An Infinite Mixture Model for Coreference Resolution in Clinical Notes.

Area of Science:

• Natural Language Processing
• Clinical Informatics
• Computational Linguistics

Background:

• Electronic Health Records (EHRs) require sophisticated Natural Language Processing (NLP) for effective utilization.
• Coreference resolution, a key relation detection task in NLP, is often suboptimal in clinical settings, impacting EHR data quality.
• Existing supervised and rule-based coreference resolution systems for clinical notes necessitate large, manually annotated datasets, limiting scalability.

Purpose of the Study:

• To develop an advanced, unsupervised method for relation detection, specifically coreference resolution, within clinical notes.
• To address the limitations of supervised methods by eliminating the need for extensive manual annotation.
• To enhance the accuracy and applicability of NLP techniques in processing large-scale EHR data.

Main Methods:

• An infinite mixture model with definite sampling was employed to resolve coreferent relations among mentions in clinical text.
• A novel similarity measure function was developed to identify and establish coreferent relationships.
• The system was evaluated on the i2b2 2011 coreference corpus.

Main Results:

• The proposed unsupervised system achieved a significant F-measure of 0.847 on the i2b2 2011 coreference corpus.
• The results demonstrate the effectiveness of the infinite mixture model for clinical coreference resolution.
• The unsupervised nature of the method shows promise for application in big-data clinical environments.

Conclusions:

• The developed infinite mixture model offers a scalable and effective solution for coreference resolution in clinical notes.
• This unsupervised approach overcomes the data annotation bottleneck, facilitating broader application in EHR processing.
• The system's high performance indicates a significant advancement in NLP for clinical text analysis and relation detection.