Suicide is a serious public health issue. Understanding the full picture of suicide risk and mental health is something researchers have long struggled with. New findings published in JAMA Network Open (2026) are beginning to change that. Using advanced technology, the study uncovers the true depth of psychological suffering behind suicide deaths in the United States. The findings are sobering and, in many ways, a call to action.

Why Suicide Risk and Mental Health Data Has Always Fallen Short

The National Violent Death Reporting System (NVDRS) is the United States’ primary source for tracking violent deaths, including suicide. Yet data from this system has long suggested something that seemed at odds with clinical experience. Fewer than half of those who die by suicide have a recorded mental health disorder. Fewer than a third are described as depressed at the time of death.

That does not match what psychological autopsy studies have found. Those studies reconstruct a person’s mental state from interviews and records after death. They show that more than 87% of people who die by suicide had a concurrent mental health or substance use disorder.

The gap between these two figures points to a significant undercounting problem. Understanding why it happens matters enormously for suicide prevention.

A New Lens: Research Domain Criteria

Researchers from UCLA took a different approach. Rather than relying solely on clinical diagnoses in the NVDRS, they used the Research Domain Criteria (RDoC) framework. Developed by the National Institute of Mental Health, RDoC looks at psychological functioning across six broad domains:

• Negative valence (distress, hopelessness, anxiety)
• Positive valence (motivation, reward, substance use patterns)
• Social processes (relationships, belonging)
• Arousal processes (agitation, sleep disturbance)
• Cognitive systems (attention, memory, decision-making)
• Sensorimotor systems

Rather than asking “does this person have a diagnosis?”, RDoC asks: “in what ways was this person’s psychological functioning disrupted?”

To extract this from NVDRS death narratives, the researchers applied two machine learning methods. One was a token-based scoring system. The other was a large language model (LLM), the technology behind modern AI tools. Both had previously been validated against psychiatric inpatient records.

What the Research Found About Psychological Dysfunction and Suicide

The study analysed death records for 72,585 people who died by suicide in 2020 and 2021. These came from all 50 US states. The results were striking.

Using the LLM scoring method, more than 90% of suicide decedents showed at least one clinically significant RDoC domain score. This means evidence of dysfunction serious enough to require treatment. It was true in both law enforcement and coroner narratives.

Compare that to what the NVDRS had recorded: only 44.4% with any mental health disorder and only 27.9% described as currently depressed.

The domains most frequently elevated were negative valence and arousal processes. These capture hopelessness, distress, anxiety, and agitation. These are emotional states that do not always lead to a formal diagnosis. Yet they are deeply relevant to suicide risk and mental health outcomes.

Female decedents and younger decedents showed consistently higher levels of dysfunction across most domains. Among younger adults aged 25 to 44, clinically relevant arousal process dysfunction appeared in around 65% of law enforcement narratives. Among those aged 65 and over, this figure dropped to around 41%. Even so, dysfunction remained widespread in that older group.

Substance Use and Psychological Dysfunction and Suicide

One finding deserves particular attention. The RDoC framework links positive valence dysfunction directly to substance use patterns and their effects on reward processing. This dysfunction was significantly more common among decedents than standard NVDRS alcohol and drug measures suggested.

The standard NVDRS measure recorded problematic alcohol or drug use in 27.5% of decedents. But RDoC positive valence dysfunction, which captures disrupted reward and motivation, showed clinically relevant levels in around 31 to 41% of decedents.

Substance use does not just create health risks in isolation. It fundamentally alters how people experience reward, motivation, and relief from distress. It reshapes emotional life in ways that heighten vulnerability. That connection is essential to understand.

Why This Changes the Conversation

The traditional approach to suicide prevention has often focused on identifying people with a clinical diagnosis. If someone has a recorded diagnosis of depression or an anxiety disorder, systems are more likely to flag them for intervention. But a large proportion of people who die by suicide do not have that flag.

The RDoC approach offers a different way in. By looking at how someone is actually functioning, it is possible to detect risk that a diagnosis alone would miss.

This is especially relevant for men. Men made up 80.6% of decedents in this study. They were consistently less likely than women to have formal mental health diagnoses recorded. The suffering was still there. It was simply less likely to have been named.

The Role of Language and Narrative

One of the more remarkable aspects of this research is what it reveals about the value of narrative. The NVDRS death records include written accounts from law enforcement officers and medical examiners. These narratives have been largely underused in research.

These accounts contain substantial information about psychological state. Even without formal clinical language, descriptions of someone’s behaviour and emotional expression carry real signal. With the right analytical tools, that signal can be read.

This has implications for how we think about documentation and early intervention systems. Information that seems peripheral or informal may carry more weight than we realise.

Suicide Risk and Mental Health: What We Still Need

This research is a proof of concept. The authors note that accuracy could be improved with more sophisticated prompting methods or by comparing results across multiple AI systems. There are also inherent limits to postmortem records. Toxicology results were absent in roughly half of cases. Reporting consistency varied across states.

What is clear is that the mental health burden surrounding suicide has been systematically underestimated. Formal diagnosis codes tell only part of the story. The full picture requires attending to the dimensions of human distress that do not fit neatly into diagnostic boxes.

That full picture is where intervention lives.

Source: jamanetwork