PhenoLines: Phenotype Comparison Visualizations for Disease Subtyping via Topic Models.

Area of Science:

• Computational biology
• Medical informatics
• Machine learning

Background:

• Topic models applied to clinical data can identify disease subtypes with evolving phenotype prevalence.
• Interpreting these complex topic models for phenotype relevance and interrelationships remains a challenge.
• Existing methods offer limited intuition for characterizing disease subtypes.

Purpose of the Study:

• To introduce PhenoLines, a visual analysis tool for interpreting disease subtypes derived from topic models.
• To facilitate the characterization of disease subtypes by comparing phenotype prevalence and identifying dominant phenotypes.
• To support the optimization of topic models in clinical data analysis.

Main Methods:

• Utilized topic models on cross-sectional patient comorbidity data (e.g., electronic health records).
• Developed a data transformation workflow using the Human Phenotype Ontology for hierarchical phenotype organization.
• Implemented a novel measure of phenotype relevance to simplify the topic model topology.
• Incorporated feedback from machine learning and clinical experts throughout the design process.

Main Results:

• PhenoLines enables comparison of phenotype prevalence within and across disease subtype topics.
• The tool supports identifying dominant phenotypes, ages of effect, and clinical validity of subtypes.
• Initial evaluations with experts suggest PhenoLines effectively aids in characterizing and optimizing topic models.
• The Human Phenotype Ontology integration and new relevance measure enhance interpretability.

Conclusions:

• PhenoLines offers a promising visual approach for interpreting complex disease subtypes identified by topic models.
• The tool addresses the challenge of understanding phenotype relevance and interrelationships in clinical data.
• PhenoLines facilitates a deeper understanding of disease heterogeneity and supports clinical decision-making.