UCLA researchers lead study that suggests new methods of analyzing existing records could provide evidence of risk before a suicide

A new study of suicides in the U.S. has found that current national reporting on these deaths underestimates the extent of “emotional dysregulation,” meaning the emotional distress that occurs before suicide, possibly providing a method to prevent future deaths.

Overall, researchers said that indicators of clinically relevant emotional distress may occur in 90% of suicides but are greatly underestimated by current measures. Using large language model (LLM) analyses of existing reports by emergency response personnel and other non-clinicians, they identified this previously underecognized effect. Use of such LLMs in the future could lead to earlier interventions, including potentially preventing both suicide attempts and resulting deaths.

“Suicide is a major contributor to early mortality, particularly among those 15 to 45 years of age; it is the second leading cause of death for men and fourth for women in the United States,” said Dr. Vickie Mays, a study co-author, clinical psychologist, and professor at UCLA. “Yet estimates drawn from the nation’s major tracking source for violent death suggest that perhaps less than half of suicide victims have a mental health disorder at the time of their death, and less than a third are described as being known to have been depressed before they commit suicide.”

The peer-reviewed study – “Research Domain Criteria and Deaths by Suicide in the National Violent Death Reporting System” – is included in the April edition of JAMA Network, published by the American Medical Association. The authors, from UCLA’s Fielding School of Public Health and Samueli School of Engineering, as well as Purdue University, analyzed 72,585 suicides drawn from the U.S. National Violent Death Reporting System (NVDRS).

“We used a large language model to detect emotional dysregulation among people who die by suicide,” said study co-author Dr. Susan Cochran, a psychologist and epidemiologist with UCLA’s Fielding School of Public Health. “While that may seem obvious, that most people who die this way are going through a great deal of upset, current methods of measuring mental health disorders in this population are undercounting the problem because they rely on people having diagnosed mental health disorders, and these disorders are very much undertreated.”

Currently, the NVDRS codes for psychiatric diagnoses that require specific symptoms; the researchers wanted to see if an analysis of the same records, provided by state and local law enforcement, coroners, and medical examiners as short narrative summaries, could detect unrecognized patterns of neurobehavioral dysfunction by using an LLM. These methods are not currently used in the NVDRS, the researchers said.

“We came at it a different way, and our results confirm expectations that suicide is accompanied by high levels of emotional dysregulation for most people, especially younger people, and women,” Cochran said. “The intensity of the dysregulation is consistent with people admitted for inpatient psychiatric treatment.”

The researchers dug into the existing databases, using standardized LLMs to go through the reports, and looking for specific language in the state and local narrative reports that go beyond the psychiatric diagnoses. The approach used NIH’s Research Domain Criteria (RDoC) framework to score mental health status indicators in the records.

“What we found is that using the RDoC framework provided better information than focusing on clinical diagnosis alone,” said Dr. Bruce Cuthbert, former acting director of the National Institute of Mental Health (NIMH) and a study co-author. “The approach used here observed more dysfunction among suicide decedents than that captured by the currently employed NVDRS measures of mental health.”

In turn, further analysis of the data found that among the 72,585 deaths, populations with significant risk factors for suicide include:

• Men (80% of the deaths)
• non-Hispanic whites (79%)
• middle-aged individuals, with a mean age of 46.3 years
• Those without a significant partnership, meaning marriages, domestic partnerships, etc. (58%)
• 32% lived in the southern U.S., and
• 19% of male decedents were military veterans.

In addition, those with evidence of any mental disorder numbered 44%, while problematic alcohol and/or drug use was prevalent among 27% of the victims. Only 28% had been classified as depressed at the time of death, and there were also significant differences between men and women, researchers said.

“Females are more likely than males to have a mental health problem and males are more likely to have an alcohol and/or drug use problem,” said study co-author Dr. Alina Arseniev-Koehler, an assistant professor of sociology at Purdue University who received her doctorate at UCLA. “However, females were no more likely than males to be depressed according to NVDRS coding, which overall, reinforces the concept that mental health diagnoses alone don’t bring out the full picture.”

Ultimately, the researchers think this work could lead to more successful interventions, before someone attempts suicide.

“This is a research project, to see what we can learn from existing data using these methods, both the RDoC framework as a model and the use of LLMs as a tool,” Mays said. “We’d like to grow this into a pilot project, focused on the high-risk populations, to see if the same methods can be used to prevent someone from doing harm to themselves.”

Methods

The authors drew on records drawn from the U.S. NVDRS (maintained by the U.S. Centers for Disease Control and Prevention) and used a standardized large language model (LLM) to find indicators of neurobehavioral dysfunction, beyond the current measures in the NVDRS. The results were then compared to those obtained by the standard measures, revealing patterns of dysfunction more intense that previously known, especially among women and younger individuals.

Funding

The work was supported, in part, with funding from the U.S. National Institutes of Health (NIH) through a National Institute on Mental Health grant (MH115334) to UCLA faculty and a National Library of Medicine Training Grant (T15LM011271) to UCSD faculty where Dr. Arseniev-Koehler did her postdoctoral training. The funders had no role in the design and conduct of the study; collection, management, analysis, and interpretation of the data; preparation, review, or approval of the manuscript; and decision to submit the manuscript for publication.