Abstract
Temporal normalization in clinical narratives is complicated by the non-linear nature of medical timelines. Unlike news articles, which predominantly anchor relative time expressions to the document creation time (DCT), clinical notes frequently reference admission dates, discharge dates, or specific procedural events. This paper proposes a normalization framework using two multi-label classifiers to identify anchor points and relations, combined with a rule-based parser. The system demonstrates significant improvements over existing methods on the 2012 i2b2 challenge dataset, achieving a relaxed normalization accuracy of 82.09%.
Introduction
To build accurate patient timelines from clinical notes, Natural Language Processing (NLP) systems must extract and normalize Temporal Expressions (TIMEXes) into a standard format, such as ISO8601. While absolute dates (e.g., "Nov 29, 2013") are straightforward, Relative and Incomplete TIMEXes (RI-TIMEXes)—such as "two days after surgery" or "at 6 AM"—require contextual clues to resolve.
Existing NLP systems usually assume these relative expressions anchor to the Document Creation Time. While this heuristic is highly effective for news articles, it fails in clinical discharge summaries where the timeline constantly shifts between admission dates, discharge dates, and various clinical events. This study addresses this gap by reframing RI-TIMEX normalization as a machine learning classification problem.
Methodology
The researchers proposed a pipeline consisting of three primary components designed to handle the unique structure of clinical narratives:
- Anchor Point Classifier: A Support Vector Machine (SVM) model that links an RI-TIMEX to a specific reference point. It classifies the anchor as either the Admission Date, Discharge Date, a Previous TIMEX, or a Previous Absolute TIMEX.
- Anchor Relation Classifier: An SVM that determines the temporal relationship between the RI-TIMEX and its identified anchor (categorized as "before," "after," or "equal/during").
- Rule-Based Parser & Integrator: This module parses the text span (e.g., converting the string "two days" to a numeric value of 2) and applies the predicted relation to the anchor point to calculate the final standardized date/time.
The system was trained using various feature sets, including bag-of-words, normalized numbers, clinical event tags, and verb tenses.
Domain Analysis and Insights
The authors conducted an annotation study across three corpora: the 2012 i2b2 Temporal Relation Challenge Corpus (clinical), TimeBank (newswire), and WikiWars (historical narratives). This comparison revealed why standard tools fail on medical text:
- Anchor Distribution: In newswire text, 96% of relative times anchor to the DCT. In clinical texts, anchors are highly distributed: 53% anchor to the admission date, 37% to a previously mentioned time, and 16% to the discharge date.
- Verb Tense Utility: While verb tense often helps predict anchor relations in general prose (e.g., past tense implying "before"), it proved useless in clinical notes. Because discharge summaries are almost entirely written in the past tense, the feature added no predictive value.
- Number Normalization: The study found that converting all numbers to a generic
<NUM>token helped the model identify anchor points but hindered its ability to find anchor relations. For example, the specific number in "day of life #1" versus "day of life #2" is the only factor determining whether the relation to the birth event is "equal" or "after."
Results
The system's performance was evaluated on a held-out test set from the 2012 i2b2 challenge:
- Classification Accuracy: The system achieved 74.68% accuracy for anchor point classification and 87.71% for anchor relation classification.
- Normalization Accuracy: Under strict matching criteria, the system achieved 57.2% accuracy. However, because clinical notes are often ambiguous (e.g., whether "post-operative day 1" refers to the day of surgery or the following day), the authors utilized a relaxed metric allowing a +/- 1 day deviation. Under this relaxed metric, accuracy rose to 82.09%.
- Comparative Performance: Compared to the top-performing systems from the original 2012 i2b2 challenge, this approach achieved a statistically significant improvement, boosting the relaxed F1-score from 45.42 to 53.77.
Conclusion
By formulating RI-TIMEX normalization as a pair of multi-label classification problems, the researchers successfully accounted for the complex anchoring patterns found in clinical narratives. While the system relies on high-quality annotations for training, its modular architecture allows for potential adaptation to other specialized domains where standard temporal heuristics are insufficient.
Summary updated