistical learning methods.
• Existing kernel-based methods are primarily designed for independent data, not longitudinal datasets.

Purpose of the Study:

• To develop a novel statistical learning method for analyzing longitudinal data in disease prediction.
• To account for within-subject correlation in longitudinal measurements using subject-specific latent effects.
• To integrate and optimally combine heterogeneous data sources for enhanced prediction power.

Main Methods:

• Developed a novel statistical learning method incorporating subject-specific short-term and long-term latent effects via a designed kernel.
• Employed a multiple kernel learning framework with random effects for nonparametric prediction rules.
• Integrated various heterogeneous data sources, including imaging and genetic data, using distinct kernels for each modality.

Main Results:

• Demonstrated a substantial gain in prediction performance by leveraging the longitudinal aspect of the data.
• Successfully applied the method to large epidemiological studies, including Huntington's disease and Alzheimer's Disease Neuroimaging Initiative (ADNI).
• Showcased the ability to combine imaging and genetic data for predicting mild cognitive impairment.

Conclusions:

• The proposed regularized multiple kernel statistical learning with random effects is effective for longitudinal disease prediction.
• The method enhances prediction power by utilizing correlations among longitudinal measures and integrating diverse data.
• This approach offers a flexible framework for combining heterogeneous data sources in chronic disease research.