IIR 12-084
Current Evidence and Early Warning Indicators of Homelessness Risk Among Veterans
Adiseshu V. Gundlapalli, MD PhD VA Salt Lake City Health Care System, Salt Lake City, UT Salt Lake City, UT Funding Period: December 2013 - May 2017 Portfolio Assignment: Healthcare Informatics |
BACKGROUND/RATIONALE:
Estimates of Veterans who are homeless are based on those who are currently receiving, have previously received, or are currently being directed to specific VA homeless services. Those 'at-risk' of homelessness, especially for the first time, are a major focus of VA prevention. Early warning indicators to identify these Veterans are currently only inferred from known risk factors for homelessness that can be gleaned from administrative data. However, references to indicators of risk in the free text of clinical narratives written by VA providers may precede the formal identification of Veterans as being homeless and potentially represent an untapped resource for early identification. This project builds on the work performed under HIR 10-002, Pro-WATCH: Homelessness as Sentinel Event. OBJECTIVE(S): The objectives of this project were to accomplish the following: (1) develop robust electronic algorithms to identify Veterans who are currently homeless or those who are at risk of homelessness using predictive modeling; (2) review the extent to which Veterans are receiving appropriate services. METHODS: The first aim of this project used text data to improve the accuracy of determination of homelessness status of Veterans. To accomplish this we first identified gaps in administrative criteria that had been used to define homelessness in the VA. We then scaled up natural language processing (NLP) algorithms and applied them to the entire population of Veterans seen within VA in FY 2013 and 2014. We first identified those with administrative evidence of homelessness and then performed NLP on 12 million TIU notes to refine the estimate of homelessness based on mentions of homelessness in the written text. The second aim of this project was to develop and apply predictive models for homelessness and homelessness outcomes in Veterans. We used both NLP processing of text and all available administrative data to develop algorithms that were predictive of future homelessness. These algorithms ranged from simple to complex involving machine learning. We focused our work on OEF/OIF/OND Veterans. Finally, we applied our prediction model to one year of data (2/01/16 - 1/31/17) to discover at-risk Veterans who had been seen at the Salt Lake City VA. FINDINGS/RESULTS: We have completed both aims of the project. We have successfully worked towards our goal of finding the simplest algorithm for identifying Veterans seen in VA facilities who are at increased risk for homelessness. In general, we have demonstrated markers based on adverse experiences in the military. We have explored factors associated with homelessness and access to care for women Veterans and developed informatics and epidemiologic methods for analyzing VA big data to refine estimates of homelessness among Veterans. 1. At the time of separation from the military: it appears that separating from the military for "misconduct" leads to 5-7 fold increase in risk of homelessness thus underscoring the need for active case management for selected groups at the time of transition from the military to civilian life (published in JAMA, 2015). 2. History of military sexual trauma (MST): We found that rates of homelessness among Veterans with a positive screen for MST were more than double the rates of Veterans with a negative screen for MST. In addition, male Veterans with a positive screen for MST were found to be at greater risk for homelessness than female Veterans with a positive screen for MST. These results underscore the importance of the MST screen conducted when a Veteran is seen VHA as a clinically important marker of reintegration outcomes among Veterans (published in JAMA Psychiatry, 2016) 3. Rural residence: We found that Veterans living in a rural area, those living between 20 and 40 miles, and 40+ miles away from the nearest VAMC were at a lower risk for homelessness (Housing, Care and Support, Vol. 20 Issue: 2, pp.45-59, https://doi.org/10.1108/HCS-10-2016-0013). 4. Service connected disability: During the first year of VHA service usage, higher levels of disability benefits were protective against homelessness among routinely-discharged Veterans, but not among disability-discharged Veterans. By 5-years, disability discharge was a risk factor for homelessness. 5. Women Veterans and access to reproductive health services (long acting reversible contraceptives): For the first time, we demonstrated that VA is successfully engaging homeless women Veterans and providing LARC access. The prevalence of perinatal risk factors in ever-homeless women Veterans highlights a need for further programmatic enhancements to improve reproductive planning (published). 6. Using social network analysis techniques for exploring patterns of homelessness among Veterans: We determined that network visualizations of large clinical datasets reveal rich, previously hidden connections between data variables related to homelessness. (published). 7. Machine learning (ML) methods to predict homelessness: We trained an ML algorithm using data from FY2013 and FY2014, using all available administrative (structured data) and relevant concepts extracted from the written notes using NLP to predict homelessness. In applying this algorithm to local VA SLC data, we were able to accurately predict homelessness status in 88% of cases, based on administrative evidence of homelessness. A human review of TIU notes from a sample of 24 Veterans who were at "very high" risk for homelessness by the algorithm and were not known to be homeless by administrative criteria revealed that eight (33%) were actually homeless. (manuscripts in preparation) 8. Using text data to improve the accuracy of determination of homelessness status of Veterans: We have shown that extracting concepts related to homelessness allows us to refine estimates of homelessness that are derived from administrative data (published and unpublished). IMPACT: Ending and preventing homelessness among Veterans are high priority areas for VA. We expect our research to have a positive impact on managing and preventing homelessness among Veterans who seek care within the VA system. When implemented, the electronic algorithms developed and validated in this research proposal will have direct impact on (1) identifying Veterans at risk for homelessness and (2) planning and provision of resources and services to prevent and end homelessness among Veterans seen in VA. This work will help answer epidemiological questions that need to be addressed with regard to homelessness and those at risk among returning OEF/OIF/OND populations. In addition, this work will contribute to the important informatics methodological domain of NLP and machine learning. External Links for this ProjectNIH ReporterGrant Number: I01HX000990-01A2Link: https://reporter.nih.gov/project-details/8597106 Dimensions for VADimensions for VA is a web-based tool available to VA staff that enables detailed searches of published research and research projects.Learn more about Dimensions for VA. VA staff not currently on the VA network can access Dimensions by registering for an account using their VA email address. Search Dimensions for this project PUBLICATIONS:Journal Articles
DRA:
Substance Use Disorders, Health Systems Science
DRE: Prognosis, Diagnosis, Prevention Keywords: none MeSH Terms: none |