Phenotyping through Semi-Supervised Tensor Factorization (PSST).

Area of Science:

• Computational biology
• Health informatics
• Machine learning

Background:

• Computational phenotyping aims to identify patient characteristics from electronic health records.
• Tensor factorization is a promising approach for automated phenotyping.
• Existing methods often fail to incorporate auxiliary patient data effectively.

Purpose of the Study:

• To introduce Phenotyping through Semi-Supervised Tensor Factorization (PSST), a novel computational phenotyping method.
• To leverage partial disease status knowledge for improved phenotype derivation.
• To demonstrate PSST's ability to uncover clinically relevant and predictive patient phenotypes.

Main Methods:

• Constructing tensors from electronic health record data.
• Applying semi-supervised tensor factorization incorporating known disease statuses for subsets of patients.
• Evaluating phenotype discriminative and meaningfulness using case studies.

Main Results:

• PSST successfully generated computational phenotypes from patient electronic health records.
• Case studies on type-2 diabetes and resistant hypertension showcased PSST's potential.
• PSST-derived phenotypes were more discriminative than unsupervised methods and more meaningful than supervised methods.

Conclusions:

• PSST offers an effective approach to computational phenotyping by integrating semi-supervised learning.
• The method enhances the discovery of clinically relevant patient phenotypes.
• PSST shows promise for improving patient stratification and understanding of diseases.