DISCOVERING PATIENT PHENOTYPES USING GENERALIZED LOW RANK MODELS

Area of Science:

• Computational biology
• Medical informatics
• Machine learning applications in healthcare

Background:

• Clinical practice relies on identifying patient phenotypes for tailored treatments.
• Traditional phenotype discovery is slow, labor-intensive, and often heuristic.
• Challenges like missing data and heterogeneity hinder machine learning-based phenotype discovery in electronic health records.

Purpose of the Study:

• To introduce Generalized Low Rank Modeling (GLRM) as a flexible framework for accelerating phenotype discovery.
• To demonstrate GLRM's capability in overcoming data limitations in electronic health records.
• To identify known and novel patient phenotypes in diverse datasets.

Main Methods:

• Applied Generalized Low Rank Modeling (GLRM) to two distinct patient datasets.
• Utilized the Healthcare Cost and Utilization Project National Inpatient Sample (NIS) with millions of hospitalization records.
• Analyzed a local autism spectrum disorder cohort dataset with granular electronic health record features, including clinical notes.

Main Results:

• GLRM effectively handled missing data, sparsity, and heterogeneity in both large-scale administrative and granular clinical datasets.
• The model successfully captured known patient phenotypes within the analyzed data.
• Putative, or potential, phenotypes were also identified, suggesting new avenues for clinical research.

Conclusions:

• Generalized Low Rank Modeling (GLRM) offers a robust solution for phenotype discovery in complex healthcare data.
• This machine learning approach can significantly accelerate the identification of patient subgroups.
• GLRM demonstrates potential for improving clinical decision-making and treatment strategies by revealing nuanced patient phenotypes.