The predictive validity and temporal characteristics of the HCR-20v3 for inpatient violence in forensic inpatient settings. An international study

Aggression and violence are common day to day problems in psychiatric settings. However, the optimal means of


Introduction
The prevention and management of violence is a central concern of mental health services (Faay and Sommer, 2021).Severe mental illness, in particular schizophrenia, is associated with a risk of violence both to self and others (Fazel et al., 2014;Whiting et al., 2022), although the nature of the association remains poorly understood.Risk factors for violence by mentally ill people overlap with risk factors for violence in the general population and include age and sex, a past history of violence, positive psychotic symptoms, the emotional impact of psychotic symptoms, drug misuse, treatment compliance, broader societal issues or some combination of these (Buchanan et al., 2019;Coid et al., 2013;Keers et al., 2014;Ullrich et al., 2014;Whiting et al., 2021).
Violence risk assessment, the process by which clinicians evaluate the probability that a patient will be violent to another person in the future, the nature of that violence, and its management are core functions of clinical psychiatry.Ideally risk assessment should be used to guide violence reduction strategies, however it is burdensome and it is often used to justify the use of restrictive practices including involuntary detention, segregation and seclusion (Buchanan et al., 2012;Buchanan and Grounds, 2011).While violence risk assessment has global acceptance (Singh et al., 2014)), there is widespread acknowledgment of its limitations and some acceptance that in reality the origins of violence are complex and subject to chance (Connors and Large, 2023).
In the context of this uncertainty and reflecting these concerns violence risk assessment has evolved over time from clinical judgement through actuarial means to the generally accepted current gold standard of structured professional judgement (SPJ) (Tully, 2017).In SPJ clinicians use their clinical judgement within an evidence-based framework to consider what risk factors are relevant and ignore what is not.
Internationally the HCR-20v3 (http://hcr-20.com/)(Douglas et al., 2014) is one of the leading violence risk assessment guides in clinical forensic psychiatry and correctional settings.Past studies in community psychiatric, adult inpatient and forensic inpatient settings have demonstrated that the HCR-20 can significantly discriminate between patients according to violence risk, with an area under the curve (AUC) that is broadly similar to other risk assessment instruments (Singh et al., 2011).
Within the EU, forensic psychiatry provisions and models of care vary widely (Salize et al., 2023), though every state now has some service to treat patients with a mental disorder and significant violence risk who need forensic psychiatric care.
This study aimed to investigate the predictive validity of the HCR-20v3 for violence in forensic hospital settings and how its properties changed over time.We used an international sample of forensic psychiatric inpatients with a schizophrenia spectrum disorder, as schizophrenia is the most commonly encountered diagnosis in such settings, and where the patients already had a history of significant interpersonal violence (De Girolamo et al., 2021).
We explored several lines of enquiry, first whether the HCR-20v3 is a valid predictor of inpatient violence.Second whether the HCR20v3 Total score would significantly predict inpatient violence at 6-and 12month follow-up.Third we predicted that the HCR-20v3 performance would fall away 6 months after its completion but that by repeating the rating, performance would improve.Finally, we wanted to test which of the HCR-20v3 Total and subscale scores were most strongly associated with future violence.

Study design
We conducted a 12-month prospective cohort study to test the predictive validity of the HCR-20v3 for inpatient violence.At baseline, T0, cases were assessed for their risk of future violence using the HCR-20v3 and then it was repeated after 6-months, T6.The dependent outcome, inpatient interpersonal physical violence was assessed over the first six months (0 to 6) and the second six months (7 to 12) using the Modified Overt Aggression Scale (MOAS).

Procedures
Research centres in each of five countries Austria, Germany, Italy, Poland and the United Kingdom, obtained ethical approval from their relevant bodies.Treating forensic psychiatrists identified potentially eligible inpatients from their caseload.Patients, with capacity, gave their written informed consent after a written and verbal description of the study aims and methods.In every country apart from Italy, where regulations prevented it, patients were reimbursed for their time.

Participants
210 patients with a primary diagnosis of a schizophrenia spectrum disorder (SSD) and a history of significant interpersonal violence agreed to take part.Significant violence was defined as having committed at least one of a homicide, attempted homicide, or a serious violent assault that resulted in significant physical injury to the victim.At recruitment all patients were resident in forensic psychiatric hospitals, in low or medium security, if that distinction existed in each country.We recruited working age adults aged between 18 and 65 years of either sex.Exclusion criteria included a diagnosis of an intellectual disability, a traumatic brain injury, an inability to speak the national language fluently and planned discharge from the inpatient ward within the next month.

Violence risk
The HCR-20v3 is a structured professional judgment guide used to predict the future risk of violence.It assesses 20 factors linked to the risk of violence divided into three subscales: ten Historical factors, H10, five current Clinical factors, C5 and five future Risk management factors, R5.Each factor was rated as present (scored 2), partially/possibly present (scored 1), or absent (scored 0).

Outcome measure
The MOAS (Kay et al., 1988) was used to record and quantify inpatient violent incidents at months 3, 6, 9, and 12 after the first HCR-20v3 assessment.In the subsequent analysis these were collapsed into two 6-month periods, 0 to 6 months, and 7 to 12 months.The MOAS rates domains of violent behaviour, verbal violence, physical violence against property, physical violence towards the self, and physical violence towards other people.In this study we focused only on incidents of interpersonal physical violence in the inpatient setting during the two 6-month follow up periods.The MOAS was rated for each patient from a wide variety of information sources, staff report, patient interviews, and review of the clinical records.

Statistical analysis
To test the predictive ability of the HCR-20v3 Total and subscale scores for incidents of physical violence during the first and second month follow up periods we computed the ROC curves and the corresponding areas under the curve (AUC) statistics, in order to create a "hierarchy" between the predictors.Then we categorized the subjects as high and low risk based on their scores above and below the median HCR-20v3 Total and subscale group scores based on a previously established method (Shepherd et al., 2018), thereby obtaining contingency tables from which we could calculate a comprehensive range of performance indicators including the measures of discrimination; AUC, Diagnostic Odds Ratio, Sensitivity, Specificity and measures of calibration, Positive Predictive Value and Negative Predictive Value as recommended by the RAGEE guidelines (Singh et al., 2015).All analysis were conducted in SPSS (IBM Corp., 2021)

Participant characteristics
All of the 210 forensic inpatients recruited into the study were included in the primary analyses of violence outcomes (Tables 1 and 2).There were 186 men (88.6 %) and 24 women (11.4 %).The majority were under the age of 41 years (64.3 %) and most had not progressed beyond secondary school education.The most common diagnoses were schizophrenia (n = 164, 78.1 %) and schizoaffective disorder (n = 21, 10.0 %), with an average age at first contact with psychiatric services of 25 years (SD 9.2) and a mean duration of illness at study recruitment of 13 years (S.D. 9.4).In total 78.1 % (n = 174) of the sample had a lifetime history of substance misuse and 29 % (n = 61) had a comorbid diagnosis of a personality disorder, with antisocial the most common.
The majority, 77.4 % (n = 161) of the participants were known to psychiatric services before their index offence, and 66.8 % (n = 139) had been hospitalised in the past.68.3 % (n = 142) had been violent before their current index violent offense and over 46 % (n = 97) had engaged in deliberate self-harm or made a suicide attempt.While 88.3 % (n = 121) had been prescribed an antipsychotic, only 13.2 % (n = 18) had been complying with that treatment at the time of their index violence.
During the first 6 months of follow up after the baseline T0 HCR-20 rating, n = 22 (10.5 %) inpatients were involved in physical violence directed towards another person.In the second 6 months, between months 7 to 12 of follow-up, n = 8 (3.8 %) inpatients exhibited violence towards others.

Predictive validity of HCR20v3
AUC values of the HCR20v3 Total and its three subscale scores, Historical (H10), Clinical (C5) and Risk (R5), ranged from 0.538 for R5 at T0 to 0.876 for HCR20v3 Total score at T6 (Table 3).
For violence over the first 6-months after the baseline HCR-20 assessment at T0, the strongest performing predictors of violence were the HCR20v3 Total score with a moderate AUC of 0.746, the Clinical (C5) subscale, AUC of 0.737, and the R5 subscale with an AUC of 0.699, while the Historical (H10) subscale yielded a lower and non-significant AUC value.
Over the second 6 months of follow up, between the seventh and twelfth months, the HCR20v3 Total score completed at baseline (T0) did not remain a strong predictor of violence, with a non-significant AUC value of 0.687.It was bettered by the second HCR20v3 Total score conducted at 6 months, T6, with a higher AUC value of 0.876.Indeed, the AUC values of all risk ratings completed at 6 months performed better, if not statistically so, over the second six months of the study than the same rating completed at baseline over the first six months.The C5 and R5 scores completed at 6 months performed very well, yielding AUC values of 0.855 and 0.767 respectively, for inpatient violence over the second six months.

Sensitivity, specificity, positive predicative value, and negative predictive value
The patient sample was then split into high and low risk subjects based on the median HCR20v3 Total and subscales scores, that was then used to calculate sensitivity, specificity, positive and negative predictive values and diagnostic odds ratios.
At the baseline (T0) median split the HCR20 Total score had the highest and statistically significant diagnostic odds ratios (DOR), 4.74 (CI 1.50-14.98)(Table 4).It had good sensitivity (0.78) and adequate specificity (0.58) and yielded a PPV of 0.16 and NPV of 0.96.In general, the baseline HCR-20v3's performance, especially for the Total score fell away over the second six months of follow up, for example the DOR decreased to 1.02 and was no longer significant (CI 0.17-6.25).However, when the HCR20v3 was updated at 6-months and looking at its performance over the second 6-months, the HCR20v3 Total score had  perfect sensitivity of 1.00, specificity of 0.57, PPV of 0.07 and NPV of 1.00.Furthermore, during the second 6-month follow-up period, every one of the second HCR-20v3 subscale scores performance improved compared to the baseline over the first 6 months.In terms of the DOR: the H10 moved from 1.99 to 6.88, the C5 from 4.24 to 9.48 and the R5 from 4.14 to 6.51.
Finally focussing on the change in HCR-20 Total, Historical, Clinical and Risk subscale ratings between the baseline and the 6-month assessment, only the change in the Historical subscale yielded a significant AUC of 0.763.Its DOR was 8.61, sensitivity 0.60 and specificity 0.85, that yielded positive and negative predicted values of 0.12 and 0.99 respectively.

General considerations
This is the first prospective cohort study to examine the predicative validity of the HCR20v3 for inpatient violence using a multi-national European sample of forensic psychiatric inpatients with SSDs and that calculated sensitivity, specificity, PPV and NPVs as well as the more usual AUC statistics.It is one of the first studies to assess the stability and performance of the HCR-20v3 over time and also to look at the performance of HCR-20v3 score changes over time.Although the number of subjects in the study was comparatively large there were still relatively few violent incidents, especially in the second six month follow up period, that probably contributed to the broad confidence intervals and some non-significant results.
As predicated, the HCR20v3 Total and Clinical subscale scores significantly predicted inpatient violence over the following 6 months of follow-up.The Historical subscale baseline (T0) score was not significant over the first six months, but when rated again at six months, its performance over the second six months improved.Indeed, the absolute change in the Historical scale scores between baseline and 6 months yielded the only significant AUC statistic, and the best DOR, sensitivity and specificity.Finally, all the scales performed better after the second rating over the second 6 months than the first rating over the first six months.
The strong AUC values over 6 months of the HCR-20 Total and Clinical subscale scores were consistent with other studies (Chen et al., 2023;de Vogel et al., 2014;O'Shea et al., 2014).Our data strongly suggests that the 6-month predictor window, in the inpatient setting, is the upper time limit for the HCR-20v3 and that after that amount of time the prediction accuracy of the HCR20v3 falls away.Furthermore, completing a second rating of the HCR-20v3 at 6 months (T6) yielded better predictors of violence over the second 6 months, than did the same items assessed at baseline.This raises the prospect that the ability of these risk factors to inform violence risk assessment improved over time.The mechanism for that was not evaluated in this study but it could be a treatment effect, or it could be that the raters were simply better at identifying the presence of the risk factors at the second rating, and their ratings were more accurate.A similar improvement in suicide risk assessment accuracy has already been found linked to better clinician awareness of risk factors (Bruer et al., 2018).The data of the performance of the change in Historical score between baseline and the second six months is consistent with the suggestion that, as this subscale relates to historical items that should not change much, the raters' awareness of their presence may improve, perhaps through the availability of new historical information about the subjects.
The data reaffirm earlier results that the C5 subscale manages to capture the status of current, potentially dynamic risk factors that are particularly germane to understand violent incidents in patients with SSDs in the inpatient setting.The Historical subscale had the lowest AUC values between 0.628 to 0.781.Other studies confirm poor to low predictive power of the Historical subscale in the inpatient setting (Dernevik, 1998;Doyle et al., 2002;O'Shea et al., 2014).However it was striking that of all the subscales and despite its historical nature, it was the change in H10 score between baseline T0 and the second rating at months, T6 that yielded the best AUC and DOR.These two findings suggest a couple of conclusions.Firstly that the impact and relevance of some violence risk factors may vary between community and inpatient settings, with for example the role of substance misuse being an obvious example.Secondly it suggests that while in the in-patient setting the dynamic risk factors may be more relevant for understanding violence risk, that nevertheless it remains critical for clinicians to also have a good understanding of the historical foundations of each patients risk.
Using the median scores of HCR20v3 Total and subscales as a cut-off revealed diagnostic odds ratios between just under 2 for the Historical subscale at baseline to over 9 for the change in HCR-20 Clinical score between baseline and 6 months, with all subscale scores at 6 months yielding diagnostic odds ratios greater than 6.These are the ratios of the odds that a subject's scores put them into a high-risk classification and that they were then violent compared to the odds that they were given a high-risk classification and they were not violent.The importance of this is that patients who score above the median score of their peers at months, on the Clinical and Risk subscales, were 6 and 9 times more likely to be violent over the second six months of follow-up.
Of equal relevance to clinicians when determining day to day violence risk may be the data on sensitivity, specificity, PPV, and NPV for the HCR20v3.Sensitivity here measures the proportion of actual positives who were correctly identified, in this case those who scored above the median on a scale and later went on to be violent in the inpatient setting.The specificity measured the proportion of negatives who were correctly identified, in this case those who scored less than the median and then were not violent.Positive and negative predictive values measure respectively the proportion of those identified at high or low risk who then did and did not go on to be violent.
At 6-months, the HCR20v3 Total and Clinical scores performed the best with a sensitivity of 78 % and 72 % respectively, low PPV of 16 %, but high NPV of 96 %.At 12-month follow-up, the Total score completed at 6 months demonstrated the best sensitivity, 100 % followed by the Historical, Clinical and Risk subscale scores, 80 %, but very low PPV of between 6 and 8 %.However once again the negative predictive validity (NPV) was universally high at 12 months based on the second HCR-20 ratings completed at 6 months, at 99 % or greater.While it is likely that in part this comes as a product of having a modest sample and the primary outcome, physical violence in the in-patient setting, was a relatively rare event, it does suggest that within the inpatient forensic setting, over a six-month window that those subjects who score below the sample median and are thus rated as low risk, then it is unlikely that they will be violent.The other important conclusion is to reiterate the importance of respecting the HCR-20v3's 6-month time window and completing a second rating at 6 months.The data showed that a second rating at 6 months will be more accurate.

Clinical implications
Earlier work {Salize, 2023 #53} has shown that internationally it is accepted that mentally disordered offenders who remain dangerous should be detained in some form of therapeutic custody and treated to reduce their risk of future violence.Equally it is accepted that people with disabilities, including serious mental illness, should not be deprived of that liberty unlawfully or arbitrarily.It is then incumbent on clinicians to ensure that they use the best available methods to support the detention of mentally disordered offenders in forensic psychiatric hospitals and also to ensure that their ongoing detention and the use of other restrictions are reduced to the absolute minimum necessary to ensure patient and public safety.
Th evidence from this study suggests a framework for clinicians that when a patient with a schizophrenia spectrum disorder and a history of significant violence scores low for risk on repeat ratings of the HCR-20v3 and assuming that other considerations allow it, then restrictions on the patient should be reviewed, before re-evaluating the HCR-20 within 6 months to determine if the patient has remained at low risk.This strategy begins to provide an evidence-based framework to work towards safe and effective relaxation of restrictions and monitoring in the forensic psychiatric setting.Equally this data suggests that clinicians' reliance on the HCR-20v3 to justify the imposition of significant restrictions on those who are rated as high risk are often unfounded or it may be that the risk management measures are so effective that the risk of violence in the high risk subjects is very effectively managed.
It is accepted that there are multiple factors that need to be considered when it comes to the assessment of violence and violence risk management in the inpatient setting.Factors can include the built environment, ward dynamics such as patient acuity and turn over, staff experience and the strength of relational and procedural security measures.All of these can play a role in influencing rates of inpatient violence beyond the individual patient specific factors captured in the HCR-20v3.Furthermore, while clinically, we strive to improve our understanding of those factors and manage them, it must equally be recognised that when focussing on violent behaviour that unmeasurable dynamic factors, that lie outside the patient will always play a role as will chance (Connors and Large, 2023).To some extent our data suggest that over time risk assessment improves, possibly by reducing unknowns or because chance factors become less influential.Equally it must be recognised that chance outcomes will occur and cannot be predicted or prevented and that a psychiatric and indeed a legal system that stives to achieve very low risk in the community must inevitably impose draconian restrictions.

Limitations
Despite the relatively large international sample, this study cannot be generalised to the wider forensic psychiatric population.Many potentially eligible patients refused to participate, and in line with good ethical practice, we were unable to collect any data on them.Assessment of the patients' symptoms was based on current interview, rather than an assessment at the time of their violent offence.MOAS ratings may have missed comparatively minor acts of violence during follow-up.Although the study was designed to be sex balanced, that was not achieved and so that issue was dealt with through a modelling adjustment.
This study was deliberately designed as an international multi-site collaboration in order to maximise the sample size of this hard to engage subject group and to take advantage of our international collaboration.Equally one of our earlier publications {Salize, 2023 #53} highlighted that across the EU while it is universally accepted that mentally disordered offenders who remain dangerous should be taken into protective custody and treated to reduce future, risk, there was considerable variation in the systems and processes for dealing with mentally disordered offenders.In order to reduce the impact of those potential source of variation we only recruited subjects from forensic psychiatric inpatient hospitals in the five countries.We applied strict diagnostic criteria to ensure that all subjects had an established schizophrenia spectrum disorder and applied strict predetermined criteria to act as a threshold to define past acts of serious violence.Thus, we were confident that our research subjects were homogeneous and representative.

Conclusions
This study investigated the predictive validity of the HCR-20v3 for inpatient violence in forensic settings and examined how its properties changed over time.It showed that the HCR-20v3 is a valid tool to assess the risk of inpatient violence, particularly over a six-month period.The HCR-20v3 Total score and its Clinical and Risk subscales, demonstrated good predictive validity for violence over six months and suggests the importance of considering more dynamic risk factors in patients with SSD in the inpatient setting.Additionally, the results suggest that HCR-20v3 performance improved over time, perhaps as awareness of the presence of risk factors improved.
The sensitivity and specificity of the HCR-20v3 varied across time and the subscales.At 6 months, the Total and Clinical scores demonstrated good sensitivity but low positive predictive value.However, the negative predictive value was consistently high, suggesting that patients rated as low risk on the HCR-20v3 were unlikely to be violent.These findings have important implications for clinicians when assessing violence risk and making decisions especially about relaxing restrictions in forensic psychiatric settings.The study also emphasizes the need to respect the six-month time rating window for the HCR-20v3.

Table 1
Socio-demographic characteristics of forensic patients.
* Frequencies and percentages (for categorical variables) and mean and standard deviations (for continuous variables) have been evaluated only for valid cases (i.e., all cases with no missing data).#This variable is related to a multiple option question: consequently, the sum of column percentages is not equal to 100 %.

Table 2
Baseline clinical characteristics of forensic patients.
* Frequencies and percentages (for categorical variables) and mean and standard deviations (for continuous variables) have been evaluated only for valid cases (i.e., all cases with no missing data).SSD Schizophrenia spectrum disorder PANSS Positive and negative syndrome scale.

Table 3
AUC for HCR20 total and subscale scores for interpersonal physical violence at 6-and 12-month follow-up.
AUC = Area under the curve.M.Picchioni et al.

Table 4
Sensitivity, specificity, and predictive values of the HCR20v3 total score and subscales using median as cut-off.
*All cases reported values above median, so Odds Ratio was not applicable.