Toward high-throughput phenotyping: unbiased automated feature extraction and selection from knowledge sources.

Area of Science:

• Biomedical Informatics
• Computational Biology
• Clinical Research

Background:

• Electronic health records (EHRs) contain valuable narrative data for population-scale phenotyping.
• Current methods for selecting text features for phenotyping algorithms are time-consuming and require expert input.
• Developing efficient and accurate phenotyping tools is crucial for advancing clinical and genetic research.

Purpose of the Study:

• To introduce an automated method for extracting and selecting informative text features from EHRs for phenotyping algorithms.
• To develop phenotyping algorithms in an unbiased manner, comparable in accuracy to expert-curated features.
• To improve the efficiency of high-throughput phenotyping for research.

Main Methods:

• Collected comprehensive medical concepts from public knowledge sources automatically.
• Utilized natural language processing (NLP) to identify concept occurrence patterns in EHR narratives.
• Selected informative features for phenotype classification and trained a penalized logistic regression model.

Main Results:

• Applied the method to identify rheumatoid arthritis (RA) and coronary artery disease (CAD) cases in EHR data.
• Achieved high classification accuracy with automated features: AUC of 0.951 for RA and 0.929 for CAD.
• Performance was comparable or slightly superior to models trained with expert-curated features.

Conclusions:

• Automated NLP-based feature selection yields phenotyping algorithms with high accuracy and efficiency.
• The developed method offers a significant advancement for high-throughput phenotyping.
• The majority of selected features were interpretable, facilitating clinical understanding.