Implementation of a Text-Mining-Based Clinical Decision Support ...

skirlorangeBiotechnology

Oct 1, 2013 (4 years and 3 months ago)

96 views

Implementation of a Text
-
Mining
-
Based Clinical Decision Support Tool to Improve the Management of
Pulmonary Nodules


Vijay Garla
1
, Steven Steinhardt
2
,3
, Forest Levin
2
,3
, Pradeep Mutalik
3,4
, Cynthia Brandt
3,4
, Caroline Taylor
3

1.

Interdepartmental Program in
Computational Biology & Bioinformatics, Yale University, New
Haven, CT

2.

Evergreen Design, Guilford, CT

3.

Connecticut VA Healthcare
System,
West Haven,
CT

4.

Section of Medical Informatics, Yale University School of Medicine, New

Haven,
CT


Objectives:

To identi
fy potentially malignant lung nodules in radiology reports to ensure their management
according to clinical practice guidelines.


Methods:

The West Haven VA has adopted the guidelines of the Fleischner society for the management of
pulmonary nodules. Radi
ologists currently manually code as Cancer Alerts reports with lung nodules
that require surveillance. Cancer alerts are forwarded to cancer care coordinators who manage
surveillance and treatment of the nodules. Internal audits have shown undercoding of

cancer alerts,
demonstrating the need for automated coding. We developed a text
-
mining system based on open
-
source Natural Language Processing tools to automatically code cancer alerts. The system extracts
information from radiology reports, applies dec
ision rules that represent the clinical practice guidelines
to the extracted data, and flags reports with mentions of nodules that require surveillance. The system
then forwards flagged reports to cancer care coordinators for review.


Results:

We applied
the system to all chest CT reports from the West Haven VA from October 2008 to August
2010

(8360 reports)
. We manually reviewed a subset of these reports and estimated the precision
(positive predictive value), recall (sensitivity), and F
-
Score of the sys
tem as 90%, 95%, and 92%
respectively. Common manual coding errors included undercoding of small nodules, stable nodules, and
ground
-
glass opacities. The system identified 710 patients for which no report had previously been
coded as a cancer alert.


Con
clusions:

The text
-
mining system we developed accurately identified lung nodules from radiology reports that
require follow
-
up according to clinical practice guidelines.


Impact Statement:

Early detection of lung cancers significantly improves outcome and
survival. The text
-
mining system we
developed will improve the surveillance of lung nodules, facilitating early detection and effective
treatment of cancerous lesions.