TABLE 2.

Features and predictors of the ML models.

References	Features	Outcome	Best performing algorithm	Validation	AUC	Accuracy (%)	Sensitivity (%)	Specificity (%)	Predictors
Gou et al. (18)	Education BPRS-4 activation BPRS-5 hostility PCL-SV total score HCR-20 total score BIS-11 total score Gray matter volume Regional homogeneity Fractional anisotropy	Violent = the Modified Overt Aggression Scale (MOAS) a score of ≥3 for item 4 (physical aggression scale) of the MOAS; Non-violent = a score of <2 for item 4 of the MOAS and were free of any severe aggressive act against property and/or themselves.	LASSO regression + support vector machine	LOOCV	0.95	90.67	90.91	90.48	Education BPRS-5: hostility PCL-SV total score HCR-20 total score Gray matter volume Regional homogeneity Fractional anisotropy
Kirchebner et al. (29)	Stressors in childhood/youth Bullying Separation/divorce of caregivers Impairment of the parent–child relationship Physical abuse by the caregiver Sexual abuse by the caregiver Poverty Separation from caregiver Rejection/being ignored by the caregiver active devaluation by the caregiver Poor parenting methods Violent physical illness of the patient Failure in school Stressors in adulthood Unemployment (at time of offense) homelessness Conflicts in the workplace Social isolation Violent victimization Psychiatric stressors Coercive psychiatric treatment At least three previous hospitalizations Compulsory psychiatric placement Positive symptoms during criminal offense	*Violent offense = homicide and attempted homicide, assault, rape, robbery, arson, and child abuse; Non-violent offense = threat, theft, damage to property, minor sexual offenses (e.g., exhibitionism), drug offenses, illegal gun possession, and other minor offenses (e.g., triggering false alarms or emergency brakes)	Boosted classification trees	5-CV	0.83	77	80.49	71.19	Social isolation in adulthood Coercive psychiatric treatment Unemployment (at time of offense) Separation from the family/caregivers in the patient’s childhood/youth Failure in school
Lu et al. (24)**	Drug Craving Scale (DCS) Buss-Warren Aggression Questionnaire Revised in China (BWAQ-RC) Impulsivity Scale Security Questionnaire (SQ) Positive Psychological Capital Questionnaire (PPCQ) Toronto Alexithymia Scale (TAS-20) Children’s Perception of Inter-parental Conflict Scale (CPIC) Interpersonal Trust Scales (ITS) Year of birth	Not clearly defined	Gradient boosted regression trees	5-CV and out-of-sample testing techniques	–	–	–	–	Interpersonal trust (ITS) Psychological security (PPCQ) Psychological capital (PPCQ) Parental conflict (CPIC) Alexithymia (TAS-20)
Menger et al. (27)	Doctor textual notes with information, such as patient history, current treatment (e.g., types of medication and therapy), and changes therein Nurse textual notes with information on the current wellbeing and activities of a patient	Violent incidents = incidents concerned violence from patients directed at staff or at other patients, including both verbal and physical aggression in the first 30 days after admission	Recurrent neural network	5-CV	0.79	–	–	–	–
Menger et al. (28)***	The 1,000 most frequent terms in the clinical notes	Violent incident = all threatening and violent behavior of a verbal or physical nature directed at another person	Support vector machine	5-CV	0.76	–	33.4–33.6	93.5–94.7	Site 1: the terms aggressive, reacts, and offered generalize Site 2: the terms verbal, threatening, and aggression
Suchting et al. (17)	328 predictor variables among which full demographic profile, patient vitals (i.e., height, weight, and blood pressure), a comprehensive psychosocial assessment, including histories of early development, education, military service, vocation/work, medical status, psychiatric status, drug/substance use and treatment, nicotine/tobacco use and counseling, abuse (victim or perpetrated physical/verbal/emotional/sexual abuse), legal status, marital status, religious beliefs, financial status, and living situation. Sleep habits, pain status, patient behavior during interview, a risk assessment, and evaluation of patient mood (via the Affective Disorders Rating Scale. General appearance (i.e., hygiene), musculoskeletal system, speech pattern, thought processes and content, perception, depression, affect, insight, judgment, skin integrity, head trauma, suicidal/homicidal/assault ideation, deterioration in function, chemical dependency, hallucinations, and delusions.	Aggressive event = it is coded into the hospital medical record following any episode of uncontrolled verbal or physical aggression that required intervention by and assistance from additional hospital staff to manage the event	Penalized generalized linear modeling	5-CV	0.78	–	–	–	Current living situation (homeless) Legal history – assault conviction Abuse history – witness (other) Abuse history – perpetrated (other) Age
Wang et al. (25)	28 predictive variables among which age, sex, age of onset of psychosis, number of previous psychiatric hospitalizations, comorbid diagnoses of lifetime alcohol, drug, and marijuana abuse or dependence, family histories of psychosis, mood disorders, suicide, ethnicity, primary language, religious identity, age of immigration, childhood trauma and five-factor personality traits from the NEO Five Factor Inventory (NEO-FFI).	Not clearly defined	Random forest	5-CV	0.63	62	32	80	–
Watts et al. (26)*	138 variables among which adverse events in childhood, income, housing, comorbidities, family history, prescribed medications, substance use, and presumed indicators of risk. Variables were transformed via one-hot encoding into new binary variables. This resulted in 156 candidate features.	Patients were divided into violent, non-violent and sexual offenses according to the most recent criminal offense for which they were found not criminally responsible. In cases where multiple crimes were committed, patients were divided according to the most serious offense committed	Elastic net	10-CV	0.88	80.34	83.26	77.42	Sexual vs. violent offenses: paraphilia; previous sexual conviction; dementia/cognitive disorder; living off family support; female support
					0.78	68.79	69.84	67.74	Sexual vs. non-violent offenses: paraphilia; schizoaffective disorder; female gender; history of sexual aggression against others; impulse control disorder
					–	–	–	–	Sexual vs. all offenses: previous absolute discharge; previous sexual convictions; female gender; anti-androgen medication; cluster A personality disorder

BPRS, Brief Psychiatric Rating Scale; PCL-SV, Psychopathy Checklist-Screening Version; HCR-20, Historical, Clinical and Risk Management-20; BIS-11, Barratt Impulsiveness Scale-11.

*Based on Swiss law.

**Five most important variables by a variable importance plot.

***The top three terms with highest within–data set generalizability (ratio).