Reformatting clinical records based on global layout statistics - Archive ouverte HAL Accéder directement au contenu
Communication Dans Un Congrès Année : 2014

Reformatting clinical records based on global layout statistics

Résumé

Texts in electronic medical records often keep traces of the layout used to display or print them in hospital information systems. This includes blank lines in double-spaced text or short lines in wrapped paragraphs. Directly inputting this format to an information extraction system may prevent it from optimally detecting entities or relations. We address this problem with a method of text normalization based on global statistics on the layout of a text. Our goal is to obtain reformatted texts in which sentences are never split across two lines. We present two evaluations: (i) intrinsic, which obtains very good results, with all double-spaced texts correctly detected, and wrapped line detection scores of R = :9877, P = :9434, F = :9651; (ii) extrinsic, which shows a small impact of text reformatting on two information extraction tasks: de-identification and risk factors identification (i2b2/UTHealth 2014 challenge). Precision is consistently improved at the expense of recall. The overall improvement on F-measure is small (+:2pt for PHI) or null (for risk factors). Closer examination for risk factors shows that it is negative (􀀀:2pt F) for singled-spaced texts which incur wrapped lines, but positive (+:7pt F) for the larger set of double-spaced texts in our corpus.
Fichier non déposé

Dates et versions

hal-01831245 , version 1 (05-07-2018)

Identifiants

  • HAL Id : hal-01831245 , version 1

Citer

Pierre Zweigenbaum, Cyril Grouin. Reformatting clinical records based on global layout statistics. International Symposium on Semantic Mining in Biomedicine, Jan 2014, Aveiro, Portugal. ⟨hal-01831245⟩
76 Consultations
0 Téléchargements

Partager

Gmail Facebook X LinkedIn More