Lead/Presenter: Adi Gundlapalli, COIN - Salt Lake City
All Authors: Pettey WB (COIN – Salt Lake City)
Todd D (COIN – Salt Lake City)
Redd A (COIN – Salt Lake City)
Carter ME (COIN – Salt Lake City)
Samore MH (COIN – Salt Lake City)
Gundlapalli AV (COIN – Salt Lake City)
Objectives:
We set out to evaluate the utility of a network projection data exploration approach to solving a clinical/operational challenge of providing cognitive support to finding patterns in health care data that could be exploited to prevent homelessness among Veterans.
Methods:
We split Veteran ICD-9-CM data into two time periods (0-59 and 60-364 days prior to the first evidence of homelessness) and then used Pajek social network analysis software to visualize these data as three different networks. A multi-relational network simultaneously displayed the magnitude of ties between the most frequent ICD-9-CM pairings. A new association network visualized ICD-9-CM pairings that greatly increased or decreased. A signed, subtraction network visualized the presence, absence, and magnitude difference between ICD-9-CM associations by time period.
Results:
A cohort of 9,468 Veterans was identified as having administrative evidence of homelessness and visits in both time periods. They were seen in 222,599 outpatient visits that generated 484,339 ICD-9-CM codes (average of 11.4 (range 1-23) visits and 2.2 (range 1-60) ICD-9-CM codes per visit). Using the three network projection methods, we were able to show distinct differences in the pattern of co-morbidities in the two time periods. In the more distant time period preceding homelessness, the network was dominated by routine health maintenance visits and physical ailment diagnoses. In the 59 days immediately prior to the homelessness identification, alcohol related diagnoses along with economic circumstances such as unemployment, legal circumstances, along with housing instability were noted.
Implications:
Network visualizations of large clinical datasets traditionally treated as tabular and difficult to manipulate reveal rich, previously hidden connections between data variables related to homelessness. A key feature is the ability to visualize changes in variables with temporality and in proximity to the event of interest. These visualizations lend support to cognitive tasks such as exploration of large clinical datasets as a prelude to hypothesis generation.
Impacts:
Network projections of VA medical diagnosis codes data for Veterans who were eventually identified as homeless reveal considerable changes in diagnosis code co-incidence, transitioning from unremarkable and typical medical visits to those consistent with substance abuse, PTSD, depression, and economic hardship.