Distant supervision for medical concept normalization.

Area of Science:

• Computational linguistics
• Medical informatics
• Natural Language Processing

Background:

• Medical Concept Normalization (MCN) maps informal medical phrases to formal concepts.
• Deep learning models perform well but struggle with large-scale MCN due to dataset limitations.
• Existing methods use knowledge bases for automatic labeling, but scaling remains a challenge.

Purpose of the Study:

• To develop a scalable MCN approach using automatically generated datasets.
• To improve MCN model performance by incorporating diverse, real-world medical language.
• To reduce the cost and effort associated with manual data labeling for MCN.

Main Methods:

• Extracted informal phrases and medical concepts from patient discussion forums.
• Utilized a synthetically trained classifier and a medical entity linker for data extraction.
• Employed pretrained sentence encoding models to identify nearest neighbor phrases for concepts.
• Combined forum-derived data with knowledge base examples to train the MCN model.

Main Results:

• Achieved a 15.9% and 17.1% increase in classification accuracy on two benchmark datasets.
• Demonstrated superior performance compared to previous state-of-the-art methods.
• Successfully avoided manual data labeling, reducing resource intensity.

Conclusions:

• The proposed method effectively scales Medical Concept Normalization to millions of concepts.
• Leveraging patient discussion forums provides a valuable source for MCN training data.
• This approach offers a cost-effective and highly accurate solution for MCN.