Unsupervised entity and relation extraction from clinical records in Italian

Area of Science:

• Natural Language Processing
• Medical Informatics
• Computational Linguistics

Background:

• Annotated clinical data is scarce and expensive for non-English languages.
• Unsupervised methods are crucial for leveraging clinical text data effectively.
• Information extraction from Italian clinical records presents unique challenges.

Purpose of the Study:

• To propose and evaluate an unsupervised text mining system for information extraction from Italian clinical records.
• To develop a two-step system for domain entity and relation extraction.
• To investigate the performance of unsupervised clustering for relation discovery.

Main Methods:

• Domain entities extracted using a metathesaurus and standard NLP tools.
• Unsupervised clustering applied to entity pairs represented by feature vectors.
• Automatic cluster labeling using significant features.
• System evaluation on a large dataset of Italian clinical records.

Main Results:

• The proposed unsupervised approach is effective for extracting information from Italian clinical records.
• Clustering methods demonstrate potential for discovering relations between entities.
• Different similarity measures impact performance, indicating areas for optimization.
• The system is well-suited for semi-automatic labeling of extracted relations.

Conclusions:

• Unsupervised text mining is a viable strategy for Italian clinical record analysis.
• The developed system offers a promising solution for semi-automatic information extraction.
• Further research can refine similarity measures and clustering techniques for improved performance.