AI Identifies Early Life Factors Predicting Suicide Attempt in Young Adults

Massimiliano Orri, Ph.D.

Early life factors and machine learning can inform on the risk of suicide two decades later.

The findings may be useful for healthcare professionals in contact with pregnant mothers or pediatricians to improve the early identification of children particularly at risk of developing future suicide risk, corresponding author Massimiliano Orri, Ph.D., said in a statement to Chief Healthcare Executive™.

Orri and an investigative team identified and evaluated the ability of early life factors to predict suicide attempt in adolescents and young adults from the general population. Study participants came from the Québec Longitudinal Study of Child Development, which initially included 2,120 singletons born in Québec, Canada in 1997 or 1998. The children were selected from the Québec Birth Registry using a stratified random procedure and were regularly assessed from ages five months to 20 years old. Overall, the study included 1,623 participants with at least one assessment of suicide attempt between ages 13 and 20 years old.

If at ages 13-, 15-, 17-, or 20-years old adolescents answered positively to the question, “In the past 12 months, did you ever seriously think of attempting suicide?” they were then asked, “In the past 12 months, how many times did you attempt suicide?” The team assessed lifetime suicide attempt at age 20 years old by asking, “In your lifetime, have you ever been to the emergency room because you tried to kill yourself,” and “In your lifetime, have you ever been hospitalized after trying to kill yourself.”

Orri and colleagues used a broad range of potential factors reported by parents when the child was five months old, along with factors extracted from hospital birth records. They assessed 150 variables encompassing sociodemographic factors and child, family, parental, and neighborhood characteristics.

A random forest machine-learning algorithm was used, as the team noted random forests are well adapted to mental health prediction. For the analysis, they randomly split the original data set into training (80% of the total cohort) and testing (20% of the total cohort) samples. Training samples were used to compute the predictive algorithms for the outcome. For preliminary analyses, sex was a factor in the models, but because the variable overshadowed all other variables, the team conducted separate analyses for males and females.

Among the 1,623 included participants aged 20 years old, 52.1% were female. Machine-learning algorithms showed moderate prediction performance. Areas under the curve for the prediction of suicide attempt were .72 (95% CI, .71-.73) for females and .62 (95% CI, .6-.62) for males. Models had low sensitivity (females, .5; males, .32), moderate positive predictive values (females, .6; males, .62), and good specificity (females, .76; males, .82) and negative predictive values (females, .75; males, .71).

Socioeconomic and demographic characteristics of the family, parents’ psychological state, and parenting practices were the most important factors contributing to the prediction. Birth-related variables such as prematurity also contributed to the prediction of suicidal behavior.

For females, family-related socioeconomic and demographic characteristics were the top factor, while parents’ antisocial behavior was the top factor for males.

Still, the study authors said the factors’ utility in the long-term prediction of suicide attempt was limited.

The study, “Machine Learning Assessment of Early Life Factors Predicting Suicide Attempt in Adolescence or Young Adulthood,” was published online in JAMA Network Open.