Machine Learning Strategies for Predicting the Risk of Suicide Using Clinical Note Text
Abstract
Preventing death by suicide is a national imperative. According to the Center for Disease Control and Prevention, suicide was the tenth-leading cause of death in the United States in 2018, claiming over 45,000 American lives. Veterans make up a substantial proportion of all suicide deaths in the U.S., accounting for approximately 18% of suicide deaths. The risk for suicide is 22% higher in the Veteran population compared to the general U.S. population. Early prediction, however, can make a difference. The development of effective tools that identify risk for suicide is vital to ensure that individuals in need receive prompt life-saving support. Given these high rates of Veteran death by suicide, the U.S. Department of Veterans Affairs (VA) has prioritized improving suicide prediction. As a means to establish more effective suicide risk prediction, the VA recently developed a prediction method that evaluates VA users’ electronic medical record (EMR) for suicide risk. Even with this innovation, however, detecting individuals who are at risk for suicide remains a major challenge. The VA’s current model automatically evaluates all users based on established risk variables. One of the problems with this method is that there are many relevant risk variables that are not uniformly measured or even assessed and therefore not included within risk evaluation. Previous research indicates that analysis of mental health providers’ clinical notes provides useful information that is not always present in numeric variables. Whereas the VA’s current suicide prediction method evaluates risk variables related to users’ demographics and health care usage, the proposed study evaluates a comprehensive network of nuanced psychological variables associated with suicide. We are able to develop these variables by utilizing Natural Language Processing (NLP), a machine learning technique that analyzes semantic patterns in written text. As VA providers write notes summarizing each patient encounter, each user’s notes spans their treatment history. Our proposed study uses NLP to analyze providers’ notes, scanning them for variables associated with suicide. This work builds on our previous research that demonstrated the feasibility and added accuracy of this innovative approach when compared to the VA’s currently used method. Within our proposed study, we will develop a dataset of VA user’s that received mental health services between 2015-2018 and died by suicide (over 25,000) and then match these individuals with a larger dataset of VA users (over 150,000) that received mental health services but did not die by suicide. These two groups will be matched on the VA’s currently used numeric suicide risk variables. We will then extract the groups’ mental health provider notes and analyze the two groups’ semantic differences. As we controlled for variables used by the VA’s current risk prediction method, any increase in predictive accuracy will be over and above current standards. Harnessing the most recent advances in machine learning, we will select the most predictive semantic variables and develop a reproducible model that can be used broadly within the VA. Moving forward, this improved model could be integrated within the VA’s suicide risk alert system, directing providers to target those with increased need. As death by suicide is a very rare event, it takes a large treatment population to accurately study suicide risk. Although we have demonstrated the feasibility and efficacy of our proposed model with a smaller VA population (250 deaths by suicide compared to 2,000 non-deaths by suicide), including a larger sample will be allow us to develop a model that is representative of VA users nationwide. The proposed study facilitates a leap forward in suicide research, leveraging advances in machine learning to extract clinically relevant information from provider’s written records. This research will improve suicide prediction, allowing the utiliz
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Jan 04, 2024
- Source ID
- HT94252310267
Entities
People
- Jiang Gui
Organizations
- Dartmouth College
- United States Army