Predicting offenses among individuals with psychiatric disorders - A machine learning approach

Area of Science:

• Forensic Psychiatry
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Traditional actuarial risk estimates lack individual prediction of criminal offense types.
• Existing methods often assess general crime likelihood in group samples.
• Improved statistical strategies are needed to enhance risk assessment predictive utility.

Purpose of the Study:

• Develop a machine learning model to predict criminal offense types in psychiatric patients.
• Enable individual-level prediction of offenses within a transdiagnostic sample.
• Advance the accuracy and clinical utility of forensic risk assessments.

Main Methods:

• Applied Random Forest, Elastic Net, and SVM algorithms to 1240 forensic mental health patients.
• Utilized clinical, historical, and sociodemographic variables for prediction.
• Developed separate models for each offense type with feature selection for interpretability.

Main Results:

• Sexual offenses predicted with 82.44% sensitivity and 60.00% specificity.
• Sexual and violent offenses predicted with 83.26% sensitivity and 77.42% specificity.
• Non-violent and sexual offenses predicted with 74.60% sensitivity and 80.65% specificity.

Conclusions:

• Machine learning models demonstrate superior accuracy (AUCs 0.70-0.80) compared to gold-standard tools.
• Individual-level offense prediction offers enhanced clinical utility.
• Future refinements are expected to further improve prospective model accuracy.