Decoding Mental Health Center
Suicide Risk Classification

Suicide Risk Classification Program

Every 11 minutes, someone in the US intentionally ends their life. Suicide is the second-leading cause of death for 10- to 34-year-olds. Suicide is the fourth-leading cause of death for those between 35 and 44 and the fifth-leading cause of death for those between the ages of 45 and 54. If we can identify suicide thought markers early, we can have a chance to prevent it.

Thought markers provide insight into what someone is thinking. They include body language, spoken and written words, and vocal tones. They are helpful in many ways—one being revealing a person’s suicidal tendencies.

We use artificial intelligence (AI) to deeply analyze what people say and how they say it. It’s our goal to use specialized data mining techniques to create evidence-based risk assessment tools that can predict suicide early and allow for positive intervention.

Our Aspirations

Our overarching goal is to cure, prevent or radically improve the treatment of pediatric and adolescent mental illness. Through our research, we hope to discover ways to:

  • Identify children at high risk for mental illness as early as possible
  • Apply evidence-based interventions to prevent suicide

Success in our work will be achieved when the rate of child or adolescent suicide decreases substantially.

The Language of Suicide

Language encompasses acoustics, facial and body expressions, and linguistics. All these expressions are markers of a person’s thoughts. Measuring thought markers (Tm) is a challenge. We can’t use the same techniques we use with DNA and other biological markers.

Our partnership with the Oak Ridge National Laboratory in Tennessee gives us access to the nation’s most powerful computer systems. Together, we lead multiple studies that incorporate complex analyses from several datasets.

Our research has included multiple unique and impactful studies including:

Development of a suicide note database. Starting with a collection of 1,300 notes, we annotated, digitized and established vectors. We identified emotions in the language used and detected signals.

Determining classification signals in spoken words. We set out to prove that by using language, subjects can be classified as either suicidal or control. We established interview questions that led to subject responses with a 93% classifier accuracy.

Studying suicide thought markers (STM). We analyzed linguistics, acoustics, facial expressions and genetics data from 379 emergency department and outpatient subjects of all ages from three sites. Results showed that subjects can be classified into three groups: suicidal, mentally ill but not suicidal, or controls. This study reveals how advanced technology can be used for suicide assessment and prevention. For example, pauses in a person’s speech were longer for those identified as suicidal. Vowel spacing—the space a vowel takes up in a spoken word—is smaller in people with certain mental illness.

Associated Research Labs