A supervised framework for resolving coreference in clinical records.

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Coreference resolution in clinical records is crucial for accurate information extraction.
• Existing methods often struggle with the complexities of medical language and record structures.

Purpose of the Study:

• To develop and evaluate an automated method for resolving coreference between medical concepts in clinical records.
• To assess the impact of lexical similarity, recency, synonymy, and context on coreference resolution performance.

Main Methods:

• A multiple-pass sieve approach employing support vector machines (SVMs) was utilized.
• Features included lexical similarity, mention recency, Wikipedia-based synonymy, and local context.
• Evaluation used standard metrics (MUC, CEAF, B(3)) on the 2011 i2b2/VA Shared Task datasets.

Main Results:

• Achieved an F score of 0.821 on the ODIE dataset, closely matching top-performing systems.
• Attained an F score of 0.906 on the i2b2 dataset, outperforming the median system.
• Demonstrated competitive performance compared to other participants in the shared task.

Conclusions:

• Simple lexical matching methods significantly contribute to achieving competitive coreference resolution performance.
• The ability to detect patient mentions in electronic medical records enhances coreference resolution.
• Performance degradation was observed on pathology reports due to complex synonymy and limited patient mentions.