Predicting criminal and violent outcomes in psychiatry: a meta-analysis of diagnostic accuracy

Abstract

Although reducing criminal outcomes in individuals with mental illness have long been a priority for governments worldwide, there is still a lack of objective and highly accurate tools that can predict these events at an individual level. Predictive machine learning models may provide a unique opportunity to identify those at the highest risk of criminal activity and facilitate personalized rehabilitation strategies. Therefore, this systematic review and meta-analysis aims to describe the diagnostic accuracy of studies using machine learning techniques to predict criminal and violent outcomes in psychiatry. We performed meta-analyses using the mada, meta, and dmetatools packages in R to predict criminal and violent outcomes in psychiatric patients (n = 2428) (Registration Number: CRD42019127169) by searching PubMed, Scopus, and Web of Science for articles published in any language up to April 2022. Twenty studies were included in the systematic review. Overall, studies used single-nucleotide polymorphisms, text analysis, psychometric scales, hospital records, and resting-state regional cerebral blood flow to build predictive models. Of the studies described in the systematic review, nine were included in the present meta-analysis. The area under the curve (AUC) for predicting violent and criminal outcomes in psychiatry was 0.816 (95% Confidence Interval (CI): 70.57-88.15), with a partial AUC of 0.773, and average sensitivity of 73.33% (95% CI: 64.09-79.63), and average specificity of 72.90% (95% CI: 63.98-79.66), respectively. Furthermore, the pooled accuracy across models was 71.45% (95% CI: 60.88-83.86), with a tau squared (τ²) of 0.0424 (95% CI: 0.0184-0.1553). Based on available evidence, we suggest that prospective models include evidence-based risk factors identified in prior actuarial models. Moreover, there is a need for a greater emphasis on identifying biological features and incorporating novel variables which have not been explored in prior literature. Furthermore, available models remain preliminary, and prospective validation with independent datasets, and across cultures, will be required prior to clinical implementation. Nonetheless, predictive machine learning models hold promise in providing clinicians and researchers with actionable tools to improve how we prevent, detect, or intervene in relevant crime and violent-related outcomes in psychiatry.

Fig. 1. Paired Forest plot of model accuracy for criminal and violent outcomes in psychiatry.

A linear mixed model with random effects was selected to produce summary estimates of sensitivity and specificity using the retisma function in mada. The average sensitivity across studies was 73.33% (95% I: 64.09–79.63) and average specificity was 72.90% (95% CI: 60.50–96.6). As such, the balanced accuracy across models (sensitivity + specificity/2) is 73.11%.

Fig. 2. Pooled effects of model accuracy.

Pooled accuracy of criminal and violent models in psychiatry across 2428 patients (two studies used the same sample n = 370) within a random effects model using a restricted maximum likelihood estimator to calculate the heterogeneity variance τ2. Reported mean accuracy across models was used, in conjunction with standard deviation, calculated by multiplying the standard error by the square root of the sample size (SD = SE×√n). Knapp-Hartung adjustments were used to calculate the confidence interval around the pooled effect. The average accuracy across models was 71.45% (95% CI: 60.88–83.86), with a heterogeneity variance τ² of 0.0424.