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

Differentiating Between Sexual Offending and Violent Non-sexual Offending in Men With Schizophrenia Spectrum Disorders Using Machine Learning

Forensic psychiatric populations commonly contain a subset of persons with schizophrenia spectrum disorders (SSD) who have committed sex offenses. A comprehensive delineation of the features that distinguish persons with SSD who have committed sex offenses from persons with SSD who have committed violent non-sex offenses could be relevant to the development of differentiated risk assessment, risk management and treatment approaches. This analysis included the patient records of 296 men with SSD convicted of at least one sex and/or violent offense who were admitted to the Centre for Inpatient Forensic Therapy at the University Hospital of Psychiatry Zurich between 1982 and 2016. Using supervised machine learning, data on 461 variables retrospectively collected from the records were compared with respect to their relative importance in differentiating between men who had committed sex offenses and men who had committed violent non-sex offenses. The final machine learning model was able to differentiate between the two types of offenders with a balanced accuracy of 71.5% (95% CI = [60.7, 82.1]) and an AUC of .80 (95% CI = [.67, .93]). The main distinguishing features included sexual behaviours and interests, psychopathological symptoms and characteristics of the index offense. Results suggest that when assessing and treating persons with SSD who have committed sex offenses, it appears to be relevant to not only address the core symptoms of the disorder, but to also take into account general risk factors for sexual recidivism, such as atypical sexual interests and sexual preoccupation.

Sociodemographic Variables in Offender and Non-Offender Patients Diagnosed with Schizophrenia Spectrum Disorders-An Explorative Analysis Using Machine Learning

With the growing availability of medical data and the enhanced performance of computers, new opportunities for data analysis in research are emerging. One of these modern approaches is machine learning (ML), an advanced form of statistics broadly defined as the application of complex algorithms. ML provides innovative methods for detecting patterns in complex datasets. This enables the identification of correlations or the prediction of specific events. These capabilities are especially valuable for multifactorial phenomena, such as those found in mental health and forensic psychiatry. ML also allows for the quantification of the quality of the emerging statistical model. The present study aims to examine various sociodemographic variables in order to detect differences in a sample of 370 offender patients and 370 non-offender patients, all with schizophrenia spectrum disorders, through discriminative model building using ML. In total, 48 variables were tested. Out of seven algorithms, gradient boosting emerged as the most suitable for the dataset. The discriminative model finally included three variables (regarding country of birth, residence status, and educational status) and yielded an area under the curve (AUC) of 0.65, meaning that the statistical discrimination of offender and non-offender patients based purely on the sociodemographic variables is rather poor.

Enhancing psychiatric rehabilitation outcomes through a multimodal multitask learning model based on BERT and TabNet: An approach for personalized treatment and improved decision-making

Evaluating the rehabilitation status of individuals with serious mental illnesses (SMI) necessitates a comprehensive analysis of multimodal data, including unstructured text records and structured diagnostic data. However, progress in the effective assessment of rehabilitation status remains limited. Our study develops a deep learning model integrating Bidirectional Encoder Representations from Transformers (BERT) and TabNet through a late fusion strategy to enhance rehabilitation prediction, including referral risk, dangerous behaviors, self-awareness, and medication adherence, in patients with SMI. BERT processes unstructured textual data, such as doctor's notes, whereas TabNet manages structured diagnostic information. The model's interpretability function serves to assist healthcare professionals in understanding the model's predictive decisions, improving patient care. Our model exhibited excellent predictive performance for all four tasks, with an accuracy exceeding 0.78 and an area under the curve of 0.70. In addition, a series of tests proved the model's robustness, fairness, and interpretability. This study combines multimodal and multitask learning strategies into a model and applies it to rehabilitation assessment tasks, offering a promising new tool that can be seamlessly integrated with the clinical workflow to support the provision of optimized patient care.

Beyond Discrimination: Generative AI Applications and Ethical Challenges in Forensic Psychiatry

The advent and growing popularity of generative artificial intelligence (GenAI) holds the potential to revolutionise AI applications in forensic psychiatry and criminal justice, which traditionally relied on discriminative AI algorithms. Generative AI models mark a significant shift from the previously prevailing paradigm through their ability to generate seemingly new realistic data and analyse and integrate a vast amount of unstructured content from different data formats. This potential extends beyond reshaping conventional practices, like risk assessment, diagnostic support, and treatment and rehabilitation plans, to creating new opportunities in previously underexplored areas, such as training and education. This paper examines the transformative impact of generative artificial intelligence on AI applications in forensic psychiatry and criminal justice. First, it introduces generative AI and its prevalent models. Following this, it reviews the current applications of discriminative AI in forensic psychiatry. Subsequently, it presents a thorough exploration of the potential of generative AI to transform established practices and introduce novel applications through multimodal generative models, data generation and data augmentation. Finally, it provides a comprehensive overview of ethical and legal issues associated with deploying generative AI models, focusing on their impact on individuals as well as their broader societal implications. In conclusion, this paper aims to contribute to the ongoing discourse concerning the dynamic challenges of generative AI applications in forensic contexts, highlighting potential opportunities, risks, and challenges. It advocates for interdisciplinary collaboration and emphasises the necessity for thorough, responsible evaluations of generative AI models before widespread adoption into domains where decisions with substantial life-altering consequences are routinely made.

Students' 2018 PISA reading self-concept: Identifying predictors and examining model generalizability for emergent bilinguals

Decades of research have indicated that reading self-concept is an important predictor of reading achievement. During this period, the population of emergent bilinguals has continued to increase within United States' schools. However, the existing literature has tended to examine native English speakers' and emergent bilinguals' reading self-concept in the aggregate, thereby potentially obfuscating the unique pathways through which reading self-concept predicts reading achievement. Furthermore, due to the overreliance of native English speakers in samples relating to theory development, researchers attempting to examine predictors of reading achievement may a priori select variables that are more aligned with native English speakers' experiences. To address this issue, we adopted Elastic Net, which is a theoretically agnostic methodology and machine learning approach to variable selection to identify the proximal and distal predictors of reading self-concept for the entire population; in our study, participants from the United States who participated in PISA 2018 served as the baseline group to determine significant predictors of reading self-concept with the intent of identifying potential new directions for future researchers. Based on Elastic Net analysis, 20 variables at the student level, three variables at the teacher level, and 12 variables at the school level were identified as the most salient predictors of reading self-concept. We then utilized a multilevel modeling approach to test model generalizability of the identified predictors of reading self-concept for emergent bilinguals and native English speakers. We disaggregated and compared findings for both emergent bilinguals and native English speakers. Our results indicate that although some predictors were important for both groups (e.g., perceived information and communications technologies competence), other predictors were not (e.g., competitiveness). Suggestions for future directions and implications of the present study are examined.

Testamentary capacity assessment in dementia using artificial intelligence: prospects and challenges

Testamentary capacity (TC), a set of capacities involved in making a valid Will, has become prominent in capacity evaluations due to the demographic increase in older persons and associated increase in cognitive impairment. The assessment of contemporaneous TC follows the criteria derived from the Banks v Goodfellow case, which do not bind capacity solely on the basis of presence of a cognitive disorder. Although effort is being made for establishing more objective criteria for TC judgment, variations in situational complexity call for incorporating the different circumstances of the testator in capacity assessment. Artificial intelligence (AI) technologies such as statistical machine learning have been used in forensic psychiatry mainly for the prediction of aggressive behavior and recidivism but little has been done in the area of capacity assessment. However, the statistical machine learning model responses are difficult to interpret and explain, which presents problems with regard to the new General Data Protection Regulation (GDPR) of the European Union. In this Perspective we present a framework for an AI decision support tool for TC assessment. The framework is based on AI decision support and explainable AI (XAI) technology.

Artificial intelligence ecosystem for computational psychiatry: Ideas to practice

Computational psychiatry is an emerging field that not only explores the biological basis of mental illness but also considers the diagnoses and identifies the underlying mechanisms. One of the key strengths of computational psychiatry is that it may identify patterns in large datasets that are not easily identifiable. This may help researchers develop more effective treatments and interventions for mental health problems. This paper is a narrative review that reviews the literature and produces an artificial intelligence ecosystem for computational psychiatry. The artificial intelligence ecosystem for computational psychiatry includes data acquisition, preparation, modeling, application, and evaluation. This approach allows researchers to integrate data from a variety of sources, such as brain imaging, genetics, and behavioral experiments, to obtain a more complete understanding of mental health conditions. Through the process of data preprocessing, training, and testing, the data that are required for model building can be prepared. By using machine learning, neural networks, artificial intelligence, and other methods, researchers have been able to develop diagnostic tools that can accurately identify mental health conditions based on a patient’s symptoms and other factors. Despite the continuous development and breakthrough of computational psychiatry, it has not yet influenced routine clinical practice and still faces many challenges, such as data availability and quality, biological risks, equity, and data protection. As we move progress in this field, it is vital to ensure that computational psychiatry remains accessible and inclusive so that all researchers may contribute to this significant and exciting field.

The HARM models: Predicting longitudinal physical aggression in patients with schizophrenia at an individual level

The prediction and prevention of aggression in individuals with schizophrenia remains a top priority within forensic psychiatric settings. While risk assessment methods are well rooted in forensic psychiatry, there are no available tools to predict longitudinal physical aggression in patients with schizophrenia within forensic settings at an individual level. In the present study, we used evidence-based risk and protective factors, as well as variables related to course of treatment assessed at baseline, to predict prospective incidents of physical aggression (4-month, 12-month, and 18-month follow-up) among 151 patients with schizophrenia within the forensic mental healthcare system. Across our HARM models, the balanced accuracy (sensitivity + specificity/2) of predicting physical aggressive incidents in patients with schizophrenia ranged from 59.73 to 87.33% at 4-month follow-up, 68.31-80.10% at 12-month follow-up, and 46.22-81.63% at 18-month follow-up, respectively. Additionally, we developed separate models, using clinician rated clinical judgement of short term and immediate violent risk, as a measure of comparison. Several modifiable evidence-based predictors of prospective physical aggression in schizophrenia were identified, including impulse control, substance abuse, impulsivity, treatment non-adherence, mood and psychotic symptoms, substance abuse, and poor family support. To the best of our knowledge, our HARM models are the first to predict longitudinal physical aggression at an individual level in patients with schizophrenia in forensic settings. However, it is important to caution that since these machine learning models were developed in the context of forensic settings, they may not be generalisable to individuals with schizophrenia more broadly. Moreover, a low base rate of physical aggression was observed in the testing set (6.0-11.6% across timepoints). As such, larger cohorts will be required to determine the replicability of these findings.

Superstorm Sandy exposure in utero is associated with neurobehavioral phenotypes and brain structure alterations in childhood: A machine learning approach

Introduction

Prenatal maternal stress (PNMS), including exposure to natural disasters, has been shown to serve as a risk factor for future child psychopathology and suboptimal brain development, particularly among brain regions shown to be sensitive to stress and trauma exposure. However, statistical approaches deployed in most studies are usually constrained by a limited number of variables for the sake of statistical power. Explainable machine learning, on the other hand, enables the study of high data dimension and offers novel insights into the prominent subset of behavioral phenotypes and brain regions most susceptible to PNMS. In the present study, we aimed to identify the most important child neurobehavioral and brain features associated with in utero exposure to Superstorm Sandy (SS).

Methods

By leveraging an explainable machine learning technique, the Shapley additive explanations method, we tested the marginal feature effect on SS exposures and examined the individual variable effects on disaster exposure.

Results

Results show that certain brain regions are especially sensitive to in utero exposure to SS. Specifically, in utero SS exposure was associated with larger gray matter volume (GMV) in the right caudate, right hippocampus, and left amygdala and smaller GMV in the right parahippocampal gyrus. Additionally, higher aggression scores at age 5 distinctly correlated with SS exposure.

Discussion

These findings suggest in utero SS exposure may be associated with greater aggression and suboptimal developmental alterations among various limbic and basal ganglia brain regions.

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

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.

Model Building in Forensic Psychiatry: A Machine Learning Approach to Screening Offender Patients with SSD

Today’s extensive availability of medical data enables the development of predictive models, but this requires suitable statistical methods, such as machine learning (ML). Especially in forensic psychiatry, a complex and cost-intensive field with risk assessments and predictions of treatment outcomes as central tasks, there is a need for such predictive tools, for example, to anticipate complex treatment courses and to be able to offer appropriate therapy on an individualized basis. This study aimed to develop a first basic model for the anticipation of adverse treatment courses based on prior compulsory admission and/or conviction as simple and easily objectifiable parameters in offender patients with a schizophrenia spectrum disorder (SSD). With a balanced accuracy of 67% and an AUC of 0.72, gradient boosting proved to be the optimal ML algorithm. Antisocial behavior, physical violence against staff, rule breaking, hyperactivity, delusions of grandeur, fewer feelings of guilt, the need for compulsory isolation, cannabis abuse/dependence, a higher dose of antipsychotics (measured by the olanzapine half-life) and an unfavorable legal prognosis emerged as the ten most influential variables out of a dataset with 209 parameters. Our findings could demonstrate an example of the use of ML in the development of an easy-to-use predictive model based on few objectifiable factors.

Predicting crime during or after psychiatric care: Evaluating machine learning for risk assessment using the Danish patient registries

BACKGROUND

Structural changes in psychiatric systems have altered treatment opportunities for patients in need of mental healthcare. These changes are possibly associated with an increase in post-discharge crime, reported in the increase of forensic psychiatric populations. As current risk-assessment tools are time-consuming to administer and offer limited accuracy, this study aims to develop a predictive model designed to identify psychiatric patients at risk of committing crime leading to a future forensic psychiatric treatment course.

METHOD

We utilized the longitudinal quality of the Danish patient registries, identifying the 45.720 adult patients who had contact with the psychiatric system in 2014, of which 474 committed crime leading to a forensic psychiatric treatment course after discharge. Four machine learning models (Logistic Regression, Random Forest, XGBoost and LightGBM) were applied over a range of sociodemographic, judicial, and psychiatric variables.

RESULTS

This study achieves a F1-macro score of 76%, with precision = 57% and recall = 47% reported by the LightGBM algorithm. Our model was therefore able to identify 47% of future forensic psychiatric patients, while making correct predictions in 57% of cases.

CONCLUSION

The study demonstrates how a clinically useful initial risk-assessment can be achieved using machine learning on data from patient registries. The proposed approach offers the opportunity to flag potential future forensic psychiatric patients, while in contact with the general psychiatric system, hereby allowing early-intervention initiatives to be activated.

A New Strategy for Identification of Coal Miners With Abnormal Physical Signs Based on EN-mRMR

Coal miners’ occupational health is a key part of production safety in the coal mine. Accurate identification of abnormal physical signs is the key to preventing occupational diseases and improving miners’ working environment. There are many problems when evaluating the physical health status of miners manually, such as too many sign parameters, low diagnostic efficiency, missed diagnosis, and misdiagnosis. To solve these problems, the machine learning algorithm is used to identify miners with abnormal signs. We proposed a feature screening strategy of integrating elastic net (EN) and Max-Relevance and Min-Redundancy (mRMR) to establish the model to identify abnormal signs and obtain the key physical signs. First, the raw 21 physical signs were expanded to 25 by feature construction technology. Then, the EN was used to delete redundant physical signs. Finally, the mRMR combined with the support vector classification of intelligent optimization algorithm by Gravitational Search Algorithm (GSA-SVC) is applied to further simplify the rest of 12 relatively important physical signs and obtain the optimal model with data of six physical signs. At this time, the accuracy, precision, recall, specificity, G-mean, and MCC of the test set were 97.50%, 97.78%, 97.78%, 97.14%, 0.98, and 0.95. The experimental results show that the proposed strategy improves the model performance with the smallest features and realizes the accurate identification of abnormal coal miners. The conclusion could provide reference evidence for intelligent classification and assessment of occupational health in the early stage.

The impact of machine learning in predicting risk of violence: A systematic review

BACKGROUND

Inpatient violence in clinical and forensic settings is still an ongoing challenge to organizations and practitioners. Existing risk assessment instruments show only moderate benefits in clinical practice, are time consuming, and seem to scarcely generalize across different populations. In the last years, machine learning (ML) models have been applied in the study of risk factors for aggressive episodes. The objective of this systematic review is to investigate the potential of ML for identifying risk of violence in clinical and forensic populations.

METHODS

Following Preferred Reporting Items for Systematic Review and Meta-Analyses (PRISMA) guidelines, a systematic review on the use of ML techniques in predicting risk of violence of psychiatric patients in clinical and forensic settings was performed. A systematic search was conducted on Medline/Pubmed, CINAHL, PsycINFO, Web of Science, and Scopus. Risk of bias and applicability assessment was performed using Prediction model Risk Of Bias ASsessment Tool (PROBAST).

RESULTS

We identified 182 potentially eligible studies from 2,259 records, and 8 papers were included in this systematic review. A wide variability in the experimental settings and characteristics of the enrolled samples emerged across studies, which probably represented the major cause for the absence of shared common predictors of violence found by the models learned. Nonetheless, a general trend toward a better performance of ML methods compared to structured violence risk assessment instruments in predicting risk of violent episodes emerged, with three out of eight studies with an AUC above 0.80. However, because of the varied experimental protocols, and heterogeneity in study populations, caution is needed when trying to quantitatively compare (e.g., in terms of AUC) and derive general conclusions from these approaches. Another limitation is represented by the overall quality of the included studies that suffer from objective limitations, difficult to overcome, such as the common use of retrospective data.

CONCLUSION

Despite these limitations, ML models represent a promising approach in shedding light on predictive factors of violent episodes in clinical and forensic settings. Further research and more investments are required, preferably in large and prospective groups, to boost the application of ML models in clinical practice.

SYSTEMATIC REVIEW REGISTRATION

[www.crd.york.ac.uk/prospero/], identifier [CRD42022310410].