Researchers in Canada have investigated how artificial intelligence could help forensic psychiatry researchers examine the severity of previously documented violent behaviour using judicial records.
The study, published on 20 November 2026 in the journal Forensic Sciences, was conducted by Sabrina Grenier and Alexandre Hudon, affiliated with the Université de Montréal.
The researchers examined 1,386 anonymized judicial records from Quebec's mental disorder review board.
Forensic psychiatry combines psychiatric assessment with legal requirements. Professionals working in this field must consider clinical history, behavioural changes, treatment requirements and public safety when conducting assessments.
The research explored whether machine learning could identify meaningful patterns within existing clinical and judicial information.
The researchers employed Extreme Gradient Boosting (XGBoost), a machine learning algorithm that combines multiple decision trees to perform classification tasks.
Historical offences were divided into two severity categories using an adapted MacArthur violence severity framework.
Of the examined records, 825 belonged to the elevated-severity category, while 561 belonged to the lower-severity category.
The model achieved an overall classification accuracy of 70.2%. For the elevated-severity category, it achieved 75.2% precision and 74.4% recall.
Researchers also applied an explainable artificial intelligence technique known as SHAP to examine which variables contributed most strongly to the classifications.
Recent aggressive behaviour, police involvement and violations of established rules emerged as influential factors.
An important distinction is that the research examined historical information rather than predicting whether an individual would commit a violent act in the future.
The findings therefore cannot establish an individual's future dangerousness or justify independent automated legal decisions.
The study also faced limitations associated with its data source. Because the judicial records were anonymized, researchers could not determine whether multiple records referred to the same individual. Consequently, the dataset cannot necessarily be treated as representing 1,386 separate people.
Further independent validation and prospective evaluation are necessary before considering practical applications in forensic psychiatric assessments.
The research highlights the potential of computational methods to examine complex forensic psychiatric information while reinforcing the importance of professional judgment, responsible interpretation and scientific validation.