Surrogate-assisted feature extraction for high-throughput phenotyping.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Health Data Science

Background:

• Electronic medical records (EMR) are crucial for identifying patient phenotypes.
• Developing phenotyping algorithms is resource-intensive due to extensive human effort.
• Scalability and robustness of EMR phenotyping remain significant challenges.

Purpose of the Study:

• Introduce a high-throughput unsupervised feature selection method for EMR phenotyping.
• Enhance the scalability and robustness of phenotyping algorithms without sacrificing accuracy.
• Reduce the reliance on extensive human resources for algorithm development.

Main Methods:

• Propose the Surrogate-Assisted Feature Extraction (SAFE) method.
• Utilize International Classification of Diseases, Ninth Revision (ICD-9) and NLP counts as silver-standard labels.
• Select candidate features highly predictive of silver-standard labels from comprehensive medical concept pools.

Main Results:

• Trained algorithms for coronary artery disease, rheumatoid arthritis, Crohn's disease, and ulcerative colitis.
• SAFE demonstrated superior performance (AUC and F-score) compared to existing methods and experts.
• SAFE's advantage was particularly pronounced with limited label sizes.

Conclusions:

• SAFE enables high-throughput phenotyping by automating informative feature selection.
• Reduces overfitting and the need for gold-standard labels in algorithm training.
• Identifies potentially crucial features overlooked by other methods or human experts.