Blog
Identifying Suicide Risk Based on the Words People Use
Researchers from the Child Mind Institute and MIT have developed a tool that can identify warning signs in text conversations, helping experts better understand who may need urgent support.
When someone is experiencing a mental health crisis, getting the right help quickly can be lifesaving. But it can be very difficult to determine who is in immediate danger.
New research suggests that the language people use during a crisis may offer important clues.
Daniel Low, PhD, a research scientist who leads the AI, Risk, and Contemplative Science (ARC) Lab at the Child Mind Institute, helped develop a tool that analyzes text conversations for language connected to known suicide risk factors. Dr. Low conducted the work while he was a graduate student at Harvard and MIT, in collaboration with Satrajit Ghosh, PhD, a senior research scientist and director of the Open Data in Neuroscience Initiative in the McGovern Institute at MIT.
The hope is that tools like this could one day help clinicians and crisis counselors spot warning signs more quickly. This way, they’ll have a better understanding of who may need immediate support.
Learning from real crisis conversations
Much of what researchers know about suicide risk comes from surveys asking people to reflect on their thoughts and experiences after a crisis has passed. But memory is not always complete, and a person’s experience may feel different once the immediate crisis is over.
To study what happens in the moment, the researchers partnered with Crisis Text Line, a nonprofit organization that provides confidential mental health support by text message. Following specialized training, the researchers were given protected access to a de-identified dataset of about 16,000 conversations. The data provided a rare opportunity to study what people were saying while actively seeking help during a crisis.
The conversations were grouped into three severity levels based on Crisis Text Line’s risk assessments: non-suicidal, suicidal ideation without imminent risk, and imminent risk. The imminent-risk group included conversations involving a plan for suicide within a 48-hour timeframe, immediate danger, or emergency service intervention.
To analyze the conversations, the team first built a library of words and phrases linked to established suicide risk factors. They used artificial intelligence to generate an initial list, then worked with expert clinicians to review and validate the results. The final lexicon contains thousands of terms spanning 49 suicide risk factors.
Researchers then trained a computer model to identify those patterns in the crisis conversations and estimate risk levels. And because the model is built on the lexicon, it can show which language contributed to a particular assessment. This allows counselors and clinicians to review the evidence behind a result rather than relying solely on a score from a computer — an important feature when decisions involve a person’s safety.
“This is such a complex space that having a human in the loop is, I think, going to be critical for a long, long time,” says Dr. Ghosh.
What the researchers found
Language related to lethal means, substance use, and active thoughts of suicide or self-harm appeared more often in imminent risk conversations. References to anxiety, post-traumatic stress disorder (PTSD), and emotional pain were also associated with risk, but were not among the highest-risk group. And while depression is a well-known risk factor for suicidal ideation, the model found expressions of depressed mood and fatigue were less closely associated with imminent risk than some of the more direct warning signs.
It’s important to note that this does not mean depression or any single warning sign is unimportant. It also does not mean that a word or phrase can predict what one person will do. Instead, the findings show which patterns of language were most strongly associated with the highest-risk conversations in this dataset.
What this could mean for mental health care
This tool is still in the research stage and requires further testing before it could be used in clinical settings.
Still, the work offers a new way to study how distress appears in language. By examining conversations that occur during a crisis rather than after the fact, researchers can gain a clearer picture of the thoughts, feelings, and experiences associated with different levels of risk.
The potential applications extend beyond crisis conversations. The suicide-risk lexicon is already being used to explore how language from sources such as social media and electronic health records might help researchers better understand and estimate suicide risk in other settings. The team has also made the lexicon and software publicly available so other researchers can build on the work and explore additional applications across a range of mental health conditions.
Dr. Low’s work is part of a broader effort at the Child Mind Institute. Through projects such as Mirror Journal, researchers at the Child Mind Institute are exploring how thoughtfully designed digital tools can help young people express what they are experiencing, process their emotions, and connect with support when they need it.
The findings were published in the Journal of Psychopathology and Clinical Science. Read the full paper.
The Suicide Risk Lexicon and tutorials to create lexicons are available here.
If you or someone you know is in immediate danger, call 911.
In the United States, call or text 988 to reach the 988 Suicide & Crisis Lifeline. The Crisis Text Line can be reached by texting HOME to 74174.
