aamc.org does not support this web browser.
  • AAMCNews

    Progress in predicting suicide

    For decades, the goal of identifying those most at risk of ending their lives has remained elusive. Finally, it may be within sight.


    Two years ago, on the day Sean Petro was to have begun his fourth year of medical school, he ended his life instead. There had been few signs, none clear. He was quieter during his third year of school, says his mother, Cheryl Collier, and looking back, she thinks her son might have been making his goodbyes during their last lunch together. Collier says she desperately wishes she had known her son was suffering. Instead, she says, she was “totally blindsided.” 

    “I think being able to predict that someone might die by suicide would be fantastic — and the earlier the better so that parents could intervene by getting help for their children,” says Collier.  

    Sean Petro is among the nearly 7% of medical students and residents who have struggled with suicidal thoughts and the thousands of people who die by suicide each year. In fact, nearly 45,000 Americans age 10 or older died by suicide in 2016, the last year for which there are statistics. And in June, the Centers for Disease Control and Prevention issued sobering data: Since 1999, suicide rates have risen in nearly every state and have climbed more than 30% nationwide.

    “Most physicians who have had a patient who died by suicide have wondered what signs they missed, and how the death could have been prevented,” says Jane Pearson, PhD, chief of Suicide Preventive Interventions Programs at the National Institute of Mental Health (NIMH). “What makes suicide so difficult to predict is that it doesn’t have just one cause, [and it] may have different causes in children, in men, in women, and in different ethnic groups.”

    But now, experts say, researchers are edging closer to the elusive goal of predicting which people may be most at risk of taking their own lives.

    Past predictions, future hopes

    Currently, physicians use various methods to identify suicidality, including asking patients and family members questions that cover such concerns as emotional stresses, feelings of hopelessness, and access to lethal means.

    But existing approaches aren't very effective. In fact, according to a 2016 meta-analysis published in the journal PLOS One, half of suicides are among patients who had been deemed at lower risk, and 95% of patients considered at high risk did not ultimately die by suicide. What's more, the paper reported no improvement in suicide prediction in the last 40 years. 

    Research suggests that providers could reach many suicidal individuals. A 2015 study found that 64% of people who attempted suicide had a health care visit in the previous month. 

    “Most physicians who have had a patient who died by suicide have wondered what signs they missed, and how the death could have been prevented.”

    Jane Pearson, PhD
    Chief of Suicide Preventive Interventions Programs
    National Institute of Mental Health

    Hope lies in creating tools — and perhaps a combination of tools — to identify suicidal individuals and then provide those patients with appropriate treatment. Several researchers say they’re making progress now, whether finding signs in electronic health records (EHRs), blood samples, or functional MRIs (fMRIs) of patients’ brain patterns.

    “This is the future of mental health,” says John Torous, MD, a researcher and director of the digital psychiatry division at Beth Israel Deaconess Medical Center. “We’re working toward that future quickly and could see progress in two or three years.” 

    Brain scans reveal suicidal thoughts 

    Functional MRIs may unlock some of the secrets of suicidality, researchers from Carnegie Mellon University, the University of Pittsburgh School of Medicine, and elsewhere have found.

    The team scanned the brains of 34 people, half of whom recently reported suicidal thoughts, while they contemplated several emotion-laden words. The scans revealed that certain words — “death” and “cruelty,” for example — caused more types of brain activity in suicidal people than in control subjects.

    Next, the researchers fed the scans into a computer and, using machine-learning techniques, it developed an algorithm for recognizing which brain patterns reflect suicidality. The computer did quite well, according to a recent Nature Human Behaviour study. In fact, it was 91% accurate in identifying people who had considered suicide. Perhaps even more helpful, it was 94% accurate in separating people who had considered but didn’t attempt suicide from those who had actually attempted suicide.

    Moving forward, the team is exploring whether brain patterns can predict future suicidality, explains David Brent, MD, endowed chair in suicide studies at the University of Pittsburgh School of Medicine.

    The goal, explains Brent, is not to use costly fMRIs on all patients who are, for example, depressed. Instead, the team hopes to figure out which thoughts are most associated with suicidality and use that information to create other tools. Clinicians would then use those tools in their practices to identify patients who have those telling thoughts.

    “Moreover,” says Brent, “knowing which thought patterns are associated with suicidal risk could guide therapists in helping suicidal people learn to modify their thinking and reduce their risk of suicide.”

    Hints hidden in electronic health records

    Electronic health records, an accumulation of vast amounts of valuable data, can be a powerful resource for identifying suicidality. That’s what a group of researchers that includes investigators from Massachusetts General Hospital and Boston Children’s Hospital reported in a 2017 study published in the American Journal of Psychiatry.

    The researchers accessed information from more than 1 million patient records spanning 15 years. Their efforts revealed several factors that were more common in the EHRs of patients who had attempted or died by suicide — including nonpsychiatric conditions such as hepatitis C.

    They then created a machine-learning model that was able to accurately identify nearly half of suicide attempts based on these factors alone, without considering later data in the EHRs explicitly indicating suicidality.

    What’s more, as the model combed through the years of records, it was able to flag suicidality an average of three to four years before the EHR notes did so.

    "Machine learning can … integrate much more information than any human being can and detect risk signals. But it doesn’t substitute for clinical judgment or expertise.”

    Jordan Smoller, MD
    Harvard School of Medicine

    Now the researchers are working to validate their model in other populations. The ultimate goal: to develop a user-friendly element to add to EHRs that would alert physicians of a patient’s potential suicide risk. 
    "Machine learning can … integrate much more information than any human being can and detect risk signals,” says Jordan Smoller, MD, a study author and professor of psychiatry at the Harvard School of Medicine. “But it doesn’t substitute for clinical judgment or expertise. It’s meant to enhance it.”

    Looking at patients’ blood

    Perhaps an objective blood test could help save lives, a third team of suicide investigators theorized.

    To test this notion, researchers at Indiana University School of Medicine enrolled hundreds of male psychiatric patients and took blood samples at several appointments. Between visits, certain patients shifted from not feeling suicidal at all to feeling highly suicidal. That enabled researchers to identify which blood biomarkers in those individuals also changed between visits. In addition, they worked with the local coroner’s office to check the identified biomarkers against those from men who had died by suicide.

    To further enhance the effectiveness of their approach, the researchers used a series of questions to create a checklist that yields a suicide risk score. They also developed an app version of it for patients to complete.

    Finally, the researchers tested the combined blood test-app tool with a new group of patients they then monitored for a year. And the tool predicted suicidality impressively well. It identified which patients would go on to have serious suicidal thoughts with 92% accuracy, and it predicted with 71% accuracy which patients would be hospitalized for suicidal behaviors.

    Alexander B. Niculescu, MD, PhD, the psychiatrist who leads the team, says he next hopes to research the general prevalence of the biomarkers to better understand their predictive value. “Now that we’ve identified ‘signatures’ to look for in people with mental health conditions who may be at risk for suicide, we want to look at larger populations who are not in psychiatric care,” he says.

    Niculescu also wants to develop an app-blood test tool that any physician could use for suicide diagnosis and prevention. But he urges patience. “These studies take time,” he notes.

    As researchers like Niculescu make progress, parents like Cheryl Collier hope others will never face a loss like the one she suffered.

    “Had I known, I would have gotten more involved, whether he wanted me to or not,” she says. “If predictive tests had said he was at risk, I could have kept an eye out.”

    Meanwhile, Collier wants to share her son’s story in the hopes it could help save someone else. “Sean will still be a healer just like the doctor he wanted to be in life.”