COPA: Constrained PARAFAC2 for Sparse & Large Datasets.

Area of Science:

• Multilinear data analysis
• Machine learning for healthcare
• Tensor decomposition methods

Background:

• PARAFAC2 (Parallel Factor Analysis 2) models irregular tensors, useful for time-varying patient data.
• Existing PARAFAC2 methods produce dense, noise-sensitive factors limiting interpretability.
• Challenges include imposing constraints (smoothness, sparsity, non-negativity) and scalability for large datasets.

Purpose of the Study:

• To develop a constrained PARAFAC2 (COPA) method for interpretable temporal modeling.
• To efficiently incorporate optimization constraints like temporal smoothness, sparsity, and non-negativity.
• To enable scalable analysis of large electronic health record (EHR) datasets.

Main Methods:

• Proposed COPA method incorporating optimization constraints into PARAFAC2.
• Utilized a hybrid optimization framework combining alternating optimization and ADMM (AO-ADMM).
• Evaluated on large-scale EHR datasets comprising hundreds of thousands of patients.

Main Results:

• COPA achieved significant speedups (up to 36x faster) compared to prior PARAFAC2 approaches.
• Outperformed baseline methods in speed while maintaining comparable accuracy.
• Demonstrated effective temporal phenotyping of medically complex children, revealing concise phenotypes and meaningful patient temporal profiles.

Conclusions:

• COPA successfully addresses the need for interpretable temporal modeling with imposed constraints.
• The hybrid AO-ADMM framework enables efficient handling of multiple constraints on large datasets.
• COPA's constrained factors provide clinically meaningful insights, validated by medical expert review.