An annotation and modeling schema for prescription regimens

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Prescription regimens contain complex information (dose, frequency, route) often embedded in unstructured patient instructions.
• Standardizing the representation of prescription data is crucial for accurate interpretation and automated processing.
• Existing methods may lack the granularity to capture all essential prescription attributes effectively.

Purpose of the Study:

• To introduce TranScriptML, a novel semantic representation schema for prescription regimens.
• To describe the manual annotation process and corpus curation for TranScriptML.
• To evaluate machine learning models for automated generation of TranScriptML representations.

Main Methods:

• Developed TranScriptML, a schema with semantic tags and attribute frameworks for prescription concepts.
• Manually annotated a corpus of 2914 ambulatory prescriptions using the MITRE Annotation Toolkit (MAT).
• Trained automated annotation models using Conditional Random Field (CRF) machine learning on the manually annotated data.

Main Results:

• Achieved high agreement (up to 0.9 F-score) for frequent tags in manual annotation.
• Machine learning models demonstrated variable but promising performance, averaging 0.748 F-measure for tag and span modeling.
• Attribute modeling accuracy exceeded 0.9 for the best-performing methods.

Conclusions:

• TranScriptML provides a structured semantic representation for prescription regimens.
• Manual annotation and machine learning models can achieve high accuracy in representing prescription data.
• This structured data facilitates comparison with pharmacist-entered data, aiding discrepancy detection and correction.