Can risk assessment predict suicide in secondary mental healthcare? Findings from the South London and Maudsley NHS Foundation Trust Biomedical Research Centre (SLaM BRC) Case Register

Can risk assessment predict suicide in secondary mental healthcare? Findings from the South... Purpose The predictive value of suicide risk assessment in secondary mental healthcare remains unclear. This study aimed to investigate the extent to which clinical risk assessment ratings can predict suicide among people receiving secondary mental healthcare. Methods Retrospective inception cohort study (n = 13,758) from the South London and Maudsley NHS Foundation Trust (SLaM) (London, UK) linked with national mortality data (n = 81 suicides). Cox regression models assessed survival from the last suicide risk assessment and ROC curves evaluated the performance of risk assessment total scores. Results Hopelessness (RR = 2.24, 95% CI 1.05–4.80, p = 0.037) and having a significant loss (RR = 1.91, 95% CI 1.03–3.55, p = 0.041) were significantly associated with suicide in the multivariable Cox regression models. However, screening statis - tics for the best cut-off point (4–5) of the risk assessment total score were: sensitivity 0.65 (95% CI 0.54–0.76), specificity 0.62 (95% CI 0.62–0.63), positive predictive value 0.01 (95% CI 0.01–0.01) and negative predictive value 0.99 (95% CI 0.99–1.00). Conclusions Although suicide was linked with hopelessness and having a significant loss, risk assessment performed poorly to predict such an uncommon outcome in a large case register of patients receiving secondary mental healthcare. Keywords Suicide · Risk assessment · Secondary mental healthcare · Mental pain Introduction Every year, almost one million people die from suicide across the world [1], which appears to have increased since Javier-David Lopez-Morinigo and Andrea C. Fernandes the start of the 2007 economic recession [2]. Indeed, sui- contributed equally to this work and should be jointly cide represents one of the three leading causes of death in acknowledged as first named authors. the most economically productive age group (15–44 years) Electronic supplementary material The online version of this article (https ://doi.org/10.1007/s0012 7-018-1536-8) contains supplementary material, which is available to authorized users. * Javier-David Lopez-Morinigo South London and Maudsley NHS Foundation Trust, javier.lopez-morinigo@kcl.ac.uk London, UK Department of Psychiatry, Marqués de Valdecilla University Department of Psychosis Studies, Institute of Psychiatry, Hospital, IFIMAV, School of Medicine, University Psychology and Neuroscience, King’s College London, De of Cantabria, Santander, Spain Crespigny Park, PO Box 68, London SE5 8AF, UK Centro Investigación Biomédica en Red de Salud Mental CAS Behavioural Health, London, UK (CIBERSAM), Madrid, Spain Department of Psychological Medicine, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, UK Vol.:(0123456789) 1 3 1162 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 [3]. Of concern, suicide rates in the UK have shown no Methods reduction over the past 5 years [4]. Up to 90% of people who complete suicide are found Participants to have had a ‘psychiatric disorder’ [5], contributing to 47–74% of the population risk of suicide, with half of peo- As stated, the sample was derived from the SLaM BRC ple completing suicide meeting retrospectively applied cri- Case Register. SLaM is one of Europe’s largest mental teria for depression [6, 7]. Although it could be envisaged health services, providing secondary mental healthcare that secondary mental health services may play a crucial to four boroughs in South-East London (UK): Lambeth, role in ‘suicide prevention’ [8], over two-thirds of those Southwark, Lewisham and Croydon. Approximately, who take their lives in the UK have not received secondary 1.23 million inhabitants reside in this geographic catch- mental healthcare in the year before death [4]. ment area, which as a whole was found to be comparable Risk assessment in mental health services might con- with other populations of London in terms of age, gen- ceivably help reduce suicide rates, and the UK Depart- der, education and socio-economic status distributions ment of Health [9] and 2004 NICE guidelines [10] recom- [20, 21]. Fully electronic health records have been in use mended the use of structured clinical risk assessments. across all SLaM services since 2006, and in 2007–2008, However, completed suicide thankfully remains a very the Clinical Record Interactive Search (CRIS) system was uncommon event, and two early studies warned of the high built which renders de-identified copies of records availa- number of false positives picked up to detect the major- ble for research use with appropriate governance structures ity of suicides using this approach [11, 12]. In particular, [20]. CRIS received ethical approval as an anonymised even a hypothetical test with a sensitivity and specificity of data resource for secondary analyses from the Oxford C 99% in a high-risk population (defined as a suicide rate of Research Ethics Committee (reference: 08/H0606/71+5), 250/100,000/years) cannot predict suicide beyond a 20% and currently accesses data on over 300,000 patients [21]. level of efficiency [ 11]. Consistent with this, recent meta- The same research ethics approval also covers the pseu- analyses have concluded that risk scales have a limited donymised linkage between CRIS data and those from the role in predicting suicidal behaviour [13–15], although Office for National Statistics (ONS) in April 2015 [22], there are important issues of between-study heteroge- which registers all deaths in the UK and the official cause neity [13]. In keeping with these meta-analyses, a 2017 of death, including suicide and the method of suicide multicentre study in the UK [16] replicated the limited according to ICD-10 classification [23]. use of risk scales to predict repeated self-harm, which is Those patients who had received SLaM care (i.e., had in line with our previous report on risk assessment and at least one face-to-face contact with a clinical member of suicide by patients with schizophrenia spectrum disorders staff) over the period from 1st January 2007 to 1st April under secondary mental healthcare [17]. However, using 2015 and had at least one suicide risk assessment docu- risk assessment scales continues to be common clinical mented within the study period were included. Those who practice [18]. died from suicide within the study period were compared In addition, the extent to which risk assessment can with those who did not. The analysis described here was predict suicide mortality (rather than ‘self-harm’, ‘sui- based on a surveillance period from 1 January 2007 to 1 cide attempts’ or ‘suicidal behaviour’) in a large sample April 2015, the rationale being that the electronic clinical of mental health service users irrespective of diagnosis, records coverage became complete across all SLaM ser- which also changes over time [19], has not been examined vices during 2006 and, at the time of the analysis, the last to date. Within this context, we investigated the perfor- death certification linkage had been accomplished in the mance of all full suicide risk assessments from the South beginning of April 2015. London and Maudsley (SLaM) Biomedical Research Centre (BRC) Case Register (in South-East London, UK) over 2007–2015 to predict suicide. Whilst anticipating that Measures some risk factors would be statistically associated with a higher risk of suicide, namely previous suicide attempts, Risk assessment suicidal ideation, hopelessness, alcohol/drugs and impul- sivity, we sought to clarify positive predictive values at ‘Full risk assessment’ is a compulsory target across the different levels of raised risk, as well as the extent to which Trust when ‘high risk’ is determined from a ‘brief risk a risk assessment might allow clinicians to rule out risk: assessment’, which is mandatory for all active cases. All i.e., the extent to which ‘low-risk’ patients would not end patients who have been seen by a member of clinical staff their lives. have a ‘brief risk assessment’ documented, which is a 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1163 narrative record of the patient’s risk: (1) to one’s self; (2) in front of a vehicle)—X80, X81; suicide by unspecified to others and (3) from others. If the patient is deemed at means—X84; and undetermined cause of death—Y10-34. ‘high’ risk in any of these domains, a ‘full risk assess- Those with an ‘undetermined cause of death’ code were con- ment’ needs to be completed and updated over time, which sidered as suicides, because in the UK most ‘open verdicts’ consists of a structured assessment taking the form of pre- have been reported as very likely to be suicides [26]. sent/absent tick-boxes enquiring about widely recognised risk factors for three major clusters: suicide, violence and Statistical analysis self-neglect. Full risk assessment is entered into CRIS as structured information separately from clinical free- First, for descriptive purposes, for all SLaM ‘active’ ser- text entries. Hence, information on those who had a ‘full vice users (i.e., at least one face-to-face contact with a staff risk assessment’ documented (compared to those without member) over the study period (2007–2015), we investigated ‘full risk assessment’, including those with ‘partial’ risk suicide rates differences between those with/without full risk assessment, i.e., only some items out of the 15-item full assessment. suicide risk assessment) can be reliably extracted from Second, in those with at least one full risk assessment CRIS. For the purposes of this study, only those patients documented (i.e., the study sample), risk assessment individ- with a ‘full suicide risk assessment’ were included, that ual items and total scores, as independent variables, entered is, those patients with ratings on the 15 items included into Kaplan–Meier survival analyses and Cox regression in the full suicide risk assessment, which is available in models [27], respectively, to investigate associations with Appendix  1 (supplementary material available online). time to suicide. Proportional hazards assumptions were Positive responses can be summed to create total scores, checked as standard for Cox procedures and no evidence of i.e., the higher the score the greater the suicide risk, which violation was found, i.e., the survival curves for two strata yielded good internal consistency (Cronbach α coefficients (determined by the values for the covariates) had ‘hazard’ of 0.69) [24]. The most recent full suicide risk assessment functions, which were ‘proportional’ (or constant) over time. was considered for this study. Age, gender, religion, employment and marital status, ethnic group, IMD and primary psychiatric diagnosis were entered Demographic and clinical covariates as covariates. For the survival analyses, the follow-up period began at the time of the last risk assessment and the end Demographic and clinical covariates included age at the date was the date of death (including suicide) or the censor- time of risk assessment, gender, ethnicity, religion, marital ing point (last face-to-face contact, date of death from non- status, employment status, social deprivation and ICD-10 suicide causes or 1st April 2015, whichever came sooner). diagnosis [23]. In addition, receiver operating characteristic (ROC) Social deprivation was scored through an anonymous link curves [28], which compare the true positive rate (i.e., sen- created in CRIS between lower super output area residence sitivity) with the false positive rate (i.e., ‘1-specificity’) at code of the latest permanent address (a geographic unit different cut-off points for the parameter (risk assessment comprising approximately 400 households) and summary total score in our study), were plotted to analyse the per- data for that area from 2007 UK Census output. Thus, the formance of risk assessment total scores to predict suicide. Index of Multiple Deprivation (IMD) is derived from seven In particular, sensitivity, specificity, positive and negative domains: income, employment, health, education, housing predictive values, the area under the curve (AUC), likeli- and services, crime and environment [25]. hood ratios (positive and negative) and diagnostic odds ratio ICD-10 diagnoses [23] were reached by consensus by (OR) were investigated for the best cut-off point, including the treating multidisciplinary team, including input from a 95% confidence intervals (CI) for each statistic at each risk senior consultant psychiatrist. Specifically, several clinically assessment total score. Positive and negative predictive val- meaningful categories were created as follows: ‘organic ues are the probability that subjects with a positive (high mental disorders’ (F0), ‘substance abuse’ (F1), ‘schizophre- risk) result will truly have the outcome of interest (in this nia spectrum disorders’ (F2), ‘mood disorders’ (F3), ‘neu- study, death from suicide), and the probability that subjects rotic disorders’ (F4) and ‘all others’ (F5–F7). with a negative (low risk) result will not have such an out- come, respectively. The AUC is a measure of how well risk Suicide method assessment total score can distinguish those who will die from suicide from those who will not. Likelihood ratios are Suicide method was ascertained using death certificate [ 22] the likelihood that a given test result would be expected in ICD-10 codes [23] and the following groups were used to a patient who took his/her life compared to the likelihood define this: poisoning—X64; hanging—X70; drowning— that same result would be expected in a patient who did not X71; cutting—X78; jumping (either from high place or end his/her life. Diagnostic ORs are the ratio of the odds of 1 3 1164 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 the test being positive (high risk) if the subject ended his/ died by suicide. Taking into account the observation time of her life relative to the odds of the test being positive if the each patient, the analysed sample contributed to 80,769.17 subject did not die from suicide. person-years, which yielded a suicide rate of 100.28/100,000 A significant level of 5% (two-tailed) was set for all the person-years. In those with a partial risk assessment (n = 1409), above analyses, which were performed using the statistical the suicide rate was 51.60/100,000 person-years, while for those package R (version 3.20) [29]. with no risk assessment completed (n = 80,340), the suicide rate was 88.34/100,000 person-years. These differences were not statistically significant (X = 6, df = 4, p = 0.19). Results The baseline demographic and clinical characteristics of the sample (n = 13,758) and differences between those who Study sample took their lives and those who did not are presented in Table 1. Although there was a higher male predominance in the suicide Over 2007–2015, there were 99,507 SLaM ‘active’ cases, i.e., completers group than in those who did not end their lives those who had at least one face-to-face contact with a SLaM (OR = 1.67, 95% CI 1.04–2.69, p = 0.03), no further signifi- staff member over that period (2007–2015), of whom 358 were cant differences emerged in age at first presentation, religion, ascertained as having died by suicide. Of all these active SLaM marital status, ethnicity, living status, employment, social dep- service users, 13,758 subjects had all suicide-related items com- rivation, first language (English vs. all others) and ICD-10 pleted on a full risk assessment, and a further 1409 had incom- diagnoses. Hanging was the most common suicide method plete data (with at least one item rated). Of the 13,758 individu- (n = 28). Twenty-one subjects received an undetermined cause als, who formed the study sample, 81 were recorded as having of death. There were no suicides by firearms. Table 1 Demographics and Suicides Non-suicides p value clinical characteristics of the N = 81 N = 13,678 sample Mean ± SD Mean ± SD Age at risk assessment (years) 41.3 ± 12.2 40.6 ± 11.5 0.60 Social deprivation 28.6 ± 13.4 28.6 ± 12.3 0.98 n (%) n (%) OR (95% CI) Gender (males) 56 (69.1) 7823 (57.2) 1.67 (1.04–2.69) 0.03 Marital status (unmarried) 72 (88.9) 11,909 (87.1) 1.10 (0.55–2.21) 0.78 Employment status (unemployed) 30 (37.0) 4662 (34.0) 1.57 (0.48–5.14) 0.46 Living status (alone) 19 (23.5) 3093 (22.6) 1.04 (0.57–1.89) 0.90 Religion (yes) 20 (24.7) 3177 (23.2) 1.00 (0.60–1.67) 0.98 Ethnicity  White 50 (61.7) 6916 (50.6) 1.85 (1.02–3.35) 0.04  Black 14 (17.3) 3525 (25.8) Ref.  Others 17 (21.0) 3237 (23.7) 1.37 (0.67–2.78) 0.38 First language (non-English) 40 (49.4) 5971 (43.7) 1.53 (0.99–2.37) 0.06 Diagnosis  Organic disorders 1 (23.0) 270 (2.0) Ref.  Substance use disorders 10 (12.3) 1232 (9.00) 1.95 (0.25–15.2) 0.53  Schizophrenia spectrum 31 (38.3) 5713 (41.8) 1.28 (0.17–9.34) 0.81  Mood disorders 24 (26.6) 2643 (19.3) 2.28 (0.30–16.9) 0.42  Neurotic disorders 5 (6.17) 860 (6.30) 1.60 (0.19–13.7) 0.67  Other diagnoses 10 (12.3) 2960 (21.6) 0.85 (0.10–6.60) 0.87 Method  Hanging 28 (34.6)  Intoxication 6 (7.4)  Jumping 3 (3.7)  Unspecified means 16 (19.7)  NA 7 (8.6)  Undetermined cause of death 21 (25.9) 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1165 Table 2 Unadjusted univariate Risk factor N Events expected Events Log-rank test RR (95% CI) p value analyses: risk assessment items observed Previous suicide attempts  Absent 7004 40 24 14.1 2.46 (1.51–4.01) < 0.001  Present 5657 33 49 Violent method  Absent 9167 50.7 37 14.7 2.46 (1.53–3.97) < 0.001  Present 3069 17.3 31 Plan to end life  Absent 11,551 64.5 53 22.5 3.37 (1.98–5.77) < 0.001  Present 1258 6.5 18 Suicidal ideation  Absent 10,583 61.7 51 9.4 2.06 (1.30–3.31) 0.002  Present 2843 15.3 26 Hopelessness  Absent 9419 55.1 37 21.7 2.79 (1.78–4.37) < 0.001  Present 3811 20.9 39 Distress  Absent 8617 51.5 42 5.1 1.66 (1.07–2.60) 0.024  Present 4651 26.5 36 No control over life  Absent 9290 55.1 42 11.3 2.13 (1.36–3.35) < 0.001  Present 3576 20.9 34 Alcohol/drugs  Absent 7834 44.1 36 3.8 1.60 (0.99–2.51) 0.051  Present 4971 27.9 36 Impulsivity  Absent 7292 43.5 34 4.7 1.64 (1.05–2.57) 0.030  Present 5636 33.5 43 Living alone  Absent 8358 48.3 46 0.3 1.12 (0.72–1.76) 0.603  Present 4949 30.7 33 Poor physical health  Absent 9616 56.2 56 0 1.01 (0.61–1.67) 0.964  Present 3431 20.8 21 Significant loss  Absent 8243 50 39 7.5 1.88 (1.20–2.96) 0.006  Present 4256 24 35 Disengagement  Absent 9841 58.3 48 7.2 1.85 (1.17–2.91) 0.007  Present 3277 19.7 30 Recent hospital discharge  Absent 10,973 65.1 56 7.3 1.93 (1.20–3.14) 0.007  Present 2399 13.9 23 Family history  Absent 5921 32.88 35 1.2 0.46 (0.11–1.90) 0.268  Present 730 4.12 2 1 3 1166 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 Risk assessment factors Differences in individual risk assessment items between suicide completers and non-completers (Kaplan–Meier survival analyses) are presented in Table 2. The following items were significantly associated with an increased risk of suicide: previous suicide attempts (RR = 2.46, 95% CI 1.51–4.01, p < 0.001), previous use of a violent method (RR = 2.46, 95% CI 1.53–3.97, p < 0.001), plans to end life (RR = 3.37, 95% CI 1.98–5.77, p < 0.001), suicidal idea- tion (RR = 2.06, 95% CI 1.30–3.31, p = 0.002), hopeless- ness (RR = 2.79, 95% CI 1.78–4.37, p < 0.001), distress (RR = 1.66, 95% CI 1.07–2.60, p = 0.024), lack of con- trol over life (RR = 2.13, 95% CI 1.36–3.35, p < 0.001), impulsivity (RR = 1.64 95% CI 1.05–2.57, p = 0.030), having a significant loss (RR = 1.88, 95% CI 1.20–2.96, p = 0.006), disengagement (RR = 1.85, 95% CI 1.17–2.91, p = 0.007) and recent hospital discharge (RR = 1.93, 95% CI 1.20–3.14, p = 0.007). Cox regression analyses of the relationship between those risk assessment factors significantly associated Fig. 1 ROC curve for risk assessment total scores with risk of suicide (see Table 2) are presented in Table 3. The left column shows the results only after adjusting strengthened coefficient but wider confidence intervals), the analyses for gender, which had been the only signifi- were substantially attenuated. cant baseline characteristic associated with suicide risk (see Table  1). All the risk factors remained significant Risk assessment overall performance (Table 3, left column). The right column of Table 3 pre- sents coefficients from a model containing all factors; in The diagnostic accuracy statistics (and 95% CI) for each risk this, only hopelessness (RR = 2.24, 95% CI 1.05–4.80, assessment total score are detailed in Table 4. ROC curve p = 0.037) and having a significant loss (RR = 1.91, 95% analyses for risk assessment total scores, which are shown CI 1.03–3.55, p = 0.041) remained statistically significant in Fig. 1, found the optimal cut-off point to be 4–5 (above predictors of suicide. All the other associations, apart which the risk would be ‘high’; below which the risk would from that with previous suicide attempts (which showed a Table 3 Adjusted Cox a b Individual items RR (95% CI) p value RR (95% CI) p value regression analyses: risk assessment items Gender (male) 1.67 (1.04–2.67) 0.03 1.70 (0.95–3.02) 0.077 Suicidal history 1.67 (1.04–2.67) < 0.001 2.00 (0.89–4.53) 0.094 Violent method 2.58 (1.58–4.21) < 0.001 1.31 (0.65–2.65) 0.453 Plan to end life 2.47 (1.53–3.98) < 0.001 1.20 (0.54–2.63) 0.657 Suicidal ideation 2.13 (1.32–3.41) 0.002 0.82 (0.40–1.67) 0.577 Hopelessness 2.90 (0.85–4.55) < 0.001 2.24 (1.05–4.80) 0.037 Distress 1.72 (1.10–2.70) 0.020 0.90 (0.48–1.68) 0.746 No control over life 2.21 (1.40–3.48) < 0.001 0.95 (0.47–1.91) 0.881 Impulsivity 1.65 (1.05–2.58) 0.029 1.05 (0.56–1.97) 0.875 Significant loss 1.95 (1.23–3.07) 0.004 1.91 (1.03–3.55) 0.041 Disengagement 1.85 (1.17–2.92) 0.008 1.16 (0.61–2.20) 0.653 Recent hospital discharge 1.97 (1.21–3.21) 0.006 1.45 (0.77–2.74) 0.247 Bold values indicate statistically significance p < 0.05 Analysis adjusted for gender only Fully adjusted analysis 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1167 Table 4 Risk assessment total scores (number of risk factors present) and diagnostic accuracy statistics with 95% confidence intervals Risk assess- Suicides Non-suicides Sensitivity, Specificity, Positive pre- Negative Likelihood Likelihood Diagnostic ment (Total % (95% CI) % (95% CI) dictive value, predictive ratio, + (95% ratio, − (95% OR (95% CI) score) % (95% CI) value, % CI) CI) (95% CI) 0 2 1113 98 (91–100) 8 (8–9) 1 (0–1) 100 1.1 (1.0–1.1) 0.3 (0.1–1.2) 3.5 (0.9–14.3) (99–100) 1 3 1748 94 (86–98) 21 (20–22) 1 (1–1) 100 (100– 1.2 (1.1–1.3) 0.3 (0.1–0.7) 4.0 (1.6–9.9) 100) 2 7 2042 85 (76–92) 36 (35–37) 1 (1–1) 100 (100– 1.3 (1.2–1.5) 0.4 (0.2–0.7) 3.2 (1.7–5.9) 100) 3 8 1909 75 (64–84) 50 (49–51) 1 (1–1) 100 (100– 1.5 (1.3–1.7) 0.5 (0.3–0.7) 3.02 (1.8–5.0) 100) 4 8 1727 65 (54–76) 62 (62–63) 1 (1–1) 100 1.7 (1.5–2.0) 0.5 (0.4–0.7) 3.1 (2.0–5.0) (99–100) 5 12 1443 51 (39–62) 73 (72–74) 1 (1–1) 100 1.9 (1.5–2.3) 0.7 (0.5–0.8) 2.8 (1.8–4.3) (99–100) 6 10 1161 38 (28–50) 81 (81–82) 1 (1–2) 100 2.1 (1.6–2.7) 0.8 (0.6–0.9) 2.7 (1.73–4.3) (99–100) 7 10 836 26 (17–37) 88 (87–88) 1 (1–2) 100 2.1 (1.4–3.0) 0.8 (0.7–1.0) 2.5 (1.5–4.1) (99–100) 8 9 712 15 (8–24) 93 (92–93) 1 (1–2) 100 2.0 (1.2–3.5) 0.9 (0.8–1.0) 2.23 (1.2–4.1) (99–100) 9 4 427 10 (4–19) 96 (96–96) 1 (1–3) 100 2.4 (1.2–4.7) 0.9 (0.8–1.0) 2.6 (1.2–5.3) (99–100) 10 4 291 5 (1–12) 98 (98–98) 1 (1–4) 100 2.5 (0.9–6.6) 1.0 (0.9–1.0) 2.6 (0.9–7.1) (99–100) 11 3 167 1 (0–7) 99 (99–99) 1 (0–5) 100 1.7 (0.2– 1.0 (1.0–1.0) 1.6 (0.2–12.1) (99–100) 11.7) 12 1 64 0 (0–4) 1 (1–1) 0 (0–9) 99 (99–100) 0 (0–0) 1.0 (1.0–1.0) NA 13 0 28 NA NA NA NA NA NA NA 14 0 10 NA NA NA NA NA NA NA be ‘low’), with a sensitivity of 0.65 (95% CI 0.65–0.76), we found that although a number of risk factors were sig- specificity of 0.62 (95% CI 0.62–0.63) and an AUC of 0.67 nificantly associated with suicide in the bivariate analyses (95% CI 0.62–0.73). The positive and negative predictive (namely, being male, previous suicide attempts, previous use values were 0.01 (95% CI 0.01–0.01) and 0.99 (95% CI of violent methods, plans to end life, suicidal ideation, dis- 0.99–1.00), respectively. tress, lack of control over life, impulsivity, disengagement from services/non-compliance with medication and recent hospital discharge), only hopelessness and having a signifi- Discussion cant loss remained independent and statistically significant predictors of suicide in the multivariable regression models. Main findings Third, overall risk assessment performed poorly to predict suicide in a large sample of mental health service users, We drew data from a large case register of patients receiving which was in line with our expectations and recent literature secondary mental healthcare in a defined catchment area [13–15]. over a prolonged period (2007–2015) linked with national mortality data and we tested the extent to which structured Comparison with previous literature risk assessment items (individual risk factors and overall scores) predicted suicide. First, as expected, we identified a Of relevance, we did not find a protective effect of com- high number of suicides in a population of patients in sec- pleting suicide risk assessment on reducing suicide rates as ondary mental healthcare (approximately 100.28/100,000 previously suggested [10], which was in line with further person-year), approximately tenfold higher than in the gen- reviews of the NICE guidelines [30]. It might seem intui- eral population (10.9/100,000 person-year) [22]. Second, tive to speculate that those patients with a risk assessment 1 3 1168 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 documented are deemed at a higher risk of suicide by the [45] based on Mann’s model of suicide [42]. Indeed, over treating team and they are more likely to receive a higher 90% of suicide attempters report ‘mental pain’ [46], which is level of input and caution in management. It remains unan- frequently the consequence of a bereavement [40], which is swerable what may have happened if these ‘high risk’ indi- of particular concern after surviving the suicide of a love one viduals had not been administered a risk assessment and to [47]. Moreover, ‘mental pain’ appears to be a contributor to research this would raise ethical issues. However, recording suicide independently of depression [48], which is in full of risk assessment is, to some degree, circular and more agreement with our results. Indeed, the relationship between likely in those patients who will be followed-up by men- suicide and ‘mental illness’ (from a psychiatric perspective) tal health services [31]. In addition, risk assessment in this may be weaker than previously thought, especially in West- cohort may have been completed due to concerns raised ern countries [44]. In keeping with this, neither medication regarding other clusters of risk, such as violence to others compliance nor (psychiatric) diagnosis were associated and/or self-neglect [24]. Moreover, risk assessment comple- with suicide in our large cohort of mental health service tion rates may have been affected by the patient’s legal status users, which may provide further support for a non-medical in some cases, which was not evaluated in this study. For approach to suicide [49]. Hence, those patients receiving instance, those receiving care under restriction, hence sub- secondary mental healthcare at risk of suicide may particu- ject to the UK Mental Health Act 1983 (amended 2007) [32] larly benefit from psychological therapies, as recommended may be more likely to have a risk assessment documented. by the UK NICE guidelines for depression [50], although not In this regard, our findings are of major relevance from a frequently offered [51]. human rights perspective, since these individuals may have In addition, overall risk assessment showed poor predic- been ‘forced’ to undertake an assessment which appears to tive validity, which was unsurprising, given the rarity of the have a limited value, which would also go against the 2015 outcome. In particular, high sensitivity was reached at the UK Code of Practice [33]. price of low specificity (i.e., a very high number of false In terms of risk factors, we replicated the association of positives to identify most of suicides) and vice versa, i.e., being male [1, 34], previous suicide attempts [35], previous reducing the number of false positives (high specificity) use of violent methods [5], plans to end life and suicidal occurred at the expenses of too many false negatives (low ideation [36], hopelessness [37, 38], distress, lack of control sensitivity), thus preventing high-risk patients from care and over life and impulsivity [39], having a significant loss [40], treating ‘healthy’ people unnecessarily, which was in full disengagement from services/non-compliance with medica- agreement with early literature [11, 12]. tion and recent hospital discharge [4] with suicide. Consist- In keeping with this, for the most optimal cut-off point ent with previous models of suicide [41, 42], only hope- (4–5), a very low positive predictive value (1%) and very lessness and having a significant loss remained significant. high negative predictive value (99%) emerged from the anal- However, it should be noted that childhood trauma, which yses. In other words, in this ‘low-risk’ group (those with was not evaluated by our risk assessment questionnaire, was less than four risk factors), there would be still 20 suicides found to have greater effects on suicidality than depression (approximately one quarter of those who ended their lives). and related variables [43], hence it should become part of On the other hand, 6988 individuals (50.8% of the total sam- routine clinical suicide risk assessment. ple) would be classified as ‘high-risk’ patients, although only Over four decades ago, hopelessness was defined as ‘the 61 of these subjects took their lives. The question arises. Is cognitive element of negative expectations’ [37], which was it, therefore, worth managing so many patients as ‘suicidal’ also demonstrated to be the strongest predictor of suicide to prevent a few deaths? More specifically, in times of finan- in outpatients [38], and this we replicated in our results. cial constraints, should so many patients receive high levels Hence, our findings agree, in part, with Mann’s diathesis- of care such as unnecessary admissions to hospital? stress model of suicide [42] regarding the role of hopeless- These findings were consistent with a 2017 literature ness, although impulsivity [39] was not significantly associ- review on ‘the limitations of epistemic uncertainty’ with ated with suicide in our cohort. It should be noted, however, regard to risk assessment, whose overall poor performance that impulsivity, as measured by the SLaM risk assessment, appears to be due to the so-called ‘aleatory uncertainty’. In might refer to a different construct. On the other hand, we short, risk factors change ‘by chance’, which is unpredictable did find that having a significant loss was a predictor of [18]. The concept of risk, therefore, requires a reformulation. suicide independently of other factors, which, in addition Specifically, while suicide risk does not appear to be predict- to the role of hopelessness, was consistent with the classic able, a more prevention-orientated approach may result in theory of ‘suicide as psychache’ [41]. Of note, the concep- better clinical outcomes [52]. tualization of suicide as the consequence of ‘mental pain or However, our findings concerning the association of psychache’ has been recently reconsidered [44] in light of hopelessness and having a significant loss with suicide, decades of relatively unsuccessful neurobiological research which as a whole provide some support for the ‘mental pain’ 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1169 model of suicide [41], which was discussed above, still sug- receiving secondary mental healthcare within a defined geo- gest that suicide may still be, to some degree, predictable. graphic catchment and time period. Since most people living in South-East London requiring secondary mental healthcare Future research receive this from NHS resources, our sample is likely to be representative. In addition, participants were followed-up While we do recommend that risk assessment should for up to 9 years and, in addition to risk assessment ratings, remain part of our routine clinical practice, a more nar- a number of covariates were taken into account. rative (free-text) approach should be taken [53] to better However, the study has some limitations to be borne in capture aspects such as ‘mental pain’, which, based on our mind when drawing conclusions from the results. First, all findings, seems to be more useful in terms of clinical risk participants were mental health service users residing in assessment. In particular, rather than categorising patients as South-East London, which is an inner urban area. Hence, ‘low–medium–high’ risk, the wide range of contributing fac- our findings may not generalise to people with mental health tors to risk should be detailed in relation to the individual’s problems under primary care or those who live in rural mental health problems and the social context and how these areas. Second, the vast majority of SLaM patients (almost factors may change dynamically over time, thus increasing or 90%) did not have a structured risk assessment completed decreasing risk for a given individual, which is what matters and those who did may have been deemed ‘at-high-risk’ by clinically [52]. Patient information from electronic records their treating teams. In other words, it could be still specu- can be easily, safely and securely de-identified to generate lated that assessing risk in all patients receiving care may large datasets for secondary research [54], such as the SLaM reduce risk. Third, although only the last suicide risk assess- CRIS [20, 21]. Specifically, naturalistic language processing ment was considered, risk factors may have changed from (NLP) tools appear to be promising research instruments the time of risk assessment to death. Also, risk assessment to extract statistically analysable clinical information from scores may have been affected by survival, hence those who narrative electronic records, hence determining risk from survived for longer (and therefore received care for a more free-text notes [55]. NLP techniques may assess suicide risk prolonged period of time) may have been rated differently. using information from unstructured questions with higher Finally, other factors such as patient’s legal status at the time precision than the classic risk assessment scales [56], thus of risk assessment or a history of childhood trauma, which potentially capturing the presence of ‘mental pain’. were not evaluated in this study, may have affected both Specifically, future studies should examine whether risk risk assessment completion rates and ratings, and the main assessment changes over time, particularly self-ratings outcome measure of this study, namely suicide. shortly before suicide may increase the predictive value of risk assessment. In this regard, mobile phone and web-based Implications and conclusions text messaging may represent a useful tool to self-monitor suicide risk [57], particularly to follow-up people attempt- Our results, therefore, support the notion that neither indi- ing suicide [58] and to assess risk shortly before suicidal vidual risk factors nor a combination of them, i.e., risk events, including suicides. For instance, the classic suicide assessment, can adequately predict suicide in a popula- note may have been substituted by a message left on this tion of patients receiving mental healthcare. Suicide is a new media, which clinicians should discuss with patients very uncommon outcome even in a high-risk group such as and carers when assessing self-harm [59]. In addition, free- patients receiving secondary mental healthcare. Our study text-based risk assessment, which can be researched through suggests that risk assessment cannot predict suicide in the NLP techniques [55], may be more accurate than psychomet- clinical setting due to its very low occurrence, which is in ric scales [56]. Future research should, therefore, switch the full agreement with recent meta-analyses [13–15], although focus from long-term risk factors to short-term risk algo- hopelessness and having a significant loss were linked with rithms, which are more relevant to the clinician [60]. How- suicide, consistent with the classic ‘mental pain’ model of ever, the evidence of the use of communication technologies suicide [41, 44, 46, 48]. in health care and public health, which is known as mobile Meanwhile, further research on suicide prevention health (mHealth) [61], on suicide prevention is limited [62]. [62–64] is warranted. In particular, means restriction remains the first-line strategy to prevent both high-risk groups such as patients receiving mental healthcare [17] Strengths and limitations and the general population [65] from suicide. The use of a large case register linked with national mortal- Acknowledgements We are extremely grateful to all the patients who received care under the SLaM teams within the study period and the ity data allowed us to investigate the role of risk assessment SLaM staff who, therefore, took part in this research project. in predicting suicide in a large sample of patients who were 1 3 1170 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 Funding The development of the SLaM BRC Case Register was 9. Department of Health (DoH) and Social Security (1984) The man- funded by two Capital Awards from the UK National Institute for agement of deliberate self-harm. HMSO, London Health Research and is further supported through the BRC Nucleus 10. National Institute for Health and Clinical Excellence (NICE) funded jointly by the Guy’s and St Thomas’ Trustees and South Lon- (2004) Clinical Guidance 16 Self-harm. http s ://www .nice.org. don and Maudsley Special Trustees. JDLM was funded by the Brit-uk/Guida nce/CG16. Accessed 7 Sep 2017 ish Medical Association via the Margaret Temple Research Award 11. MacKinnon DR, Farberow NL (1976) An assessment of the utility for Schizophrenia. RD is funded by a Clinician Scientist Fellowship of suicide prediction. Suicide Life Threat Behav 6:86–91 awarded by the Academy of Medical Sciences in partnership with The 12. Pokorny AD (1983) Prediction of suicide in psychiatric patients. Health Foundation. JDLM, ACF, HS, RB and RD are part-funded Arch Gen Psychiatry 40:249–257 by the National Institute for Health Research (NIHR) Biomedical 13. Large M, Kaneson M, Myles N, Myles H, Gunaratne P, Ryan C Research Centre and Dementia Biomedical Research Unit at South (2016) Meta-analysis of longitudinal cohort studies of suicide risk London and Maudsley NHS Foundation Trust and King’s College assessment among psychiatric patients: heterogenity in results and London. JDLM and AB also acknowledge funding support from CAS lack of improvement over time. PLoS One 11:e0156322 Behavioural Health. The funders had no role in hypothesis generation, 14. Quinlivan L, Cooper J, Davies L, Hawton K, Gunnell D, Kapur N study design, data collection and analysis, decision to publish or the (2016) Which are the most useful scales for predicting repeat self- manuscript writing. harm? A systematic review evaluating risk scales using measures of diagnostic accuracy. BMJ Open 6:e009297 15. Carter G, Milner A, McGill K, Pirkis J, Kapur N, Spittal M (2017) Compliance with ethical standards Predicting suicidal behaviours using clinical instruments: system- atic review and meta-analysis of positive predictive values for risk Conflict of interest The authors disclosed no competing interests con- scales. Br J Psychiatry 210:387–395 cerning the subject of the study. 16. Quinlivan L, Cooper J, Meehan D et al (2017) Predictive accuracy of risk scales following self-harm: multicentre, prospective cohort Ethical statement This study was performed in accordance with the study. Br J Psychiatry 210:429–436 ethical standards laid down in the 1964 Declaration of Helsinki and its 17. Lopez-Morinigo JD, Ayesa-Arriola R, Torres-Romano B et al later amendments. Under UK law anonymised data can be used without (2016) Risk assessment and suicide by patients with schizophrenia prior consent and data for this study came from CRIS (see “Methods” in secondary mental healthcare: a case–control study. BMJ Open section for details), which received ethical approval as an anonymised 6:e011929 data resource for secondary analyses from the Oxford C Research Eth- 18. Large M, Galletly C, Myles N, Ryan CJ, Myles H (2017) Known ics Committee (reference: 08/H0606/71+5). unknowns and unknown unknowns in suicide risk assessment: evidence from meta-analyses of aleatory and epistemic uncer- tainty. Br J Psychiatry Bull 41:160–163 Open Access This article is distributed under the terms of the Crea- 19. Heslin M, Lomas B, Lappin JM et  al (2015) Diagnostic tive Commons Attribution 4.0 International License (http://creat iveco change 10 years after a first episode of psychosis. Psychol Med mmons.or g/licenses/b y/4.0/), which permits unrestricted use, distribu- 45:2757–2769 tion, and reproduction in any medium, provided you give appropriate 20. Stewart R, Soremekun M, Perera G et al (2009) The South London credit to the original author(s) and the source, provide a link to the and Maudsley NHS Foundation Trust Biomedical Research Centre Creative Commons license, and indicate if changes were made. (SLAM BRC) case register: development and descriptive data. BMC Psychiatry 9:51 21. Perera G, Broadbent M, Callard F et al (2016) Cohort profile of the South London and Maudsley NHS Foundation Trust Biomedi- cal Research Centre (SLaM BRC) Case Register: current status References and recent enhancement of an Electronic Mental Health Record- derived data resource. BMJ Open 6:e008721 1. World Health Organization (WHO) (2014) Preventing suicide: a 22. Office for National Statistics (ONS) (2015) Suicide statistics: global imperative. World Health Organization, Geneva deaths registered in England and Wales. http://www.ons.gov. 2. Oyesanya M, Lopez-Morinigo J, Dutta R (2015) Systematic uk/ons/dat as e ts-andt a bles/inde x .html?pag eS ize=50&sor tB review of suicide in economic recession. World J Psychiatry y=none&sor tD ir ect ion=none&ne wq u er y=suici de&conte nt- 5:243–254 type=Refer ence+table &conte nt-type=Datas et. Accessed 1 Apr 3. Patton GC, Coffey C, Sawyer SM et al (2009) Global patterns of mortality in young people: a systematic analysis of population 23. World Health Organization (WHO), The ICD-10 (1993) Clas- health data. Lancet 374:881–892 sification of mental and behavioural disorders: diagnostic criteria 4. The National Confidential Inquiry into Suicide and Homicide for research. World Health Organization, Geneva by People with Mental Illness (2016) (NCISH) Annual Report: 24. Wu CY, Chang CK, Hayes M, Broadbent M, Hotopf M, Stewart England, Northern Ireland, Scotland and Wales. October 2017. R (2012) Clinical risk assessment rating and all-cause mortality University of Manchester in secondary mental healthcare: the South London and Mauds- 5. Arsenault-Lapierre G, Kim C, Turecki G (2004) Psychiatric diag- ley NHS Foundation Trust Biomedical Research Centre (SLAM noses in 3275 suicides: a meta-analysis. BMC Psychiatry 4:37 BRC) Case Register. Psychol Med 42:1581–1590 6. Cavanagh JT, Carson AJ, Sharpe M, Lawrie SM (2003) Psycho- 25. Office for National Statistics (ONS) (2007) Indices of depriva- logical autopsy studies of suicide: a systematic review. Psychol tion 2007 for super output areas. http://www .neigh bour h ood. Med 33:395–405 s t a t i s t i c s . g o v. uk / d i s s e m i n a t i o n / i n s t a n c e S e l e c t i o n . d o ? J S A l l 7. Hawton K, Harris L, Hall S, Smikin S, Bale E, Bond A (2003) o w ed=tr ue&F unct ion=&%24ph=60_61&Cur r e ntP ag Deliberate self-harm in Oxford, 1990–2000: a time of change in eId=61&s tep=2&dat as e tF am il yId =1893&ins t a nceSe lecti patient characteristics. Psychol Med 33:987–995 on=12283 8&Next.x=19&Next.y=17. Accessed 10 Oct 2009 8. World Health Organization (WHO) (2013) Public health action 26. Linsley KR, Schapira K, Kelly TP (2001) Open verdict v. sui- for the prevention of suicide. World Health Organization, Geneva cide—importance to research. Br J Psychiatry 178:465–468 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1171 27. Cox D (1972) Regression models and life tables (with discus- 49. Shahtahmasebi S (2013) Examining the claim that 80–90% of sions). J R Stat Soc 34:187–220 suicide cases had depression. Front Public Health 1:62 28. Hanley JA, McNeil BJ (1982) The meaning and use of the area 50. National Institute of Care and Clinical Excellence (2016) Depres- under a receiver operating characteristic (ROC) curve. Radiology sion in adults: recognition and management. Clinical guideline 143:29–36 [cg90]. https ://www.nice.org.uk/guida nce/cg90 29. R Core Team (2015) A language and environment for statistical 51. Geekie J, Read J, Renton J, Harrop C (2017) Do English mental computing. R Foundation for Statistical Computing health services know whether they followed N.I.C.E. guidelines 30. National Institute for Health and Clinical Excellence (NICE) with patients who killed themselves? Psychol Psychother Theory (2011) Clinical guidance 133 self-harm. https ://www.nice.org. Res Pract 90:797–800 uk/donotdo/do-no t-use-risk-assessment -t ools-and-scale s-t o-deter 52. Pisani AR, Murrie DC, Silverman MM (2016) Reformulating mine-who-shoul d-and-shoul d-not-be-offer ed-treat ment-or-who- suicide risk formulation: from prediction to prevention. Acad shoul d-be-disch arged . Accessed 3 Sep 2017 Psychiatry 40:623–629 31. Kapur N, Murphy E, Cooper J et al (2008) Psychosocial assess- 53. Carroll R, Metcalfe C, Steeg S et al (2016) Psychosocial assess- ment following self-harm: results from the multi-centre monitor- ment of self-harm patients and risk of repeat presentation: an ing of self-harm project. J Affect Disord 106:285–293 instrumental variable analysis using time of hospital presentation. 32. Department of Health (DoH) (2008) Mental Health Act 2007: PLoS One 11(2):e0149713 patients on aftercare under supervision (ACUS): transitional 54. Fernandes AC, Cloete D, Broadbent MT et al (2013) Development arrangements. Department of Health, London and evaluation of a de-identification procedure for a case register 33. Department of Health (DoH) (2008) Code of practice: Mental coursed from mental health electronic records. BMC Med Inform Health Act 1983. TSO (The Stationer Office) Decis Mak 13:71 34. Hawton K, van Heeringen K (2009) Suicide. Lancet 55. Velupillai S, Mowery D, South BR, Kvist M, Dalianis H (2015) 373:1372–1381 Recent advances in clinical natural language processing in support 35. Coryell W, Young EA (2005) Clinical predictors of suicide in of semantic analysis. Yearb Med Inform 10:183–193 primary major depressive disorder. J Clin Psychiatry 66:412–417 56. Cook BJ, Progovac AM, Chen P, Mullin B, Hou S, Baca-Garcia E 36. Bebbington PE, Minot S, Cooper C et al (2010) Suicidal ideation, (2016) Novel use of natural language processing (NLP) to predict self-harm and attempted suicide: results from the British psychi- suicidal ideation and psychiatric symptoms in a text-based mental atric morbidity survey 2000. Eur Psychiatry 25:427–431 health intervention in Madrid. Comput Math Methods Med. https 37. Minkoff K, Bergman E, Beck AT, Beck R (1973) Hopelessness, ://doi.org/10.1155/2016/87084 34 depression, and attempted suicide. Am J Psychiatry 30:455–459 57. Berrouiguet S, Baca-Garcia E, Brandt S, Walter M, Courtet P 38. Beck AT, Brown G, Berchick RJ, Stewart BL, Steer RA (1990) (2016) Fundamentals for future mobile-health (mHealth): a sys- Relationship between hopelessness and ultimate suicide: a replica- tematic review of mobile phone and web-based text messaging in tion with psychiatric outpatients. Am J Psychiatry 147(2):190–195 mental health. J Med Internet Res 18:e135 39. Baca-Garcia E, Diaz-Sastre C, Garcia-Resa E et al (2005) Sui- 58. Berrouiguet S, Gravey M, Le Galudec M, Alavi Z, Walter M cide attempts and impulsivity. Eur Arch Psychiatry Clin Neurosci (2014) Post-acute crisis text messaging outreach for suicide pre- 255:152–156 vention: a pilot study. Psychiatry Res 217:154–157 40. Guldin MB, Ina Siegismund Kjaersgaard M, Fenger-Grøn M et al 59. Barrett JA, Shetty H, Broadbent M, Cross S, Hotopf M, Stewart (2017) Risk of suicide, deliberate self-harm and psychiatric illness R, Lee W (2016) ‘He left me a message on Facebook’: comparing after the loss of a close relative: a nationwide cohort study. World the risk profiles of self-harming patients who leave paper suicide Psychiatry 16:193–199 notes with those who leave messages on new media. Br J Psychia- 41. Shneidman ES (1993) Suicide as psychache. J Nerv Ment Dis try Open 2:217–220 181:145–147 60. Franklin JC, Ribeiro JD, Fox KR et al (2017) Risk factors for 42. Mann JJ, Waternaux C, Haas GL, Malone KM (1999) Toward a suicidal thoughts and behaviors: a meta-analysis of 50 years of clinical model of suicidal behavior in psychiatric patients. Am J research. Psychol Bull 143:187–232 Psychiatry 156:181–189 61. Free C, Phillips G, Felix L, Galli L, Patel V, Edwards P (2010) 43. Read J, Agar K, Barker-Collo S (2001) Assessing suicidality in The effectiveness of m-health technologies for improving health adults: integrating childhood trauma as a major risk factor. Prof and health services: a systematic review protocol. BMC Res Notes Psychol Res Pract 32:367–372 3:250 44. de Leon J, Baca-García E, Blasco-Fontecilla H (2015) From the 62. Zalsman G, Hawton K, Wasserman D et al (2016) Suicide pre- serotonin model of suicide to a mental pain model of suicide. vention strategies revisited: 10-year systematic review. Lancet Psychother Psychosom 84:323–329 Psychiatry 3:646–659 45. Mann JJ (2003) Neurobiology of suicidal behaviour. Nat Rev Neu- 63. Mann JJ, Apter A, Bertolote J et al (2005) Suicide prevention rosci 4:819–828 strategies: a systematic review. JAMA 294:2064–2074 46. Blasco-Fontecilla H, Baca-Garcia E, Courtet P, Garcia Nieto R, de 64. Riblet NBV, Shiner B, Young-Xu Y, Watts BV (2017) Strategies to Leon J (2015) Horror vacui: emptiness might distinguish between prevent death by suicide: meta-analysis of randomised controlled major suicide repeaters and nonmajor suicide repeaters: a pilot trials. Br J Psychiatry 210:396–402 study. Psychother Psychosom 84:117–119 65. World Health Organization (WHO) (2016) First WHO world sui- 47. Pompili M, Shrivastava A, Serafini G et  al (2013) Bereave- cide report. World Health Organization, Geneva ment after the suicide of a significant other. Indian J Psychiatry 55:256–263 48. Verrocchio MC, Carrozzino D, Marchetti D et al (2016) Mental pain and suicide: a systematic review of the literature. Front Psy- chiatry 7:108 1 3 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Social Psychiatry and Psychiatric Epidemiology Springer Journals

Can risk assessment predict suicide in secondary mental healthcare? Findings from the South London and Maudsley NHS Foundation Trust Biomedical Research Centre (SLaM BRC) Case Register

Free
11 pages

Loading next page...
 
/lp/springer_journal/can-risk-assessment-predict-suicide-in-secondary-mental-healthcare-m7WgnAseLA
Publisher
Springer Journals
Copyright
Copyright © 2018 by The Author(s)
Subject
Medicine & Public Health; Psychiatry
ISSN
0933-7954
eISSN
1433-9285
D.O.I.
10.1007/s00127-018-1536-8
Publisher site
See Article on Publisher Site

Abstract

Purpose The predictive value of suicide risk assessment in secondary mental healthcare remains unclear. This study aimed to investigate the extent to which clinical risk assessment ratings can predict suicide among people receiving secondary mental healthcare. Methods Retrospective inception cohort study (n = 13,758) from the South London and Maudsley NHS Foundation Trust (SLaM) (London, UK) linked with national mortality data (n = 81 suicides). Cox regression models assessed survival from the last suicide risk assessment and ROC curves evaluated the performance of risk assessment total scores. Results Hopelessness (RR = 2.24, 95% CI 1.05–4.80, p = 0.037) and having a significant loss (RR = 1.91, 95% CI 1.03–3.55, p = 0.041) were significantly associated with suicide in the multivariable Cox regression models. However, screening statis - tics for the best cut-off point (4–5) of the risk assessment total score were: sensitivity 0.65 (95% CI 0.54–0.76), specificity 0.62 (95% CI 0.62–0.63), positive predictive value 0.01 (95% CI 0.01–0.01) and negative predictive value 0.99 (95% CI 0.99–1.00). Conclusions Although suicide was linked with hopelessness and having a significant loss, risk assessment performed poorly to predict such an uncommon outcome in a large case register of patients receiving secondary mental healthcare. Keywords Suicide · Risk assessment · Secondary mental healthcare · Mental pain Introduction Every year, almost one million people die from suicide across the world [1], which appears to have increased since Javier-David Lopez-Morinigo and Andrea C. Fernandes the start of the 2007 economic recession [2]. Indeed, sui- contributed equally to this work and should be jointly cide represents one of the three leading causes of death in acknowledged as first named authors. the most economically productive age group (15–44 years) Electronic supplementary material The online version of this article (https ://doi.org/10.1007/s0012 7-018-1536-8) contains supplementary material, which is available to authorized users. * Javier-David Lopez-Morinigo South London and Maudsley NHS Foundation Trust, javier.lopez-morinigo@kcl.ac.uk London, UK Department of Psychiatry, Marqués de Valdecilla University Department of Psychosis Studies, Institute of Psychiatry, Hospital, IFIMAV, School of Medicine, University Psychology and Neuroscience, King’s College London, De of Cantabria, Santander, Spain Crespigny Park, PO Box 68, London SE5 8AF, UK Centro Investigación Biomédica en Red de Salud Mental CAS Behavioural Health, London, UK (CIBERSAM), Madrid, Spain Department of Psychological Medicine, Institute of Psychiatry, Psychology and Neuroscience, King’s College London, London, UK Vol.:(0123456789) 1 3 1162 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 [3]. Of concern, suicide rates in the UK have shown no Methods reduction over the past 5 years [4]. Up to 90% of people who complete suicide are found Participants to have had a ‘psychiatric disorder’ [5], contributing to 47–74% of the population risk of suicide, with half of peo- As stated, the sample was derived from the SLaM BRC ple completing suicide meeting retrospectively applied cri- Case Register. SLaM is one of Europe’s largest mental teria for depression [6, 7]. Although it could be envisaged health services, providing secondary mental healthcare that secondary mental health services may play a crucial to four boroughs in South-East London (UK): Lambeth, role in ‘suicide prevention’ [8], over two-thirds of those Southwark, Lewisham and Croydon. Approximately, who take their lives in the UK have not received secondary 1.23 million inhabitants reside in this geographic catch- mental healthcare in the year before death [4]. ment area, which as a whole was found to be comparable Risk assessment in mental health services might con- with other populations of London in terms of age, gen- ceivably help reduce suicide rates, and the UK Depart- der, education and socio-economic status distributions ment of Health [9] and 2004 NICE guidelines [10] recom- [20, 21]. Fully electronic health records have been in use mended the use of structured clinical risk assessments. across all SLaM services since 2006, and in 2007–2008, However, completed suicide thankfully remains a very the Clinical Record Interactive Search (CRIS) system was uncommon event, and two early studies warned of the high built which renders de-identified copies of records availa- number of false positives picked up to detect the major- ble for research use with appropriate governance structures ity of suicides using this approach [11, 12]. In particular, [20]. CRIS received ethical approval as an anonymised even a hypothetical test with a sensitivity and specificity of data resource for secondary analyses from the Oxford C 99% in a high-risk population (defined as a suicide rate of Research Ethics Committee (reference: 08/H0606/71+5), 250/100,000/years) cannot predict suicide beyond a 20% and currently accesses data on over 300,000 patients [21]. level of efficiency [ 11]. Consistent with this, recent meta- The same research ethics approval also covers the pseu- analyses have concluded that risk scales have a limited donymised linkage between CRIS data and those from the role in predicting suicidal behaviour [13–15], although Office for National Statistics (ONS) in April 2015 [22], there are important issues of between-study heteroge- which registers all deaths in the UK and the official cause neity [13]. In keeping with these meta-analyses, a 2017 of death, including suicide and the method of suicide multicentre study in the UK [16] replicated the limited according to ICD-10 classification [23]. use of risk scales to predict repeated self-harm, which is Those patients who had received SLaM care (i.e., had in line with our previous report on risk assessment and at least one face-to-face contact with a clinical member of suicide by patients with schizophrenia spectrum disorders staff) over the period from 1st January 2007 to 1st April under secondary mental healthcare [17]. However, using 2015 and had at least one suicide risk assessment docu- risk assessment scales continues to be common clinical mented within the study period were included. Those who practice [18]. died from suicide within the study period were compared In addition, the extent to which risk assessment can with those who did not. The analysis described here was predict suicide mortality (rather than ‘self-harm’, ‘sui- based on a surveillance period from 1 January 2007 to 1 cide attempts’ or ‘suicidal behaviour’) in a large sample April 2015, the rationale being that the electronic clinical of mental health service users irrespective of diagnosis, records coverage became complete across all SLaM ser- which also changes over time [19], has not been examined vices during 2006 and, at the time of the analysis, the last to date. Within this context, we investigated the perfor- death certification linkage had been accomplished in the mance of all full suicide risk assessments from the South beginning of April 2015. London and Maudsley (SLaM) Biomedical Research Centre (BRC) Case Register (in South-East London, UK) over 2007–2015 to predict suicide. Whilst anticipating that Measures some risk factors would be statistically associated with a higher risk of suicide, namely previous suicide attempts, Risk assessment suicidal ideation, hopelessness, alcohol/drugs and impul- sivity, we sought to clarify positive predictive values at ‘Full risk assessment’ is a compulsory target across the different levels of raised risk, as well as the extent to which Trust when ‘high risk’ is determined from a ‘brief risk a risk assessment might allow clinicians to rule out risk: assessment’, which is mandatory for all active cases. All i.e., the extent to which ‘low-risk’ patients would not end patients who have been seen by a member of clinical staff their lives. have a ‘brief risk assessment’ documented, which is a 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1163 narrative record of the patient’s risk: (1) to one’s self; (2) in front of a vehicle)—X80, X81; suicide by unspecified to others and (3) from others. If the patient is deemed at means—X84; and undetermined cause of death—Y10-34. ‘high’ risk in any of these domains, a ‘full risk assess- Those with an ‘undetermined cause of death’ code were con- ment’ needs to be completed and updated over time, which sidered as suicides, because in the UK most ‘open verdicts’ consists of a structured assessment taking the form of pre- have been reported as very likely to be suicides [26]. sent/absent tick-boxes enquiring about widely recognised risk factors for three major clusters: suicide, violence and Statistical analysis self-neglect. Full risk assessment is entered into CRIS as structured information separately from clinical free- First, for descriptive purposes, for all SLaM ‘active’ ser- text entries. Hence, information on those who had a ‘full vice users (i.e., at least one face-to-face contact with a staff risk assessment’ documented (compared to those without member) over the study period (2007–2015), we investigated ‘full risk assessment’, including those with ‘partial’ risk suicide rates differences between those with/without full risk assessment, i.e., only some items out of the 15-item full assessment. suicide risk assessment) can be reliably extracted from Second, in those with at least one full risk assessment CRIS. For the purposes of this study, only those patients documented (i.e., the study sample), risk assessment individ- with a ‘full suicide risk assessment’ were included, that ual items and total scores, as independent variables, entered is, those patients with ratings on the 15 items included into Kaplan–Meier survival analyses and Cox regression in the full suicide risk assessment, which is available in models [27], respectively, to investigate associations with Appendix  1 (supplementary material available online). time to suicide. Proportional hazards assumptions were Positive responses can be summed to create total scores, checked as standard for Cox procedures and no evidence of i.e., the higher the score the greater the suicide risk, which violation was found, i.e., the survival curves for two strata yielded good internal consistency (Cronbach α coefficients (determined by the values for the covariates) had ‘hazard’ of 0.69) [24]. The most recent full suicide risk assessment functions, which were ‘proportional’ (or constant) over time. was considered for this study. Age, gender, religion, employment and marital status, ethnic group, IMD and primary psychiatric diagnosis were entered Demographic and clinical covariates as covariates. For the survival analyses, the follow-up period began at the time of the last risk assessment and the end Demographic and clinical covariates included age at the date was the date of death (including suicide) or the censor- time of risk assessment, gender, ethnicity, religion, marital ing point (last face-to-face contact, date of death from non- status, employment status, social deprivation and ICD-10 suicide causes or 1st April 2015, whichever came sooner). diagnosis [23]. In addition, receiver operating characteristic (ROC) Social deprivation was scored through an anonymous link curves [28], which compare the true positive rate (i.e., sen- created in CRIS between lower super output area residence sitivity) with the false positive rate (i.e., ‘1-specificity’) at code of the latest permanent address (a geographic unit different cut-off points for the parameter (risk assessment comprising approximately 400 households) and summary total score in our study), were plotted to analyse the per- data for that area from 2007 UK Census output. Thus, the formance of risk assessment total scores to predict suicide. Index of Multiple Deprivation (IMD) is derived from seven In particular, sensitivity, specificity, positive and negative domains: income, employment, health, education, housing predictive values, the area under the curve (AUC), likeli- and services, crime and environment [25]. hood ratios (positive and negative) and diagnostic odds ratio ICD-10 diagnoses [23] were reached by consensus by (OR) were investigated for the best cut-off point, including the treating multidisciplinary team, including input from a 95% confidence intervals (CI) for each statistic at each risk senior consultant psychiatrist. Specifically, several clinically assessment total score. Positive and negative predictive val- meaningful categories were created as follows: ‘organic ues are the probability that subjects with a positive (high mental disorders’ (F0), ‘substance abuse’ (F1), ‘schizophre- risk) result will truly have the outcome of interest (in this nia spectrum disorders’ (F2), ‘mood disorders’ (F3), ‘neu- study, death from suicide), and the probability that subjects rotic disorders’ (F4) and ‘all others’ (F5–F7). with a negative (low risk) result will not have such an out- come, respectively. The AUC is a measure of how well risk Suicide method assessment total score can distinguish those who will die from suicide from those who will not. Likelihood ratios are Suicide method was ascertained using death certificate [ 22] the likelihood that a given test result would be expected in ICD-10 codes [23] and the following groups were used to a patient who took his/her life compared to the likelihood define this: poisoning—X64; hanging—X70; drowning— that same result would be expected in a patient who did not X71; cutting—X78; jumping (either from high place or end his/her life. Diagnostic ORs are the ratio of the odds of 1 3 1164 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 the test being positive (high risk) if the subject ended his/ died by suicide. Taking into account the observation time of her life relative to the odds of the test being positive if the each patient, the analysed sample contributed to 80,769.17 subject did not die from suicide. person-years, which yielded a suicide rate of 100.28/100,000 A significant level of 5% (two-tailed) was set for all the person-years. In those with a partial risk assessment (n = 1409), above analyses, which were performed using the statistical the suicide rate was 51.60/100,000 person-years, while for those package R (version 3.20) [29]. with no risk assessment completed (n = 80,340), the suicide rate was 88.34/100,000 person-years. These differences were not statistically significant (X = 6, df = 4, p = 0.19). Results The baseline demographic and clinical characteristics of the sample (n = 13,758) and differences between those who Study sample took their lives and those who did not are presented in Table 1. Although there was a higher male predominance in the suicide Over 2007–2015, there were 99,507 SLaM ‘active’ cases, i.e., completers group than in those who did not end their lives those who had at least one face-to-face contact with a SLaM (OR = 1.67, 95% CI 1.04–2.69, p = 0.03), no further signifi- staff member over that period (2007–2015), of whom 358 were cant differences emerged in age at first presentation, religion, ascertained as having died by suicide. Of all these active SLaM marital status, ethnicity, living status, employment, social dep- service users, 13,758 subjects had all suicide-related items com- rivation, first language (English vs. all others) and ICD-10 pleted on a full risk assessment, and a further 1409 had incom- diagnoses. Hanging was the most common suicide method plete data (with at least one item rated). Of the 13,758 individu- (n = 28). Twenty-one subjects received an undetermined cause als, who formed the study sample, 81 were recorded as having of death. There were no suicides by firearms. Table 1 Demographics and Suicides Non-suicides p value clinical characteristics of the N = 81 N = 13,678 sample Mean ± SD Mean ± SD Age at risk assessment (years) 41.3 ± 12.2 40.6 ± 11.5 0.60 Social deprivation 28.6 ± 13.4 28.6 ± 12.3 0.98 n (%) n (%) OR (95% CI) Gender (males) 56 (69.1) 7823 (57.2) 1.67 (1.04–2.69) 0.03 Marital status (unmarried) 72 (88.9) 11,909 (87.1) 1.10 (0.55–2.21) 0.78 Employment status (unemployed) 30 (37.0) 4662 (34.0) 1.57 (0.48–5.14) 0.46 Living status (alone) 19 (23.5) 3093 (22.6) 1.04 (0.57–1.89) 0.90 Religion (yes) 20 (24.7) 3177 (23.2) 1.00 (0.60–1.67) 0.98 Ethnicity  White 50 (61.7) 6916 (50.6) 1.85 (1.02–3.35) 0.04  Black 14 (17.3) 3525 (25.8) Ref.  Others 17 (21.0) 3237 (23.7) 1.37 (0.67–2.78) 0.38 First language (non-English) 40 (49.4) 5971 (43.7) 1.53 (0.99–2.37) 0.06 Diagnosis  Organic disorders 1 (23.0) 270 (2.0) Ref.  Substance use disorders 10 (12.3) 1232 (9.00) 1.95 (0.25–15.2) 0.53  Schizophrenia spectrum 31 (38.3) 5713 (41.8) 1.28 (0.17–9.34) 0.81  Mood disorders 24 (26.6) 2643 (19.3) 2.28 (0.30–16.9) 0.42  Neurotic disorders 5 (6.17) 860 (6.30) 1.60 (0.19–13.7) 0.67  Other diagnoses 10 (12.3) 2960 (21.6) 0.85 (0.10–6.60) 0.87 Method  Hanging 28 (34.6)  Intoxication 6 (7.4)  Jumping 3 (3.7)  Unspecified means 16 (19.7)  NA 7 (8.6)  Undetermined cause of death 21 (25.9) 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1165 Table 2 Unadjusted univariate Risk factor N Events expected Events Log-rank test RR (95% CI) p value analyses: risk assessment items observed Previous suicide attempts  Absent 7004 40 24 14.1 2.46 (1.51–4.01) < 0.001  Present 5657 33 49 Violent method  Absent 9167 50.7 37 14.7 2.46 (1.53–3.97) < 0.001  Present 3069 17.3 31 Plan to end life  Absent 11,551 64.5 53 22.5 3.37 (1.98–5.77) < 0.001  Present 1258 6.5 18 Suicidal ideation  Absent 10,583 61.7 51 9.4 2.06 (1.30–3.31) 0.002  Present 2843 15.3 26 Hopelessness  Absent 9419 55.1 37 21.7 2.79 (1.78–4.37) < 0.001  Present 3811 20.9 39 Distress  Absent 8617 51.5 42 5.1 1.66 (1.07–2.60) 0.024  Present 4651 26.5 36 No control over life  Absent 9290 55.1 42 11.3 2.13 (1.36–3.35) < 0.001  Present 3576 20.9 34 Alcohol/drugs  Absent 7834 44.1 36 3.8 1.60 (0.99–2.51) 0.051  Present 4971 27.9 36 Impulsivity  Absent 7292 43.5 34 4.7 1.64 (1.05–2.57) 0.030  Present 5636 33.5 43 Living alone  Absent 8358 48.3 46 0.3 1.12 (0.72–1.76) 0.603  Present 4949 30.7 33 Poor physical health  Absent 9616 56.2 56 0 1.01 (0.61–1.67) 0.964  Present 3431 20.8 21 Significant loss  Absent 8243 50 39 7.5 1.88 (1.20–2.96) 0.006  Present 4256 24 35 Disengagement  Absent 9841 58.3 48 7.2 1.85 (1.17–2.91) 0.007  Present 3277 19.7 30 Recent hospital discharge  Absent 10,973 65.1 56 7.3 1.93 (1.20–3.14) 0.007  Present 2399 13.9 23 Family history  Absent 5921 32.88 35 1.2 0.46 (0.11–1.90) 0.268  Present 730 4.12 2 1 3 1166 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 Risk assessment factors Differences in individual risk assessment items between suicide completers and non-completers (Kaplan–Meier survival analyses) are presented in Table 2. The following items were significantly associated with an increased risk of suicide: previous suicide attempts (RR = 2.46, 95% CI 1.51–4.01, p < 0.001), previous use of a violent method (RR = 2.46, 95% CI 1.53–3.97, p < 0.001), plans to end life (RR = 3.37, 95% CI 1.98–5.77, p < 0.001), suicidal idea- tion (RR = 2.06, 95% CI 1.30–3.31, p = 0.002), hopeless- ness (RR = 2.79, 95% CI 1.78–4.37, p < 0.001), distress (RR = 1.66, 95% CI 1.07–2.60, p = 0.024), lack of con- trol over life (RR = 2.13, 95% CI 1.36–3.35, p < 0.001), impulsivity (RR = 1.64 95% CI 1.05–2.57, p = 0.030), having a significant loss (RR = 1.88, 95% CI 1.20–2.96, p = 0.006), disengagement (RR = 1.85, 95% CI 1.17–2.91, p = 0.007) and recent hospital discharge (RR = 1.93, 95% CI 1.20–3.14, p = 0.007). Cox regression analyses of the relationship between those risk assessment factors significantly associated Fig. 1 ROC curve for risk assessment total scores with risk of suicide (see Table 2) are presented in Table 3. The left column shows the results only after adjusting strengthened coefficient but wider confidence intervals), the analyses for gender, which had been the only signifi- were substantially attenuated. cant baseline characteristic associated with suicide risk (see Table  1). All the risk factors remained significant Risk assessment overall performance (Table 3, left column). The right column of Table 3 pre- sents coefficients from a model containing all factors; in The diagnostic accuracy statistics (and 95% CI) for each risk this, only hopelessness (RR = 2.24, 95% CI 1.05–4.80, assessment total score are detailed in Table 4. ROC curve p = 0.037) and having a significant loss (RR = 1.91, 95% analyses for risk assessment total scores, which are shown CI 1.03–3.55, p = 0.041) remained statistically significant in Fig. 1, found the optimal cut-off point to be 4–5 (above predictors of suicide. All the other associations, apart which the risk would be ‘high’; below which the risk would from that with previous suicide attempts (which showed a Table 3 Adjusted Cox a b Individual items RR (95% CI) p value RR (95% CI) p value regression analyses: risk assessment items Gender (male) 1.67 (1.04–2.67) 0.03 1.70 (0.95–3.02) 0.077 Suicidal history 1.67 (1.04–2.67) < 0.001 2.00 (0.89–4.53) 0.094 Violent method 2.58 (1.58–4.21) < 0.001 1.31 (0.65–2.65) 0.453 Plan to end life 2.47 (1.53–3.98) < 0.001 1.20 (0.54–2.63) 0.657 Suicidal ideation 2.13 (1.32–3.41) 0.002 0.82 (0.40–1.67) 0.577 Hopelessness 2.90 (0.85–4.55) < 0.001 2.24 (1.05–4.80) 0.037 Distress 1.72 (1.10–2.70) 0.020 0.90 (0.48–1.68) 0.746 No control over life 2.21 (1.40–3.48) < 0.001 0.95 (0.47–1.91) 0.881 Impulsivity 1.65 (1.05–2.58) 0.029 1.05 (0.56–1.97) 0.875 Significant loss 1.95 (1.23–3.07) 0.004 1.91 (1.03–3.55) 0.041 Disengagement 1.85 (1.17–2.92) 0.008 1.16 (0.61–2.20) 0.653 Recent hospital discharge 1.97 (1.21–3.21) 0.006 1.45 (0.77–2.74) 0.247 Bold values indicate statistically significance p < 0.05 Analysis adjusted for gender only Fully adjusted analysis 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1167 Table 4 Risk assessment total scores (number of risk factors present) and diagnostic accuracy statistics with 95% confidence intervals Risk assess- Suicides Non-suicides Sensitivity, Specificity, Positive pre- Negative Likelihood Likelihood Diagnostic ment (Total % (95% CI) % (95% CI) dictive value, predictive ratio, + (95% ratio, − (95% OR (95% CI) score) % (95% CI) value, % CI) CI) (95% CI) 0 2 1113 98 (91–100) 8 (8–9) 1 (0–1) 100 1.1 (1.0–1.1) 0.3 (0.1–1.2) 3.5 (0.9–14.3) (99–100) 1 3 1748 94 (86–98) 21 (20–22) 1 (1–1) 100 (100– 1.2 (1.1–1.3) 0.3 (0.1–0.7) 4.0 (1.6–9.9) 100) 2 7 2042 85 (76–92) 36 (35–37) 1 (1–1) 100 (100– 1.3 (1.2–1.5) 0.4 (0.2–0.7) 3.2 (1.7–5.9) 100) 3 8 1909 75 (64–84) 50 (49–51) 1 (1–1) 100 (100– 1.5 (1.3–1.7) 0.5 (0.3–0.7) 3.02 (1.8–5.0) 100) 4 8 1727 65 (54–76) 62 (62–63) 1 (1–1) 100 1.7 (1.5–2.0) 0.5 (0.4–0.7) 3.1 (2.0–5.0) (99–100) 5 12 1443 51 (39–62) 73 (72–74) 1 (1–1) 100 1.9 (1.5–2.3) 0.7 (0.5–0.8) 2.8 (1.8–4.3) (99–100) 6 10 1161 38 (28–50) 81 (81–82) 1 (1–2) 100 2.1 (1.6–2.7) 0.8 (0.6–0.9) 2.7 (1.73–4.3) (99–100) 7 10 836 26 (17–37) 88 (87–88) 1 (1–2) 100 2.1 (1.4–3.0) 0.8 (0.7–1.0) 2.5 (1.5–4.1) (99–100) 8 9 712 15 (8–24) 93 (92–93) 1 (1–2) 100 2.0 (1.2–3.5) 0.9 (0.8–1.0) 2.23 (1.2–4.1) (99–100) 9 4 427 10 (4–19) 96 (96–96) 1 (1–3) 100 2.4 (1.2–4.7) 0.9 (0.8–1.0) 2.6 (1.2–5.3) (99–100) 10 4 291 5 (1–12) 98 (98–98) 1 (1–4) 100 2.5 (0.9–6.6) 1.0 (0.9–1.0) 2.6 (0.9–7.1) (99–100) 11 3 167 1 (0–7) 99 (99–99) 1 (0–5) 100 1.7 (0.2– 1.0 (1.0–1.0) 1.6 (0.2–12.1) (99–100) 11.7) 12 1 64 0 (0–4) 1 (1–1) 0 (0–9) 99 (99–100) 0 (0–0) 1.0 (1.0–1.0) NA 13 0 28 NA NA NA NA NA NA NA 14 0 10 NA NA NA NA NA NA NA be ‘low’), with a sensitivity of 0.65 (95% CI 0.65–0.76), we found that although a number of risk factors were sig- specificity of 0.62 (95% CI 0.62–0.63) and an AUC of 0.67 nificantly associated with suicide in the bivariate analyses (95% CI 0.62–0.73). The positive and negative predictive (namely, being male, previous suicide attempts, previous use values were 0.01 (95% CI 0.01–0.01) and 0.99 (95% CI of violent methods, plans to end life, suicidal ideation, dis- 0.99–1.00), respectively. tress, lack of control over life, impulsivity, disengagement from services/non-compliance with medication and recent hospital discharge), only hopelessness and having a signifi- Discussion cant loss remained independent and statistically significant predictors of suicide in the multivariable regression models. Main findings Third, overall risk assessment performed poorly to predict suicide in a large sample of mental health service users, We drew data from a large case register of patients receiving which was in line with our expectations and recent literature secondary mental healthcare in a defined catchment area [13–15]. over a prolonged period (2007–2015) linked with national mortality data and we tested the extent to which structured Comparison with previous literature risk assessment items (individual risk factors and overall scores) predicted suicide. First, as expected, we identified a Of relevance, we did not find a protective effect of com- high number of suicides in a population of patients in sec- pleting suicide risk assessment on reducing suicide rates as ondary mental healthcare (approximately 100.28/100,000 previously suggested [10], which was in line with further person-year), approximately tenfold higher than in the gen- reviews of the NICE guidelines [30]. It might seem intui- eral population (10.9/100,000 person-year) [22]. Second, tive to speculate that those patients with a risk assessment 1 3 1168 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 documented are deemed at a higher risk of suicide by the [45] based on Mann’s model of suicide [42]. Indeed, over treating team and they are more likely to receive a higher 90% of suicide attempters report ‘mental pain’ [46], which is level of input and caution in management. It remains unan- frequently the consequence of a bereavement [40], which is swerable what may have happened if these ‘high risk’ indi- of particular concern after surviving the suicide of a love one viduals had not been administered a risk assessment and to [47]. Moreover, ‘mental pain’ appears to be a contributor to research this would raise ethical issues. However, recording suicide independently of depression [48], which is in full of risk assessment is, to some degree, circular and more agreement with our results. Indeed, the relationship between likely in those patients who will be followed-up by men- suicide and ‘mental illness’ (from a psychiatric perspective) tal health services [31]. In addition, risk assessment in this may be weaker than previously thought, especially in West- cohort may have been completed due to concerns raised ern countries [44]. In keeping with this, neither medication regarding other clusters of risk, such as violence to others compliance nor (psychiatric) diagnosis were associated and/or self-neglect [24]. Moreover, risk assessment comple- with suicide in our large cohort of mental health service tion rates may have been affected by the patient’s legal status users, which may provide further support for a non-medical in some cases, which was not evaluated in this study. For approach to suicide [49]. Hence, those patients receiving instance, those receiving care under restriction, hence sub- secondary mental healthcare at risk of suicide may particu- ject to the UK Mental Health Act 1983 (amended 2007) [32] larly benefit from psychological therapies, as recommended may be more likely to have a risk assessment documented. by the UK NICE guidelines for depression [50], although not In this regard, our findings are of major relevance from a frequently offered [51]. human rights perspective, since these individuals may have In addition, overall risk assessment showed poor predic- been ‘forced’ to undertake an assessment which appears to tive validity, which was unsurprising, given the rarity of the have a limited value, which would also go against the 2015 outcome. In particular, high sensitivity was reached at the UK Code of Practice [33]. price of low specificity (i.e., a very high number of false In terms of risk factors, we replicated the association of positives to identify most of suicides) and vice versa, i.e., being male [1, 34], previous suicide attempts [35], previous reducing the number of false positives (high specificity) use of violent methods [5], plans to end life and suicidal occurred at the expenses of too many false negatives (low ideation [36], hopelessness [37, 38], distress, lack of control sensitivity), thus preventing high-risk patients from care and over life and impulsivity [39], having a significant loss [40], treating ‘healthy’ people unnecessarily, which was in full disengagement from services/non-compliance with medica- agreement with early literature [11, 12]. tion and recent hospital discharge [4] with suicide. Consist- In keeping with this, for the most optimal cut-off point ent with previous models of suicide [41, 42], only hope- (4–5), a very low positive predictive value (1%) and very lessness and having a significant loss remained significant. high negative predictive value (99%) emerged from the anal- However, it should be noted that childhood trauma, which yses. In other words, in this ‘low-risk’ group (those with was not evaluated by our risk assessment questionnaire, was less than four risk factors), there would be still 20 suicides found to have greater effects on suicidality than depression (approximately one quarter of those who ended their lives). and related variables [43], hence it should become part of On the other hand, 6988 individuals (50.8% of the total sam- routine clinical suicide risk assessment. ple) would be classified as ‘high-risk’ patients, although only Over four decades ago, hopelessness was defined as ‘the 61 of these subjects took their lives. The question arises. Is cognitive element of negative expectations’ [37], which was it, therefore, worth managing so many patients as ‘suicidal’ also demonstrated to be the strongest predictor of suicide to prevent a few deaths? More specifically, in times of finan- in outpatients [38], and this we replicated in our results. cial constraints, should so many patients receive high levels Hence, our findings agree, in part, with Mann’s diathesis- of care such as unnecessary admissions to hospital? stress model of suicide [42] regarding the role of hopeless- These findings were consistent with a 2017 literature ness, although impulsivity [39] was not significantly associ- review on ‘the limitations of epistemic uncertainty’ with ated with suicide in our cohort. It should be noted, however, regard to risk assessment, whose overall poor performance that impulsivity, as measured by the SLaM risk assessment, appears to be due to the so-called ‘aleatory uncertainty’. In might refer to a different construct. On the other hand, we short, risk factors change ‘by chance’, which is unpredictable did find that having a significant loss was a predictor of [18]. The concept of risk, therefore, requires a reformulation. suicide independently of other factors, which, in addition Specifically, while suicide risk does not appear to be predict- to the role of hopelessness, was consistent with the classic able, a more prevention-orientated approach may result in theory of ‘suicide as psychache’ [41]. Of note, the concep- better clinical outcomes [52]. tualization of suicide as the consequence of ‘mental pain or However, our findings concerning the association of psychache’ has been recently reconsidered [44] in light of hopelessness and having a significant loss with suicide, decades of relatively unsuccessful neurobiological research which as a whole provide some support for the ‘mental pain’ 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1169 model of suicide [41], which was discussed above, still sug- receiving secondary mental healthcare within a defined geo- gest that suicide may still be, to some degree, predictable. graphic catchment and time period. Since most people living in South-East London requiring secondary mental healthcare Future research receive this from NHS resources, our sample is likely to be representative. In addition, participants were followed-up While we do recommend that risk assessment should for up to 9 years and, in addition to risk assessment ratings, remain part of our routine clinical practice, a more nar- a number of covariates were taken into account. rative (free-text) approach should be taken [53] to better However, the study has some limitations to be borne in capture aspects such as ‘mental pain’, which, based on our mind when drawing conclusions from the results. First, all findings, seems to be more useful in terms of clinical risk participants were mental health service users residing in assessment. In particular, rather than categorising patients as South-East London, which is an inner urban area. Hence, ‘low–medium–high’ risk, the wide range of contributing fac- our findings may not generalise to people with mental health tors to risk should be detailed in relation to the individual’s problems under primary care or those who live in rural mental health problems and the social context and how these areas. Second, the vast majority of SLaM patients (almost factors may change dynamically over time, thus increasing or 90%) did not have a structured risk assessment completed decreasing risk for a given individual, which is what matters and those who did may have been deemed ‘at-high-risk’ by clinically [52]. Patient information from electronic records their treating teams. In other words, it could be still specu- can be easily, safely and securely de-identified to generate lated that assessing risk in all patients receiving care may large datasets for secondary research [54], such as the SLaM reduce risk. Third, although only the last suicide risk assess- CRIS [20, 21]. Specifically, naturalistic language processing ment was considered, risk factors may have changed from (NLP) tools appear to be promising research instruments the time of risk assessment to death. Also, risk assessment to extract statistically analysable clinical information from scores may have been affected by survival, hence those who narrative electronic records, hence determining risk from survived for longer (and therefore received care for a more free-text notes [55]. NLP techniques may assess suicide risk prolonged period of time) may have been rated differently. using information from unstructured questions with higher Finally, other factors such as patient’s legal status at the time precision than the classic risk assessment scales [56], thus of risk assessment or a history of childhood trauma, which potentially capturing the presence of ‘mental pain’. were not evaluated in this study, may have affected both Specifically, future studies should examine whether risk risk assessment completion rates and ratings, and the main assessment changes over time, particularly self-ratings outcome measure of this study, namely suicide. shortly before suicide may increase the predictive value of risk assessment. In this regard, mobile phone and web-based Implications and conclusions text messaging may represent a useful tool to self-monitor suicide risk [57], particularly to follow-up people attempt- Our results, therefore, support the notion that neither indi- ing suicide [58] and to assess risk shortly before suicidal vidual risk factors nor a combination of them, i.e., risk events, including suicides. For instance, the classic suicide assessment, can adequately predict suicide in a popula- note may have been substituted by a message left on this tion of patients receiving mental healthcare. Suicide is a new media, which clinicians should discuss with patients very uncommon outcome even in a high-risk group such as and carers when assessing self-harm [59]. In addition, free- patients receiving secondary mental healthcare. Our study text-based risk assessment, which can be researched through suggests that risk assessment cannot predict suicide in the NLP techniques [55], may be more accurate than psychomet- clinical setting due to its very low occurrence, which is in ric scales [56]. Future research should, therefore, switch the full agreement with recent meta-analyses [13–15], although focus from long-term risk factors to short-term risk algo- hopelessness and having a significant loss were linked with rithms, which are more relevant to the clinician [60]. How- suicide, consistent with the classic ‘mental pain’ model of ever, the evidence of the use of communication technologies suicide [41, 44, 46, 48]. in health care and public health, which is known as mobile Meanwhile, further research on suicide prevention health (mHealth) [61], on suicide prevention is limited [62]. [62–64] is warranted. In particular, means restriction remains the first-line strategy to prevent both high-risk groups such as patients receiving mental healthcare [17] Strengths and limitations and the general population [65] from suicide. The use of a large case register linked with national mortal- Acknowledgements We are extremely grateful to all the patients who received care under the SLaM teams within the study period and the ity data allowed us to investigate the role of risk assessment SLaM staff who, therefore, took part in this research project. in predicting suicide in a large sample of patients who were 1 3 1170 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 Funding The development of the SLaM BRC Case Register was 9. Department of Health (DoH) and Social Security (1984) The man- funded by two Capital Awards from the UK National Institute for agement of deliberate self-harm. HMSO, London Health Research and is further supported through the BRC Nucleus 10. National Institute for Health and Clinical Excellence (NICE) funded jointly by the Guy’s and St Thomas’ Trustees and South Lon- (2004) Clinical Guidance 16 Self-harm. http s ://www .nice.org. don and Maudsley Special Trustees. JDLM was funded by the Brit-uk/Guida nce/CG16. Accessed 7 Sep 2017 ish Medical Association via the Margaret Temple Research Award 11. MacKinnon DR, Farberow NL (1976) An assessment of the utility for Schizophrenia. RD is funded by a Clinician Scientist Fellowship of suicide prediction. Suicide Life Threat Behav 6:86–91 awarded by the Academy of Medical Sciences in partnership with The 12. Pokorny AD (1983) Prediction of suicide in psychiatric patients. Health Foundation. JDLM, ACF, HS, RB and RD are part-funded Arch Gen Psychiatry 40:249–257 by the National Institute for Health Research (NIHR) Biomedical 13. Large M, Kaneson M, Myles N, Myles H, Gunaratne P, Ryan C Research Centre and Dementia Biomedical Research Unit at South (2016) Meta-analysis of longitudinal cohort studies of suicide risk London and Maudsley NHS Foundation Trust and King’s College assessment among psychiatric patients: heterogenity in results and London. JDLM and AB also acknowledge funding support from CAS lack of improvement over time. PLoS One 11:e0156322 Behavioural Health. The funders had no role in hypothesis generation, 14. Quinlivan L, Cooper J, Davies L, Hawton K, Gunnell D, Kapur N study design, data collection and analysis, decision to publish or the (2016) Which are the most useful scales for predicting repeat self- manuscript writing. harm? A systematic review evaluating risk scales using measures of diagnostic accuracy. BMJ Open 6:e009297 15. Carter G, Milner A, McGill K, Pirkis J, Kapur N, Spittal M (2017) Compliance with ethical standards Predicting suicidal behaviours using clinical instruments: system- atic review and meta-analysis of positive predictive values for risk Conflict of interest The authors disclosed no competing interests con- scales. Br J Psychiatry 210:387–395 cerning the subject of the study. 16. Quinlivan L, Cooper J, Meehan D et al (2017) Predictive accuracy of risk scales following self-harm: multicentre, prospective cohort Ethical statement This study was performed in accordance with the study. Br J Psychiatry 210:429–436 ethical standards laid down in the 1964 Declaration of Helsinki and its 17. Lopez-Morinigo JD, Ayesa-Arriola R, Torres-Romano B et al later amendments. Under UK law anonymised data can be used without (2016) Risk assessment and suicide by patients with schizophrenia prior consent and data for this study came from CRIS (see “Methods” in secondary mental healthcare: a case–control study. BMJ Open section for details), which received ethical approval as an anonymised 6:e011929 data resource for secondary analyses from the Oxford C Research Eth- 18. Large M, Galletly C, Myles N, Ryan CJ, Myles H (2017) Known ics Committee (reference: 08/H0606/71+5). unknowns and unknown unknowns in suicide risk assessment: evidence from meta-analyses of aleatory and epistemic uncer- tainty. Br J Psychiatry Bull 41:160–163 Open Access This article is distributed under the terms of the Crea- 19. Heslin M, Lomas B, Lappin JM et  al (2015) Diagnostic tive Commons Attribution 4.0 International License (http://creat iveco change 10 years after a first episode of psychosis. Psychol Med mmons.or g/licenses/b y/4.0/), which permits unrestricted use, distribu- 45:2757–2769 tion, and reproduction in any medium, provided you give appropriate 20. Stewart R, Soremekun M, Perera G et al (2009) The South London credit to the original author(s) and the source, provide a link to the and Maudsley NHS Foundation Trust Biomedical Research Centre Creative Commons license, and indicate if changes were made. (SLAM BRC) case register: development and descriptive data. BMC Psychiatry 9:51 21. Perera G, Broadbent M, Callard F et al (2016) Cohort profile of the South London and Maudsley NHS Foundation Trust Biomedi- cal Research Centre (SLaM BRC) Case Register: current status References and recent enhancement of an Electronic Mental Health Record- derived data resource. BMJ Open 6:e008721 1. World Health Organization (WHO) (2014) Preventing suicide: a 22. Office for National Statistics (ONS) (2015) Suicide statistics: global imperative. World Health Organization, Geneva deaths registered in England and Wales. http://www.ons.gov. 2. Oyesanya M, Lopez-Morinigo J, Dutta R (2015) Systematic uk/ons/dat as e ts-andt a bles/inde x .html?pag eS ize=50&sor tB review of suicide in economic recession. World J Psychiatry y=none&sor tD ir ect ion=none&ne wq u er y=suici de&conte nt- 5:243–254 type=Refer ence+table &conte nt-type=Datas et. Accessed 1 Apr 3. Patton GC, Coffey C, Sawyer SM et al (2009) Global patterns of mortality in young people: a systematic analysis of population 23. World Health Organization (WHO), The ICD-10 (1993) Clas- health data. Lancet 374:881–892 sification of mental and behavioural disorders: diagnostic criteria 4. The National Confidential Inquiry into Suicide and Homicide for research. World Health Organization, Geneva by People with Mental Illness (2016) (NCISH) Annual Report: 24. Wu CY, Chang CK, Hayes M, Broadbent M, Hotopf M, Stewart England, Northern Ireland, Scotland and Wales. October 2017. R (2012) Clinical risk assessment rating and all-cause mortality University of Manchester in secondary mental healthcare: the South London and Mauds- 5. Arsenault-Lapierre G, Kim C, Turecki G (2004) Psychiatric diag- ley NHS Foundation Trust Biomedical Research Centre (SLAM noses in 3275 suicides: a meta-analysis. BMC Psychiatry 4:37 BRC) Case Register. Psychol Med 42:1581–1590 6. Cavanagh JT, Carson AJ, Sharpe M, Lawrie SM (2003) Psycho- 25. Office for National Statistics (ONS) (2007) Indices of depriva- logical autopsy studies of suicide: a systematic review. Psychol tion 2007 for super output areas. http://www .neigh bour h ood. Med 33:395–405 s t a t i s t i c s . g o v. uk / d i s s e m i n a t i o n / i n s t a n c e S e l e c t i o n . d o ? J S A l l 7. Hawton K, Harris L, Hall S, Smikin S, Bale E, Bond A (2003) o w ed=tr ue&F unct ion=&%24ph=60_61&Cur r e ntP ag Deliberate self-harm in Oxford, 1990–2000: a time of change in eId=61&s tep=2&dat as e tF am il yId =1893&ins t a nceSe lecti patient characteristics. Psychol Med 33:987–995 on=12283 8&Next.x=19&Next.y=17. Accessed 10 Oct 2009 8. World Health Organization (WHO) (2013) Public health action 26. Linsley KR, Schapira K, Kelly TP (2001) Open verdict v. sui- for the prevention of suicide. World Health Organization, Geneva cide—importance to research. Br J Psychiatry 178:465–468 1 3 Social Psychiatry and Psychiatric Epidemiology (2018) 53:1161–1171 1171 27. Cox D (1972) Regression models and life tables (with discus- 49. Shahtahmasebi S (2013) Examining the claim that 80–90% of sions). J R Stat Soc 34:187–220 suicide cases had depression. Front Public Health 1:62 28. Hanley JA, McNeil BJ (1982) The meaning and use of the area 50. National Institute of Care and Clinical Excellence (2016) Depres- under a receiver operating characteristic (ROC) curve. Radiology sion in adults: recognition and management. Clinical guideline 143:29–36 [cg90]. https ://www.nice.org.uk/guida nce/cg90 29. R Core Team (2015) A language and environment for statistical 51. Geekie J, Read J, Renton J, Harrop C (2017) Do English mental computing. R Foundation for Statistical Computing health services know whether they followed N.I.C.E. guidelines 30. National Institute for Health and Clinical Excellence (NICE) with patients who killed themselves? Psychol Psychother Theory (2011) Clinical guidance 133 self-harm. https ://www.nice.org. Res Pract 90:797–800 uk/donotdo/do-no t-use-risk-assessment -t ools-and-scale s-t o-deter 52. Pisani AR, Murrie DC, Silverman MM (2016) Reformulating mine-who-shoul d-and-shoul d-not-be-offer ed-treat ment-or-who- suicide risk formulation: from prediction to prevention. Acad shoul d-be-disch arged . Accessed 3 Sep 2017 Psychiatry 40:623–629 31. Kapur N, Murphy E, Cooper J et al (2008) Psychosocial assess- 53. Carroll R, Metcalfe C, Steeg S et al (2016) Psychosocial assess- ment following self-harm: results from the multi-centre monitor- ment of self-harm patients and risk of repeat presentation: an ing of self-harm project. J Affect Disord 106:285–293 instrumental variable analysis using time of hospital presentation. 32. Department of Health (DoH) (2008) Mental Health Act 2007: PLoS One 11(2):e0149713 patients on aftercare under supervision (ACUS): transitional 54. Fernandes AC, Cloete D, Broadbent MT et al (2013) Development arrangements. Department of Health, London and evaluation of a de-identification procedure for a case register 33. Department of Health (DoH) (2008) Code of practice: Mental coursed from mental health electronic records. BMC Med Inform Health Act 1983. TSO (The Stationer Office) Decis Mak 13:71 34. Hawton K, van Heeringen K (2009) Suicide. Lancet 55. Velupillai S, Mowery D, South BR, Kvist M, Dalianis H (2015) 373:1372–1381 Recent advances in clinical natural language processing in support 35. Coryell W, Young EA (2005) Clinical predictors of suicide in of semantic analysis. Yearb Med Inform 10:183–193 primary major depressive disorder. J Clin Psychiatry 66:412–417 56. Cook BJ, Progovac AM, Chen P, Mullin B, Hou S, Baca-Garcia E 36. Bebbington PE, Minot S, Cooper C et al (2010) Suicidal ideation, (2016) Novel use of natural language processing (NLP) to predict self-harm and attempted suicide: results from the British psychi- suicidal ideation and psychiatric symptoms in a text-based mental atric morbidity survey 2000. Eur Psychiatry 25:427–431 health intervention in Madrid. Comput Math Methods Med. https 37. Minkoff K, Bergman E, Beck AT, Beck R (1973) Hopelessness, ://doi.org/10.1155/2016/87084 34 depression, and attempted suicide. Am J Psychiatry 30:455–459 57. Berrouiguet S, Baca-Garcia E, Brandt S, Walter M, Courtet P 38. Beck AT, Brown G, Berchick RJ, Stewart BL, Steer RA (1990) (2016) Fundamentals for future mobile-health (mHealth): a sys- Relationship between hopelessness and ultimate suicide: a replica- tematic review of mobile phone and web-based text messaging in tion with psychiatric outpatients. Am J Psychiatry 147(2):190–195 mental health. J Med Internet Res 18:e135 39. Baca-Garcia E, Diaz-Sastre C, Garcia-Resa E et al (2005) Sui- 58. Berrouiguet S, Gravey M, Le Galudec M, Alavi Z, Walter M cide attempts and impulsivity. Eur Arch Psychiatry Clin Neurosci (2014) Post-acute crisis text messaging outreach for suicide pre- 255:152–156 vention: a pilot study. Psychiatry Res 217:154–157 40. Guldin MB, Ina Siegismund Kjaersgaard M, Fenger-Grøn M et al 59. Barrett JA, Shetty H, Broadbent M, Cross S, Hotopf M, Stewart (2017) Risk of suicide, deliberate self-harm and psychiatric illness R, Lee W (2016) ‘He left me a message on Facebook’: comparing after the loss of a close relative: a nationwide cohort study. World the risk profiles of self-harming patients who leave paper suicide Psychiatry 16:193–199 notes with those who leave messages on new media. Br J Psychia- 41. Shneidman ES (1993) Suicide as psychache. J Nerv Ment Dis try Open 2:217–220 181:145–147 60. Franklin JC, Ribeiro JD, Fox KR et al (2017) Risk factors for 42. Mann JJ, Waternaux C, Haas GL, Malone KM (1999) Toward a suicidal thoughts and behaviors: a meta-analysis of 50 years of clinical model of suicidal behavior in psychiatric patients. Am J research. Psychol Bull 143:187–232 Psychiatry 156:181–189 61. Free C, Phillips G, Felix L, Galli L, Patel V, Edwards P (2010) 43. Read J, Agar K, Barker-Collo S (2001) Assessing suicidality in The effectiveness of m-health technologies for improving health adults: integrating childhood trauma as a major risk factor. Prof and health services: a systematic review protocol. BMC Res Notes Psychol Res Pract 32:367–372 3:250 44. de Leon J, Baca-García E, Blasco-Fontecilla H (2015) From the 62. Zalsman G, Hawton K, Wasserman D et al (2016) Suicide pre- serotonin model of suicide to a mental pain model of suicide. vention strategies revisited: 10-year systematic review. Lancet Psychother Psychosom 84:323–329 Psychiatry 3:646–659 45. Mann JJ (2003) Neurobiology of suicidal behaviour. Nat Rev Neu- 63. Mann JJ, Apter A, Bertolote J et al (2005) Suicide prevention rosci 4:819–828 strategies: a systematic review. JAMA 294:2064–2074 46. Blasco-Fontecilla H, Baca-Garcia E, Courtet P, Garcia Nieto R, de 64. Riblet NBV, Shiner B, Young-Xu Y, Watts BV (2017) Strategies to Leon J (2015) Horror vacui: emptiness might distinguish between prevent death by suicide: meta-analysis of randomised controlled major suicide repeaters and nonmajor suicide repeaters: a pilot trials. Br J Psychiatry 210:396–402 study. Psychother Psychosom 84:117–119 65. World Health Organization (WHO) (2016) First WHO world sui- 47. Pompili M, Shrivastava A, Serafini G et  al (2013) Bereave- cide report. World Health Organization, Geneva ment after the suicide of a significant other. Indian J Psychiatry 55:256–263 48. Verrocchio MC, Carrozzino D, Marchetti D et al (2016) Mental pain and suicide: a systematic review of the literature. Front Psy- chiatry 7:108 1 3

Journal

Social Psychiatry and Psychiatric EpidemiologySpringer Journals

Published: Jun 2, 2018

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off