Can suicide theory-guided natural language processing of clinical progress notes improve existing prediction models of Veteran suicide mortality?

点击翻译按钮获取中文摘要

Background: Reducing suicide and suicide attempts among U.S. Veterans is a major national priority, as more than 6,000 Veterans die by suicide every year and many more attempt suicide. In 2017, the most recent year for which data are available, the suicide rate among Veterans was 1.5 times the rate of non-Veterans, and the suicide rate among female Veterans was 2.2 times the rate of non-Veteran females. Current VHA suicide risk prediction models suffer from high numbers of false negatives - Veterans not deemed at high risk of suicide who do attempt or die by suicide. These suicide prediction models have not incorporated the rich information from clinical progress notes that may improve our ability to predict suicidal behavior. Much of this information in clinical progress notes is unstructured free text. A suicide-specific ontology and information extraction system that can extract suicide-related information from unstructured clinical progress notes is not available. Significance/Impact: Enhancing VHA's ability to identify Veterans who are most likely to attempt suicide ensures that limited intervention resources can be focused on Veterans with the highest risk, before they attempt suicide or die by suicide. The proposed study is well-aligned with priorities for HSR&D research and with VA strategic goals for 2018 – 2024 set out by VA leadership, who listed suicide prevention as “VA's highest clinical priority.” Innovation: Our key methodological innovation is to pair a state-of-the-art theoretical framework (3-step Theory of Suicide) to predict who is most likely to act on their suicidal thoughts with state-of-the-art data science methods (NLP, machine learning). Since our suicide-theory concepts, that is hopelessness, connectedness, psychological pain, and capacity for suicide, are not represented in structured patient data, we will develop novel NLP and information extraction tools and apply them to clinical progress notes, the potential of which has not been fully levied to improve suicide prediction models. Specific Aims: We have three specific aims: 1. Develop a suicide-specific ontology for machine recognition of hopelessness, connectedness, psychological pain, and capacity for suicide in progress notes of clinical encounters with Veterans who attempted or died by suicide. 2. Extract information on the presence and intensity of hopelessness, connectedness, psychological pain, and capacity for suicide in clinical progress notes and describe change in these concepts in proximity of a suicide or suicide attempt. 3. Determine the predictive validity of hopelessness, connectedness, psychological pain, and capacity for suicide regarding Veteran suicide attempts and mortality in two prediction models that VA currently uses in clinical practice: STORM and REACHVET. Methodology: The proposed mixed-methods study has an exploratory sequential design where a qualitative component (Aim 1) informs quantitative analyses (Aims 2 and 3). Data collection will be from existing clinical progress notes in VHA's Corporate Data Warehouse, VA's Suicide Prevention Applications Network and from the VA/DoD Suicide Data Repository. We will use linguistic annotation and thematic analysis for Aim 1 and natural language processing and machine learning models for Aims 2 and 3. The target population is Veterans who receive care through VHA. Next Steps/Implementation: Our most important next step is to be in regular contact with local and national colleagues at the VA Office of Mental Health and Suicide Prevention (OMHSP) to facilitate implementation of our results in the operational versions of STORM and REACHVET.