A temporal constraint structure for extracting temporal information from clinical narrative

Li Zhou, Genevieve B. Melton, Simon Parsons, George HripcsakJournal of Biomedical Informatics 2005 · DOI

TL;DR

This paper introduces a novel "temporal constraint structure" designed to capture the rich and complex temporal information found in clinical narratives, such as discharge summaries. By providing a structured and expressive way to represent temporal relationships, this work lays the groundwork for more sophisticated temporal reasoning in medical informatics, ultimately aiming to improve patient care and medical decision-making.

What Problem Does This Paper Address?

Understanding the timeline of medical events is crucial for effective medical reasoning, diagnosis, and treatment. While electronic health records (EHRs) contain vast amounts of temporal data, much of this information is embedded within unstructured clinical narratives (like doctor's notes). Extracting and representing this temporal information in a way that machines can understand and reason with is a significant challenge. Existing methods often struggle with the nuances of clinical language, such as relative time expressions ("two days ago"), qualitative relationships ("fever before rash"), and vague timeframes ("early morning"). This paper aims to bridge this gap by proposing a structured representation that can accurately capture this diverse temporal information from clinical narratives.

Methods

The core of this paper is the development of a "temporal constraint structure." This structure was built using a grounded, bottom-up approach, analyzing a corpus of clinical discharge summaries to identify recurring patterns in temporal expressions. The structure itself is designed to model time over which an event occurs by constraining its start and end times. It includes fields for:

  • Event endpoints: Specifying the start and/or finish of an event.
  • Anchor information: Identifying a reference point (like a specific date or another event) for the temporal expression.
  • Qualitative and metric temporal relations: Capturing relationships like "before," "after," "equal," and specifying durations or time differences.
  • Vagueness: Indicating uncertainty in the temporal information.

The authors then evaluated this structure by categorizing temporal expressions and assessing how effectively their model could represent them. They also conducted inter-coder reliability tests to ensure the consistency of their approach.

Datasets

The study utilized a corpus of 200 random electronic discharge summaries from the Columbia University Medical Center data repository, spanning from 1987 to 2004. This corpus contained a total of 134,322 words. A subset of 100 summaries was used for detailed analysis and frequency tallying of temporal expression categories, and a further subset of 50 expressions was used for inter-coder agreement evaluation.

Key Insights and Ideas

  • Clinical narratives are temporally rich and complex: The analysis revealed six main categories of temporal expressions, highlighting the diversity beyond simple dates and times.
  • Relative and event-dependent expressions are prevalent: Clinical narratives heavily rely on relative time ("yesterday") and temporal relationships anchored to other events ("before admission"), which are often challenging for automated systems.
  • A structured representation is key: The proposed temporal constraint structure provides a unified and expressive way to encode these diverse temporal elements, moving beyond simple time-stamping.
  • Bridging NLP and reasoning: This structure acts as a crucial intermediate step, allowing NLP-extracted temporal information to be integrated with domain knowledge and used for higher-level temporal reasoning, such as Simple Temporal Problems (STPs).
  • Domain-specific "key events" as anchors: The study identified specific events like "admission," "discharge," and "operation" that frequently serve as anchors for temporal expressions, suggesting their importance in clinical temporal reasoning.

Key Results

  • The temporal constraint structure effectively modeled 97% (1961 out of 2022) of identified temporal expressions in the evaluated corpus.
  • Inter-coder evaluation showed high agreement: 90% exact match, 8% partial match with trivial differences, and 2% partial match with large differences. Total mismatch was 0%.
  • Six major categories of temporal expressions were identified: "date and time," "relative date and time," "duration," "event-dependent temporal expression," "fuzzy time," and "recurring times."
  • The structure successfully captured temporal information that was previously difficult to process, such as relative dates, durations, and event-anchored times.

Limitations and Future Work

The authors acknowledge several limitations:

  • Focus on historical events: The evaluation primarily focused on past events described in discharge summaries. Hypothetical events (like treatment plans) were modeled as statements of the plan rather than fully resolved temporal assertions.
  • Inter-coder reliability bias: The evaluation involved co-developers as coders, which might have inflated the agreement scores.
  • Exclusion of certain expressions: Some fuzzy time expressions (e.g., "non-specific time") and temporal adjectives/adverbs describing event characteristics (e.g., "gradual," "occasional") were not modeled due to their inability to be localized on a timeline or link events.
  • Vagueness handling: While the structure has a "vagueness" field, the initial mapping uses exact numbers, with vagueness being addressed in a later post-processing stage.

Future work suggested includes:

  • Integrating the temporal constraint structure with existing medical NLP systems (like MedLEE) to automate annotation and extraction.
  • Developing rules for resolving temporal co-references and handling vagueness more robustly, potentially leveraging textual context and domain knowledge.
  • Extending the model to handle hypothetical events and temporal discourse analysis.
  • Applying the structure to other types of clinical texts (radiology reports, pathology reports) and exploring its applicability to different clinical domains.

Summary updated

BibTeX
@article{zhou2005temporal,
  author = {Li Zhou and Genevieve B. Melton and Simon Parsons and George Hripcsak},
  title = {{A temporal constraint structure for extracting temporal information from clinical narrative}},
  journal = {Journal of Biomedical Informatics},
  year = {2005},
  doi = {10.1016/j.jbi.2005.07.002},
}
Figure