Electronic Surveillance of Invasive Fungal Diseases in Hematology-oncology Patients using Natural Language Processing of Computed Tomography Reports: Towards Real-time Prospective surveillance

Ref ID: 18795

Author:

M. Ananda-Rajah, FRACP – Infectious diseases physician1, D. Martinez, PhD – Research Fellow 2, H. Suominen, PhD – Research Fellow 3, K. Thursky, FRACP, MD – Infectious diseases physician 4, M. Slavin, FRACP, MD – Infectious diseases physician 4, L. C

Author address:

1Alfred Health, Melbourne, Australia, 2NICTA, Melbourne, Australia, 3NICTA, Canberra, Australia, 4Melbourne Health, Melbourne, Australia.

Full conference title:

52nd Annual ICAAC

Date: 9 September 2014

Abstract:

Background: Continuous prospective surveillance of invasive fungal diseases (IFDs) in hematology-oncology patients should be the standard of care but is unfeasible for many institutions due to resource limitations and an absence of, in the case of invasive aspergillosis, an easily identifiable electronic trigger. Here we report natural language processing (NLP) of computed tomography (CT) reports as a means of IFD detection. Methods: CT scan reports were collected for IFD-case and uninfected control patients identified from completed studies performed at Alfred Health, Melbourne Health, and Peter MacCallum Cancer Institute from 2003 to 2010. IFDs were confirmed using EORTC/MSG (2008) criteria. An expert annotated subset was used to build a machine learning system for IFD detection at individual scan and patient level i.e. an episode of care inclusive of index admission and 12-weeks thereafter. Results: We collected 2129 CT reports from 288 IFD-case and 291 control patients comprising, proven/probable & possible IFDs in 151/288, 52% & 137/288, 48% respectively involving predominantly pulmonary (232/288, 81%) and sinus (25/288, 8.7%) sites. Mold infections either suspected or microbiologically confirmed, accounted for 178/288 (62%). The annotated subset comprising chest/sinus reports from 398 case and 83 control patients with 10-fold cross-validation resulted in a sensitivity/specificity/positive/negative predictive value (%) for IFD-detection at scan & patient level of 93.6/82/89.2/89 and 98.2/61/77.5/96.2 respectively. Earlier detection than expert annotators occurred in 10% of cases. Conclusion: NLP of CT-scan reports may enable sustainable contemporaneous IFD surveillance with potential to define local epidemiology, expedite outbreak detection and facilitate intra and inter-facility comparisons. Acknowledgment: NICTA is federally funded

Abstract Number: M-315

Conference Poster: y

Conference Year: 2012

Link to conference website: NULL

New link: NULL

Conference abstracts, posters & presentations

Showing 10 posts of 17325 posts found.

  • Title

    Author

    Year

    Number

    Poster