A flexible framework for deriving assertions from electronic medical records.

Area of Science:

• Computational linguistics
• Medical informatics
• Machine learning

Background:

• Clinical text processing involves identifying medical concepts and their status (e.g., present, absent, uncertain).
• Leveraging diverse resources for clinical text analysis necessitates optimal feature selection for machine learning models.

Purpose of the Study:

• To develop and evaluate natural language processing (NLP) techniques for medical concept identification and assertion classification in clinical text.
• To devise and assess feature-selection methods for optimizing machine learning classifiers in clinical text analysis.

Main Methods:

• Utilized two machine learning classifiers: Support Vector Machines (SVMs) and Conditional Random Fields (CRFs).
• Developed novel feature-selection techniques to reduce manual experimentation and enhance classifier performance.
• Evaluated approaches on the 2010 i2b2/VA challenge dataset.

Main Results:

• Achieved a 79.59 micro F-measure for medical concept extraction.
• Achieved a 93.94 micro F-measure for assertion classification.

Conclusions:

• Machine learning-based approaches offer adaptability for new clinical data and tasks.
• Optimal feature selection is crucial for maximizing the performance of ML techniques in clinical text processing.
• The proposed feature-selection methods yielded state-of-the-art results, outperforming existing approaches.