Phenotype Concept Set Construction from Concept Pair Likelihoods.

Area of Science:

• Computational biology
• Biomedical informatics
• Clinical research informatics

Background:

• Phenotyping algorithms are crucial for clinical research using observational data.
• Manually developed algorithms, like those in the eMERGE Network, are high-quality but labor-intensive.
• Efficient construction of phenotype definitions is needed to advance research.

Purpose of the Study:

• To propose a framework for learning from existing phenotype concept sets to aid in creating new phenotype definitions.
• To leverage the structure of established phenotypes for automated or semi-automated definition generation.
• To reduce the time and effort required for developing clinical phenotyping algorithms.

Main Methods:

• Utilized eMERGE Network phenotypes as reference concept sets.
• Engineered features to characterize pairwise relationships between concepts.
• Modeled these relationships as a concept graph and employed edge prediction.
• Identified candidate phenotype concept sets as highly connected subgraphs.

Main Results:

• Developed a framework that learns from the structural properties of existing phenotype concept sets.
• Demonstrated a method for identifying potential novel phenotype definitions through graph-based analysis.
• The approach facilitates the discovery and composition of new phenotype definitions.

Conclusions:

• The proposed framework offers a scalable approach to assist in the construction of novel phenotype definitions.
• This method has the potential to significantly expedite the development of phenotyping algorithms for clinical research.
• Learning from existing concept sets can streamline the process of defining complex phenotypes from electronic health records.