Background: Actuarial risk estimates are considered the gold-standard way to assess whether forensic psychiatry patients are likely to commit prospective criminal offences. However, these risk estimates cannot individually predict the type of criminal offence a patient will subsequently commit, and often simply assess the general likelihood of crime occurring in a group sample. In order to advance the predictive utility of risk assessments, better statistical strategies are required. Aim: To develop a machine learning model to predict the type of criminal offense committed in forensic psychiatry patients, at an individual level. Method: Machine learning algorithms (Random Forest, Elastic Net, SVM), were applied to a representative and diverse sample of 1240 patients in the forensic mental health system. Clinical, historical, and sociodemographic variables were considered as potential predictors and assessed in a data-driven way. Separate models were created for each type of criminal offence, and feature selection methods were used to improve the interpretability and generalizability of our findings. Results: Sexual and violent crimes can be predicted at an individual level with 83.26% sensitivity and 77.42% specificity using only 20 clinical variables. Likewise, nonviolent, and sexual crimes can be individually predicted with 74.60% sensitivity and 80.65% specificity using 30 clinical variables. Conclusion: The current results suggest that machine learning models have accuracy comparable to existing risk assessment tools (AUCs .70-.80). However, unlike existing risk tools, this approach allows for the prediction of cases at an individual level, which is more clinically useful. The accuracy of prospective models is expected to only improve with further refinement. / Thesis / Master of Science (MSc) / Individuals end up in the forensic mental health system when they commit crimes and are found to be not criminality responsible because of a mental disorder. They are released back into the community when deemed to be low risk. However, it is important to consider the accuracy of the method we use to determine risk at the level of an individual person. Currently, we use group average to assess individual risk, which does not work very well. The range of our predictions become so large, that they are virtually meaningless. In other words, the average of a group is meaningless with respect to you. Instead, statistical models can be developed that can make predictions accurately, and at an individual level. Therefore, the current work sought to predict the types of criminal offences committed, among 1240 forensic patients. Making accurate predictions of the crimes people may commit in the future is urgently needed to identify better strategies to prevent these crimes from occurring in the first place. Here, we show that it is possible to predict the type of criminal offense an individual will later commit, using data that is readily available by clinicians. These models perform similarly to the best risk assessment tools available, but unlike these risk assessment tools, can make predictions at an individual level. It is suggested that similar approaches to the ones outlined in this paper could be used to improve risk prediction models, and aid crime prevention strategies.

In this thesis, we developed prognostic models of clinical outcomes, specific to violent and criminal outcomes in psychiatry, and predictive models of treatment response at an individual level. Overall, we demonstrate that evidence-based risk factors, protective factors, and treatment status variables were able to prognosticate prospective physical aggression at an individual level; 2) prognostic models of clinical and violent outcomes in psychiatry have largely focused on clinical and sociodemographic variables, show similar performance between identifying true positives and true negatives, although the error rate of models are still high, and further refinement is needed; 3) within treatment response prediction models in MDD using EEG, greater performance was observed in predicting response to rTMS, relative to antidepressants, and across models, greater sensitivity (true positives), were observed relative to specificity (true negatives), suggesting that EEG prediction models thus far better identify non-responders than responders; and 4) across randomized clinical trials using data-driven biomarkers in predictive models, based on the consistency of performance across models with large sample sizes, the highest degree of evidence was in predicting response to sertraline and citalopram using fMRI features. / Dissertation / Doctor of Philosophy (PhD)