Introduction

Unconscious violence is defined as aggressive or harmful behavior that occurs without the perpetrator's conscious intent to injure (ResearchBank.ac.nz). For example, a confused patient may physically lash out under delirium or intoxication without deliberate intent (ResearchBank.ac.nz). Such acts often arise from acute cognitive or emotional impairments (e.g. head injury, substance effects, acute stress) that undermine judgment and self-control (ResearchBank.ac.nz). The concept emerged in healthcare; for instance, a New Zealand study noted that pain, nausea, or inebriation can precipitate "unconscious violence" where intent is absent (ResearchBank.ac.nz). Despite its significance, this form of violence is not well-addressed by standard risk tools. Traditional threat assessments focus on rational planning or historical factors, overlooking abrupt loss of control. Threat assessment as a discipline arose in the late 20th century to prevent targeted violence (e.g. school shootings, assaults). Early models from law enforcement and the Secret Service advocated structured approaches combining investigation and intervention (Fein et al.). For example, Fein et al. described threat assessment as identifying individuals who may harm others and devising plans to divert them from violence. Over time, formal tools were developed to evaluate violence risk more systematically. These include actuarial instruments like the Violence Risk Appraisal Guide (VRAG) and the Psychopathy Checklist–Revised (PCL-R), which use statistical algorithms to score known risk factors (Ling et al. 56), and Structured Professional Judgment (SPJ) guides like the HCR-20 that blend research-based factors with clinical judgment ("Structured Professional Judgment Tools"; "General Violence Risk"). The HCR-20, for instance, is an SPJ tool encompassing 20 items across Historical, Clinical, and Risk-Management domains; it is regarded as a "leading violence risk assessment instrument" with extensive validation ("Structured Professional Judgment Tools"; "General Violence Risk"). Despite these advances, significant gaps remain. First, most tools focus on conscious, deliberate violence – e.g. revenge or psychopathic attacks – and may miss impulsive or context-dependent aggression (Ling et al. 55-56). The PCL-R, while predictive of recidivism, mainly captures enduring traits of psychopathy (lack of empathy, antisocial behavior) (Ling et al. 56). Actuarial scales emphasize static factors (e.g. age, prior offences) and seldom incorporate transient emotional states (Ling et al. 55-56). Second, many instruments require extensive data and expert administration, limiting their timeliness and usability. For example, completing the HCR-20 involves hours of interviews and record review ("General Violence Risk"), and its predictive accuracy, though solid (AUC ≈0.70–0.75) (Ling et al. 58), is comparable to other tools but not outstanding ("General Violence Risk"; Ling et al. 58). Third, few tools consider novel risk domains like emotional dysregulation, recent trauma, or acute physiological stressors – factors increasingly implicated in violent behavior. Research in forensic psychology and neuroscience suggests that dynamic psychological and situational factors play a key role in spontaneous aggression. Emotional dysregulation, for example, has been linked to aggression in youth and adults. Adolescents exposed to stressors who exhibit poor emotion regulation show higher rates of later violence (Herts et al. 1111). Similarly, trauma survivors may react violently when triggered by fear or frustration; one review notes that intense emotions (fear, anger, anxiety) can escalate to maladaptive aggression (ResearchBank.ac.nz). Impulsivity is another major risk factor – individuals who act rashly under stress are more prone to sudden violence. Chronic social isolation is also connected to extreme violence: isolated "loners" are overrepresented among mass shooters and assassins (Lankford and Silva). Neuroscientific studies highlight brain mechanisms: dysregulation of prefrontal inhibitory circuits and hyperreactive amygdala responses have been found in aggressive offenders, suggesting unconscious emotional triggers (Ling et al. 56; Herts et al. 1112).

Moreover, machine learning (ML) research indicates promise for improving threat assessment by detecting complex patterns in data that human raters may miss. Recent systematic reviews report that ML-based models in forensic settings often outperform traditional risk tools. Parmigiani et al. found that ML models frequently achieved AUCs above 0.80 in violence prediction, generally surpassing instruments like the HCR-20. For example, Menger et al. applied ML to psychiatric electronic health records to predict inpatient assaults and obtained AUC ≈0.80. More recently, Dobbins et al. developed deep-learning classifiers on clinical notes to forecast violence against healthcare workers, achieving an F1 score of 0.75 versus 0.50 for human clinicians. These advances suggest that data-driven models can capture subtle cues and temporal dynamics beyond static checklists (Parmigiani et al.; Menger et al.; Dobbins et al.). In sum, conventional risk tools emphasize historical and static variables, leaving a critical gap in real-time, situational risk indicators. No widely-used instrument directly targets unconscious violence – violence emerging from impaired states rather than deliberate intent. Therefore, we propose the Unconscious Violence Risk Index (UVRI), a hybrid approach that encodes behavioral cues, emotional states, and environmental triggers into an ML-based threat assessment framework. The rationale is twofold: (1) to fill a conceptual gap by quantifying risk factors linked to involuntary violence (e.g. acute distress, fear responses), and (2) to leverage ML for improved predictive accuracy. This paper reports a comprehensive study of UVRI's development and evaluation. The following sections first survey relevant literature on risk assessment tools, psychological predictors of impulsive aggression, and prior ML applications (Literature Review). We then detail our mixedmethods methodology for building UVRI, present quantitative and qualitative results, discuss implications and limitations, and conclude with recommendations for future research.