Evaluating temporal relations in clinical text: 2012 i2b2 Challenge.

Area of Science:

• Clinical Informatics
• Natural Language Processing
• Biomedical Data Science

Background:

• The Sixth Informatics for Integrating Biology and the Bedside (i2b2) challenge addressed temporal relations in clinical narratives.
• A corpus of discharge summaries with temporal annotations was provided for developing and evaluating temporal reasoning systems.
• 18 international teams participated, presenting system analyses and future research directions.

Purpose of the Study:

• To evaluate systems for extracting clinically significant events from clinical narratives.
• To assess the extraction and normalization of temporal expressions (dates, times, durations) to ISO standards.
• To determine temporal relations between clinical events and expressions.

Main Methods:

• Information extraction tasks including event detection, temporal expression identification, and temporal relation classification.
• Evaluation of statistical machine learning (ML) and rule-based methods for event and temporal expression detection.
• Assessment of normalization techniques and hybrid approaches for temporal relation classification.

Main Results:

• Statistical machine learning methods excelled in event detection.
• Both ML and rule-based methods performed comparably in detecting temporal expressions, with rule-based approaches favored for normalization.
• Hybrid methods combining ML and heuristics achieved the best performance in temporal relation classification.

Conclusions:

• Machine learning is effective for event detection in clinical narratives.
• Rule-based systems are robust for temporal expression normalization.
• Hybrid approaches offer the most promising strategy for accurate temporal relation extraction in clinical text.