Natural Language Mapping of Electrocardiogram Interpretations to a Standardized Ontology

Area of Science:

• Cardiology
• Medical Informatics
• Natural Language Processing

Background:

• Electrocardiogram (ECG) interpretations are often external to electronic health records (EHRs), requiring natural language processing (NLP) for computable data extraction.
• Challenges in NLP for ECGs include misspellings, nonstandard abbreviations, jargon, and equivocation.

Purpose of the Study:

• To develop a reliable and efficient NLP algorithm for extracting quantitative and diagnostic information from ECG narrative reports.
• To map extracted ECG data to a standardized ontology developed by major cardiology societies.
• To ensure the algorithm is modifiable for various EHR and ECG reporting systems.

Main Methods:

• Developed an NLP algorithm using structured query language to extract and map ECG data.
• Trained the algorithm on 43,861 ECG reports and tested it on 46,873 reports.
• Validated the algorithm externally using data from a different institution and reporting system.

Main Results:

• Achieved 100% accuracy, precision, recall, and F1-measure for quantitative ECG data extraction.
• Exceeded 99% performance for mapping diagnostic categories to the standardized ECG ontology.
• Demonstrated a processing speed of approximately 20,000 reports per minute with similar performance in external validation.

Conclusions:

• The developed algorithm effectively creates computable representations of ECG interpretations with high performance.
• The provided software and lookup tables allow for easy customization and use with different EHR and ECG systems.
• The algorithm offers significant utility for research and clinical decision-support applications incorporating ECG findings.