A team of researchers from Brazil has developed a new, “highly accurate” algorithm that is able to predict suicide attempts in the general population — including identifying psychological disorders or life crises that may put people at greater risk of attempting suicide.
In a paper published Jan. 14 in the Psychological Medicine journal, the team detailed how it created models to accurately predict suicide attempts among U.S. adults using machine-learning techniques and sociodemographic and clinical data. It is the first study to evaluate the prediction of suicide attempts in a nationally representative sample of the U.S. population.
Ives Passos, a professor at the Federal University of Rio Grande do Sul in Brazil and co-author of the paper, told The Academic Times that there is a growing body of scientific knowledge that has put forward several risk factors associated with patients who attempt suicide.
“These efforts have largely reported average group-level differences between people that attempted suicide and the ones that did not,” Passos said. “However, what was not known until recently is how to integrate these variables to build models to estimate the probability of an individual attempting suicide.”
This is important to investigate because suicide is a highly preventable event, Passos emphasized. Correctly identifying those at risk can allow for earlier intervention with prevention methods such as cognitive behavior therapy and lithium medication, which have been proven to reduce suicide attempts.
Passos and the research team used machine learning approaches to predict suicide attempts among survey participants in a nationally representative sample of U.S. adults, and among participants with lifetime major depressive episodes.
“Machine learning comes from the field of artificial intelligence and uses mathematical functions to give computer systems the ability to ‘learn’ from past experiences and make predictions on data, without being explicitly programmed,” Passos said. “These techniques can find hidden non-linear patterns of association between the predictor and the outcome and ultimately build a risk calculator to predict a clinical outcome of interest, such as a suicide attempt, at an individual level.”
In the nationally representative sample, the most important predictor variables were a diagnosis of borderline personality disorder or post-traumatic stress disorder. In the sample of participants who had major depressive episodes, the most important predictor variables were previous suicide attempts or diagnosis of borderline personality disorder. Additional key risk factors included other psychiatric diagnoses, death of a loved one, job loss, divorce and being the victim of a crime.
The study analyzed gender, age, race, marital status, education, income, prior suicide attempts, stressful life events and body mass index in data from the National Epidemiologic Survey on Alcohol and Related Conditions. This survey included about 50,000 U.S. adults and was conducted by the National Institute on Alcohol Abuse and Alcoholism.
The data were collected at two points — the first wave took place between 2001-2002, and the second wave consisted of follow-up interviews with the same participants between 2004-2005. Notably, it was the largest comorbidity study ever conducted in the U.S.
For the current paper, Passos and the research team assessed answers from nearly 35,000 participants, looking at whether the individuals had attempted suicide in the three-year follow-up period, and what factors could have predicted this outcome.
According to the authors, the models developed by the team achieved good performance results, and all tested algorithms achieved greater than 50% accuracy in distinguishing people who attempted suicide from those who did not.
Notably, the sample of 6,350 adults who had experienced lifetime major depressive episodes included 150 who had attempted suicide within the past three years. Of those 150 people, 35.6% had a drug use disorder, 46.2% had bipolar disorder, 43.2% had post-traumatic stress disorder and about 66% had borderline personality disorder.
The authors said the results may indicate that trauma is a significant predictor of a suicide attempt, but only for those who develop a trauma-related disorder.
Suicide was the 10th-leading cause of death overall in the U.S. in 2018, according to the Centers for Disease Control and Prevention. About 48,000 Americans died by suicide that year, and close to 800,000 people around the world commit suicide annually.
The team proposed that governments should use initiatives such as the National Epidemiologic Survey on Alcohol and Related Conditions to collect large datasets and make predictions of important clinical outcomes related to public health, such as suicide attempts.
“Machine learning models can be displayed as a user-friendly calculator and incorporated into the clinical workflows including electronic medical records,” Passos said. “In the case that the calculator predicts that a patient is likely to attempt suicide, the mental health professional can consider preventive strategies.”
The authors recommended that the model for the general population sample should be used as a screening tool to identify people at higher risk of attempting suicide. Health authorities could then potentially contact those at risk, or their families, to advise future mental health assessments.
For those who have already had a major depressive episode and are at risk for suicide attempts in the future, preventive strategies like the use of lithium or cognitive behavior therapy should be implemented.
Passos and his co-authors recently formed a research group, known as the Artificial Intelligence in Neuroscience group, that studies the interface between mental health and artificial intelligence. Over the past five years, his group and others have started to build machine-learning models to predict suicide attempts.
The lead author of the current paper, Cristiane dos Santos Machado, is a master’s degree in science student under the supervision of Passos. The group is working on additional projects to predict suicide during the COVID-19 pandemic and among people with PTSD.
For future studies, Passos suggested integrating data from different biological levels, such as genetics, metabolomics and digital health data, to potentially build more accurate models. Longer follow-up periods would also be beneficial to reduce the number of false positives, as the current study only followed patients for three years.
“We believe that significant advances in science need collaboration between experts in different fields of knowledge. Our vision is that health sciences will benefit from the field of artificial intelligence by using machine learning models to predict important clinical outcomes such as mortality and treatment response,” Passos said.
The study, “Prediction of suicide attempts in a prospective cohort study with a nationally representative sample of the US population” was published in the Psychological Medicine journal on Jan. 14. Cristiane dos Santos Machado of the Federal University of Rio Grande do Sul was the lead author. Ives Passos, Flávio Kapczinski and Marco Antonio Caldieraro, of the Federal University of Rio Grande do Sul, served as co-authors. Pedro L. Ballester, of McMaster University, Bo Cao, of the University of Alberta, and Benson Mwangi, of the University of Texas Health Science Center, also served as co-authors.