Rubik: Knowledge Guided Tensor Factorization and Completion for Health Data Analytics | JoVE Visualize

Area of Science:

Computational health informatics
Machine learning for healthcare
Biomedical data science

Background:

Electronic health records (EHRs) contain vast patient data but are heterogeneous and often noisy.
Unsupervised phenotyping methods aim to discover clinical concepts from EHRs but lack medical knowledge integration and robust data handling.
Existing methods struggle with missing data and cannot effectively leverage current medical understanding.

Purpose of the Study:

To introduce Rubik, a novel constrained non-negative tensor factorization and completion method for computational phenotyping.
To enhance phenotype discovery by integrating existing medical knowledge and ensuring distinct, non-overlapping phenotypes.
To address challenges of noisy and missing data in EHRs through built-in tensor completion.

Main Methods:

Developed Rubik, a constrained non-negative tensor factorization and completion approach.
Incorporated guidance constraints for medical knowledge alignment and pairwise constraints for phenotype distinctness.
Utilized the Alternating Direction Method of Multipliers (ADMM) for scalable tensor factorization and completion.
Employed tensor completion to mitigate the impact of missing and noisy EHR data.

Main Results:

Rubik demonstrated superior performance in discovering more meaningful and distinct phenotypes compared to baseline methods.
Knowledge guidance constraints enabled Rubik to identify sub-phenotypes for major diseases.
The method showed significant speed improvements, running approximately seven times faster than state-of-the-art tensor methods.
Rubik proved scalable to large EHR datasets, processing millions of records efficiently.

Conclusions:

Rubik offers an advanced computational phenotyping solution by effectively integrating medical knowledge and handling data imperfections.
The method enhances the discovery of clinically relevant phenotypes, including detailed sub-phenotypes.
Rubik presents a computationally efficient and scalable approach for large-scale EHR analysis.