Machine learning-based coreference resolution of concepts in clinical documents.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Machine Learning

Background:

• Coreference resolution is crucial for understanding clinical narratives but is underutilized in healthcare.
• The 2011 i2b2 competition provided a valuable challenge for advancing this area.
• Identifying coreferent concepts (person, problems, treatments, tests) is key for clinical data analysis.

Purpose of the Study:

• To develop and evaluate a machine learning approach for coreference resolution in clinical documents.
• To collate coreferent chains of concepts within a clinical text corpus.
• To improve the extraction of structured information from unstructured clinical notes.

Main Methods:

• Utilized a machine learning approach employing graphical models for concept clustering.
• Features were categorized into domain-independent and domain-specific sets.
• Trained on 489 documents (6949 chains) and tested on 322 documents from the i2b2 dataset.

Main Results:

• Achieved an F-measure of 0.8423 without domain-specific features.
• Improved to an F-measure of 0.8483 when incorporating both domain-independent and domain-specific features.
• Demonstrated the effectiveness of the machine learning model in resolving coreferent concepts.

Conclusions:

• The proposed machine learning approach shows significant promise for coreference resolution in clinical text.
• Accurate coreference resolution facilitates practical applications like automated problem and medication list generation.
• This method enhances the utility of clinical documents for data mining and clinical decision support.