Exploiting Unlabeled Texts with Clustering-based Instance Selection for Medical Relation Classification.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Machine Learning

Background:

• Accurate classification of medical concept relations in clinical texts is vital for patient care.
• Limited labeled data and imbalanced classes hinder performance, especially for rare but important relation types.

Purpose of the Study:

• To improve medical relation classification performance, particularly for underrepresented relation types.
• To address challenges posed by limited labeled data and class imbalance in clinical text analysis.

Main Methods:

• Development of two novel clustering-based instance selection methods for weakly supervised learning.
• Method 1: Selects representative instances from unlabeled data clusters.
• Method 2: Selects counterpart instances from mixed labeled/unlabeled data clusters.

Main Results:

• Achieved substantial recall improvements for minority relation classes compared to traditional supervised learning.
• Maintained comparable performance on majority relation classes.

Conclusions:

• Weakly supervised learning with proposed instance selection methods effectively enhances medical relation classification.
• These methods are valuable for extracting critical information from clinical texts, especially for rare relations.