Enabling phenotypic big data with PheNorm.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Health Data Science

Background:

• Electronic health record (EHR)-based phenotyping is crucial for disease inference.
• Traditional phenotyping algorithms require time-intensive manual annotation of EHR data, limiting scalability.
• Automating feature curation has been achieved, but annotation remains a significant bottleneck.

Purpose of the Study:

• To introduce PheNorm, a novel phenotyping algorithm that bypasses the need for expert-labeled training samples.
• To demonstrate that PheNorm can achieve high-throughput phenotyping without manual annotation.

Main Methods:

• PheNorm normalizes predictive EHR features to a normal mixture distribution.
• Transformed features undergo denoising and are combined into a classification score.
• The algorithm automates the generation of phenotyping algorithms.

Main Results:

• PheNorm was validated for coronary artery disease, rheumatoid arthritis, Crohn's disease, and ulcerative colitis.
• Achieved Area Under the Curve (AUC) scores of 0.90, 0.94, 0.95, and 0.94, respectively.
• Performance was comparable to supervised algorithms trained on 100-300 samples, with no statistically significant difference.

Conclusions:

• PheNorm algorithms demonstrate accuracy on par with those trained using annotated samples.
• The algorithm fully automates the generation of accurate phenotyping models.
• PheNorm enables EHR-driven annotations to scale for phenotypic big data analysis.