April 16, 2021
2 min read
Watts reports no relevant financial disclosures. Please see the study for all other authors’ relevant financial disclosures.
Machine learning models may have greater accuracy than gold-standard risk assessment tools for predicting criminal offense among people with psychiatric disorders, according to study results published in Journal of Psychiatric Research.
“Knowing the type of crime an individual is likely to commit, before the offense occurs, is urgently needed in order to guide more targeted and precise risk assessment strategies and frontline therapeutic interventions,” Devon Watts, of the department of psychiatry and behavioral neurosciences at McMaster University in Canada, and colleagues wrote. “Furthermore, the vast majority of work thus far has focused on predicting recidivism in non-psychiatric prison populations. Importantly, it is largely unclear whether such models can be appropriately extrapolated to offenses committed by those with severe mental illness.”
Current, actuarial risk estimates are unable to individually predict criminal offense type a patient will go on to commit, and they frequently simply evaluate the general risk for crime occurring among a group sample, according to the researchers. In the current study, Watts and colleagues sought to create a machine learning model able to predict criminal offense type committed among a large transdiagnostic sample of psychiatry patients, on the individual level. They applied three machine learning algorithms to a representative and diverse sample of 1,240 individuals in the forensic mental health system and evaluated clinical, historical and sociodemographic variables, which they considered as possible predictors, via a data-driven manner. Further, they created separate models for each criminal offense type and utilized feature selection methods to bolster the interpretability and generalizability of the results.
When using 36 variables and a sensitivity of 82.44% and specificity of 60%, the researchers were able to predict sexual offenses from nonviolent and violent offenses on the individual level. Results of a binary classification model showed sensitivity of 83.26% and specificity of 77.42% allowed for prediction of sexual and violent offenses at an individual level using 20 clinical variables. Similarly, sensitivity of 74.6% and specificity of 80.65% using 30 clinical variables allowed for individual prediction of non-violent and sexual offenses.
“Within subsequent work it is important to note that model accuracy should be considered alongside several other factors, such as the input features used, the preprocessing pipeline, feature selection method, model optimization strategy, and the validation procedure,” Watts and colleagues wrote. “The importance of these factors for model replicability and utility cannot be overstated. Furthermore, data-driven approaches to feature selection can be useful in many cases, since it does not require knowledge derived from pre-existing literature to manually select important variables. However, hypothesis-driven designs based on scientific evidence are also important for the field.”
Credit: Google News