SUSTain: Scalable Unsupervised Scoring for Tensors and its Application to Phenotyping

Area of Science:

• Data Science
• Machine Learning
• Computational Statistics

Background:

• Integer-valued data, common in event counts and ordinal measures, poses challenges for standard real-valued factorization methods.
• Conventional approaches treat integer data as real, leading to loss of crucial data characteristics and difficult interpretation.
• Existing methods struggle to preserve the inherent discrete nature of count or ordinal data.

Purpose of the Study:

• To introduce SUSTain, a new method for real-valued matrix and tensor factorization tailored for integer datasets.
• To develop an approach that extracts interpretable factor values as scores from small integer sets, enhancing data understanding.
• To provide efficient and accurate analysis of integer-valued data, particularly in domains like Electronic Health Records (EHR).

Main Methods:

• SUSTain partitions problems into efficiently solvable integer-constrained subproblems.
• It optimizes the order of subproblem solutions to maximize the reuse of intermediate results.
• Two variants, SUSTain_matrix and SUSTain_tensor, are proposed for matrix and tensor inputs, respectively.

Main Results:

• SUSTain demonstrates superior or comparable fit to state-of-the-art methods on synthetic and real EHR datasets.
• The method achieves significant speedups, up to 425x faster than baselines while maintaining comparable data fit.
• Application to EHR data successfully extracted patient phenotypes, with 87% validated as clinically meaningful heart failure related clusters.

Conclusions:

• SUSTain offers an effective and efficient solution for factorizing integer-valued data, overcoming limitations of real-valued methods.
• The interpretable scoring system facilitates a deeper understanding of feature contributions and data patterns.
• The method's successful application in EHR analysis highlights its potential for clinical data mining and phenotype discovery.