A Hybrid System for Temporal Relation Extraction from Discharge Summaries - Technologies and Applications of Artificial Intelligence - page 377

A Hybrid System for Temporal Relation
Extraction from Discharge Summaries
Yueh-Lin Yang 1 ,Po-TingLai 2 , and Richard Tzong-Han Tsai 3 ,
1 Department of Computer Science and Engineering, Yuan Ze University,
Taoyuan, Taiwan
[e-mail]
2 Department of Computer Science, National Tsing-Hua University, HsinChu, Taiwan
[e-mail]
3 Department of Computer Science and Information Engineering, National Central
University, Taoyuan, Taiwan
[e-mail]
Abstract. Automatically detecting temporal relations among
dates/times and events mentioned in patient records has much potential
to help medical staff in understanding disease progression and patients
response to treatments. It can also facilitate evidence-based medicine
(EBM) research. In this paper, we propose a hybrid temporal relation
extraction approach which combines patient-record-specific rules and the
Conditional Random Fields (CRFs) model to process patient records. We
evaluate our approach on i2b2 dataset, and the results show our approach
achieves an F-score of 61%.
Keywords: discharge summaries, conditional random fields, temporal
relation extraction.
1 Introduction
Temporal information extraction (TIE), an important task in natural language
processing (NLP), is the extraction of temporal relations among the events and
dates/times found in plain text. TIE has been used for a variety of NLP appli-
cations in several domains. It is especially useful for processing patient records
in the clinical domain. To improve TIE in the clinical domain, the Sixth Infor-
matics for Integrating Biology and the Bedside (i2b2) NLP challenge [7] invites
participants to develop TIE systems for processing patient records. The chal-
lenge released a dataset, which contains the TLINK, EVENT and TIMEX3 tags
defined in TimeBank [4], and proposed three tracks: (1) EVENT/TIMEX3 track:
recognize event and time expressions; (2) TLINK track: identify the temporal
relations of the given EVENT and TIMEX3 tags; (3) End-to-End track: perform
the above two tasks on raw discharge summaries. In this paper, we will focus on
the TLINK track.
Corresponding author.