Using natural language processing methods to classify use status of dietary supplements in clinical notes

Area of Science:

• Pharmacovigilance
• Clinical Informatics
• Natural Language Processing

Background:

• Dietary supplement safety is a concern due to potential interactions with prescribed medications.
• Electronic Health Records (EHRs) contain valuable data on supplement use within clinical notes.
• Extracting supplement use status from EHRs is crucial for ongoing safety research.

Purpose of the Study:

• To develop and compare rule-based and machine learning classifiers for extracting dietary supplement use status from clinical notes.
• To evaluate the performance and scalability of different classification methods.

Main Methods:

• Collected and annotated 2500 sentences related to 25 common dietary supplements into four use status categories: Continuing (C), Discontinued (D), Started (S), and Unclassified (U).
• Developed and evaluated both rule-based and machine learning-based (Maximum Entropy) classifiers on the annotated dataset.
• Compared the performance metrics (F-measure) of the developed classifiers.

Main Results:

• The rule-based classifier achieved F-measures ranging from 0.85 to 0.90 across the four categories.
• The optimal machine learning classifier (Maximum Entropy) achieved higher F-measures, ranging from 0.90 to 0.92.
• Machine learning demonstrated superior performance, efficiency, and scalability, especially with larger datasets.

Conclusions:

• Machine learning-based classifiers significantly outperform rule-based methods for categorizing dietary supplement use status in clinical notes.
• Future research will explore deep learning and hybrid systems for enhanced classification accuracy.