The University Network

AI Predicts Risk Of School Violence

In a recent pilot study, researchers from the Cincinnati Children’s Hospital Medical Center (CCHMC) have demonstrated artificial intelligence as a useful tool in predicting which students are more likely to perpetrate school violence.

The researchers determined that machine learning is as accurate as a team of child, adolescent and forensic psychiatrists in determining a young person’s risk of committing violence at school.

The full study is published online in the journal Psychiatric Quarterly.

This research succeeds a history of the hospital evaluating violence in children.

“We were successful in predicting violence by children and adolescents within the hospital,” said Drew Barzman, a child forensic psychiatrist at CCHMC and lead author of the study. “Therefore, we decided to apply our successful methods to the school setting.”

Because of the rise in school violence over the past 10 years, the researchers wanted to use a more sensitive method to evaluate students and determine those who are at high risk for school violence.

In the study, Barzman and his team picked 103 students between ages 12 and 18 from 74 schools throughout the U.S. All of the chosen students had a history of minor or major behavioral change or aggression towards themselves or others.

All of them were recruited from psychiatry emergency departments, outpatient clinics, and inpatient clinics.

The team performed traditional school risk assessments with all of the participating students. They wrote down notes and transcribed audio recordings from the interviews.

Using two scales that the researchers validated in previous research, the team determined the students to be relatively equally divided. Out of the 103 participating students, 55 were grouped as moderate to high risk, and 48 were found to be low risk, based on paper risk assessments.

The researchers used the recorded interview content to develop a machine-learning algorithm capable of predicting risk of school violence.

The algorithm showed an accuracy rate of 91.02 percent, which is considered excellent. When demographic and socioeconomic data was added, the accuracy rate increased to 91.45 percent.

“The machine learning algorithm, based only on the participant’s interview, was almost as accurate in assessing risk levels as a full assessment by our research team, including gathering information from parents and the school, a review of records when available, and scoring on the two scales we developed,” Yizhao Ni, a computational scientist in the Division of Biomedical Informatics at CCHMC and co-author of the study, said in a statement.

Through this study, the team “will be able to build artificial intelligence to augment human clinical judgment,” said Barzman.

The researchers’ assessments were strictly based on predicting any type of physical aggression at school. They did not gather data to prove whether machine learning could help prevent school violence.

That is their next goal. But for that, they need more funding.

“Funding is our biggest hurdle and challenge,” said Barzman. “In the future, we will be able to complete a study which compares the effectiveness of an expert forensic team versus the artificial intelligence in violence prevention by tracking outcomes in a large prospective randomized study. Outcomes on violence and aggression can be easily gathered via schools, web-based surveys, and phone interviews.”

Ultimately, the research team’s goal is “to spread the use of the machine learning technology to schools in the future to augment structures, professional judgment to more efficiently and effectively prevent school violence,” Barzman said in a statement.