A flexible framework for recognizing events, temporal expressions, and temporal relations in clinical text.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Biomedical Text Mining

Background:

• Clinical records contain valuable information for patient care and research.
• Automated extraction of events and temporal information from these records is crucial for clinical decision support and data analysis.
• Existing methods often struggle with the complexity and nuances of clinical language.

Purpose of the Study:

• To develop and evaluate a natural language processing (NLP) method for automatic recognition of events, temporal expressions, and temporal relations in clinical records.
• To assess the performance of a hybrid approach combining supervised, unsupervised, and rule-based techniques.

Main Methods:

• Employed a combination of supervised learning (Conditional Random Fields, Support Vector Machines) and unsupervised learning (Brown clustering) for event and temporal expression recognition.
• Utilized an automated feature selection technique to optimize supervised models.
• The integration of unsupervised methods rendered the approach semi-supervised.

Main Results:

• Achieved an F1-measure of 0.8045 for event recognition and 0.6154 for temporal expression recognition on the i2b2 shared task data.
• The event recognition system ranked third among 14 participants.
• Temporal link detection achieved an F1-measure of 0.5594 (overall) and 0.5258 (end-to-end).

Conclusions:

• A hybrid NLP approach effectively recognizes events and temporal expressions in clinical text with minimal medical knowledge.
• Temporal normalization and relation recognition necessitate more sophisticated modeling of medical knowledge.
• Future work should focus on enhancing temporal relation extraction with specialized medical domain understanding.