Meta‐analysis of clinical risk factors for suicide among people presenting to emergency departments and general hospitals with suicidal thoughts and behaviours

Suicidal thoughts and behaviours (STB) are a common reason for presentation to emergency departments and general hospitals. A meta‐analysis of the strength of clinical risk factors for subsequent suicide might aid understanding of suicidal behaviour and help suicide prevention.

categorised as higher risk using suicide prediction scales or any other method that combined risk factors had moderately increased odds of suicide (OR = 2.58).Younger age, Black and Hispanic ethnicity, overdose, a diagnosis of adjustment disorder, and the absence of any psychiatric diagnosis were protective against suicide.
Conclusions: Most risk factors for suicide among people who have presented with STB are not strongly associated with later suicide.The strongest risk factors relate to self-harm methods.In the absence of clear indicators of future suicide, all people presenting with suicidality warrant a thorough assessment of their needs, and further research is needed before we can meaningfully categorise people with STB according to suicide risk.
emergency departments, risk assessment, self-harm, suicide

| INTRODUCTION
The World Health Organisation estimates that there are currently about 700,000 annual deaths by suicide and 20 times more suicide attempts. 1About half of all suicides are preceded by self-harm or a suicide attempt. 2 While not all people who self-harm present for medical treatment, 3 self-harm that is severe or distressing enough to bring a person to hospital is associated with a subsequent suicide rate almost 50 times the global suicide rate. 4Hence, people attending a hospital with suicidal thoughts and behaviours (STB) present an obvious opportunity for assessment and care planning. 5In clinical practice, clinicians inevitably consider suicide risk among people with STB and sometimes categorise risk according to lower, medium, or higher suicide risk. 6Some primary research studies have considered whether it is feasible to categorise people as being at increased risk of suicide or repeated self-harm 7,8 however the current view is that available methods of assessment provide insufficient information to guide clinical decision making. 8,9A comprehensive meta-analysis of observational studies reporting on risk factors for suicide among people presenting with STB in emergency departments and general hospitals, could explore the range and statistical strength of clinical risk factors and might assist clinicians and patients in making more informed decisions about their care and aid suicide prevention.

| Objectives of the study
To identify replicated clinical risk factors reported in primary research, and to estimate a pooled effect size for later suicide among people with STB presenting to emergency departments and general hospitals.

Summations
• We report a meta-analysis of a wide range of clinical risk factors for later suicide among people presenting to hospitals with STB.• The strongest statistical risk factors for later suicide were aspects of self-harm, most notably violent methods, rather than other clinical or demographic risk factors.• High-risk categorizations based on suicide risk scales and various ways of combining risk factors was less strongly associated with later suicide than violent methods of self-harm.

Limitations
• There is a conspicuous absence of primary research examining the clinical associations with suicide among people presenting with STB in low-and middleincome countries.• The between study heterogeneity in the metaanalysed variables was generally high, suggesting that the results might not be generalised to all settings.• A number of potentially important risk factors for suicide, including those related to trauma, sexual orientation, and gender identity were not included in the meta-analysis due to lack of primary data.
We performed a meta-analysis according to the Metaanalysis of Observation Studies in Epidemiology (MOOSE) and Preferred Reporting Items for Systematic Reviews and Meta-analysis (PRISMA) guidelines. 10,11he meta-analysis was registered prospectively with PROSPERO (CRD42020211634). 121 | Definitions of self-harm, risk factors, and suicide We examined primary literature that described people with STB using terms including 'self-harm,' 'suicide attempt,' 'deliberate self-harm,' 'non-suicidal self-injury,' 'parasuicide,' and 'suicidal thoughts and behaviour.'We included studies using this broad range of terms because of the clinical overlap of these concepts and populations. 13he dependent variable was later death by suicide.Suicide was defined according to the primary research papers, many of which relied on external mortality data.We included studies that included a proportion of deaths with open verdicts that were considered to be likely suicides. 14he independent variables were 'clinical risk factors,' subsuming factors increasing suicide risk and protective factors.We extracted the independent variables according to the definitions in the primary research.The data synthesis aggregated similar risk factors as described in the primary research.This aggregation of terms (for example, terms like depressed mood, depression, depressive disorder, and major depression) was by consensus between three researchers (CG, JH and ML).Broader groups of clinical risk factors were further aggregated by consensus into six aggregated risk factors: (i) reduced social connectedness (living alone or homeless, divorced or separated, single, widowed, unemployed), (ii) social connectedness (employed, connected to family, married), (iii) psychiatric diagnoses (adjustment disorder, alcohol use disorder, anxiety disorder, bipolar disorder, depressive disorder, affective disorder, personality disorder, psychotic disorders, substance use disorders, and other mental disorders), (iv) mood disorder (affective disorder, bipolar disorder, depressive disorder), (v) any form of hospital admission after self-harm (psychiatric, somatic, or both), and (vi) self-harm by any violent method (including defined as violent in the primary research and self-harm by drowning, firearm, hanging/asphyxiation, and jumping).The statistical strength of higher-risk models derived from suicide risk scales or as defined by the presence of two or more clinical risk factors was also examined, in line with previous research into suicide prediction modelling. 15

| Study inclusion and exclusion criteria
We included English-language peer-reviewed cohort or case-control studies that reported clinical risk factors among suicide decedents and survivors after discharge from emergency departments and general hospitals after presenting with STB.We included studies reporting the numbers or proportions of people in suicide and survivor groups or other effect-size data allowing a data synthesis.
We did not include studies of the effects of suicide prevention interventions, studies comparing the characteristics of suicides with other fatalities, studies reporting risk factors for all-cause mortality, studies reporting on risk factors for repeated suicide attempts, studies aggregating suicides with suicide attempts, nor studies of the characteristics of people who died in hospital because of suicide attempts.Studies that exclusively sampled people discharged from psychiatric inpatient facilities were excluded because they might be atypical of the vast majority of patients who are not admitted.Suicides after discharge from psychiatric hospitals were also the topic of an earlier and related meta-analysis. 16We excluded studies of non-clinical risk factors, including genetic, neuroimaging, and neuro-endocrine data because of their lack of relevance in clinical settings.
Studies with overlapping populations were included if one or more clinical risk factor was unique to each study.Where the same clinical risk factor was reported in the same population, for example, after different follow-up periods, we included the risk factor from the study with the largest number of suicides.

| Search strategy
Embase, MEDLINE ® , PsycInfo, and PubMed were searched for English language papers published between 1 January 1960 and 10 October 2022 using the terms ((suicide*).m_titl.AND (emergency* OR accident and emergency OR casualty OR general hospital OR toxicology service).mp.) or were indexed in PubMed and had titles located with the search terms (suicide* OR self-harm OR self-harm OR self-injury OR self-injury OR self-poisoning OR self-poisoning OR overdose OR para-suicide OR parasuicide [title/abstract]) AND (Emergency department OR emergency room OR Casualty OR general hospital OR toxicology OR accident and emergency [all fields]).The search terms were the same as those used in a related study of rates of suicide among people with STB discharged from non-psychiatric settings. 4ML and CG performed searches of the titles and abstracts.Handsearching the reference lists of identified studies identified additional studies.The resulting sample of full-text papers was screened for inclusion in the meta-analysis by at least two authors, with discordant decisions about inclusion and exclusion resolved by consensus between CG and ML.

| Effect size data
Two authors (CG and AB) independently extracted the total numbers of suicides and survivors and the numbers or proportions with clinical risk factors in each group.When this count data could not be extracted or calculated, effect size data in the form of odds ratios (OR), confidence intervals (CI), and chi-square statistics were extracted.The mean and standard deviation of continuous variables, such as age in both suicide and survivor groups, were extracted.All data about clinical factors reported in each primary research paper was extracted.Disputed data points were reconciled by ML.Papers reporting hazard ratios were analysed separately.

| Study design and strength of reporting data
We used a 12-item purpose-designed Strength of Reporting scale to evaluate study quality based on the Newcastle-Ottawa Scale (NOS) 17 and items used in a related study. 4he scale included seven items relating to 'Representativeness and Selection', two items for the 'Exposure' (independent variables or risk factors), and three items for the 'Ascertainment of Outcome' (death by suicide) (Supplementary Material 1. Strength of reporting scale).Two authors (JH and MB) independently assessed Strength of Reporting; discrepancies were reconciled by discussion, and a third author (ML) resolved unreconciled data.Two authors (JH and CG) extracted length of follow-up data.Where length of follow-up was not reported directly, an estimate was calculated by taking the average of the minimum and maximum follow-up periods (Supplementary Material 2. Length of follow-up calculation).

| Coding of risk factors
Psychiatric diagnoses were coded using ICD-10 18 and DSM-5. 19Bipolar affective and depressive disorders were analysed both separately and together as mood disorders.Obsolete diagnostic labels were mapped to their best current equivalents, decided by discussion between ML and CG.Alcohol use during self-harm events was coded separately from alcohol use disorder.
Somatic or physical health diagnoses were coded as physical comorbidity.
Means of self-harm or suicide attempt were coded broadly as cutting/stabbing, overdoses/self-poisoning, jumping from a height/into traffic, drowning, hanging/ asphyxiation, gassing, firearms, and other methods.Violent methods, lethality and suicide intent were coded according to the primary studies.
Age was examined using three methods.First, in order to examine an effect size associated with increasing age, we dichotomised the reported age groups and combined this with data derived from the continuous variable of age in suicide and non-suicide groups.Second, younger age (<25 years) was compared to all subjects 25 years and over.Third, older age (≥65 years) was compared to subjects under 65.Employment and relationship status were coded according to the primary studies.High-risk groups, determined by a validated risk assessment tool, were categorised together as high-risk regardless of the specific tool used.

| Statistical analysis
A random-effects meta-analysis was used to synthesise the data associated with variables reported in more than three studies using Comprehensive Meta-Analysis (CMA). 20ooled Odds Ratio (OR), 95% Confidence Intervals (CI), Q-value, and I 2 were calculated for each clinical risk variable with a significance level of 0.05.The strength of the meta-analytic estimates was classified as weak if the OR was about 1.5, moderate at 2.5, strong at 4, and very strong at 10 according to established qualitative descriptors of effect size. 21A separate random-effects meta-analysis of hazard ratios was also performed.
For clinical risk factors reported in 10 or more primary studies, we assessed publication bias using a funnel plot and Egger's regression intercept.If the Egger's regression yielded a two-tailed p-value < 0.05, indicating potential publication bias, we used Duval and Tweedie's trim and fill method to estimate the effect size after accounting for this bias.Given the large number of risk factors examined, some factors were expected to show apparent but chance publication bias.Additionally, we calculated sensitivity, specificity, and positive predictive value (for cohort studies) using random-effects metaanalysis for the subset of variables reported in 10 or more primary studies.
We analysed clinical risk factors reported in 10 or more studies using random-effects (method of moments) meta-regression to explore the relationship between study characteristics (sample type, strength of reporting, publication year and follow-up duration) and the pooled effect-size.

| Study selection
Seventy-five relevant studies, published between 1965 and 2022, were found 7, (Figure 1. Flowchart of searches). Threeadditional papers reported hazard ratios that could not be included in the main analysis.[96][97][98] Most studies (except one) were conducted in high-income countries: 31 from Nordic countries, 21 from the UK, 11 from the United States/Canada, four from Australia/ New Zealand, four from high-income Asian countries, and three from mainland Europe.The 75 studies comprised 72 cohort studies and three controlled studies, with 19,649 suicides among 741,624 people. Amongthese, 31 studies focussed on suicide attempts, 25 examined self-harm, 14 studied self-poisoning or overdose, and five used the term "parasuicide" (Table 1.Included studies).A total of 667 risk factors reported in primary research were extracted.Of these, 34 were excluded because they were reported in only one or two studies and 48 were excluded because they contained redundant information included in other variables.Sixtytwo separate clinical risk factors for later suicide were reported in three or more studies.On average there were 9.4 reported effect sizes per risk factor (median 7, total 585 effect sizes) included in the meta-analysis.(Table 2. Pooled odds ratio of clinical risk factors).The mean duration of study follow-up was 7.3 years (SD 7 years, median 5 years, range 1-37 years).

| Strength of reporting
The median strength of reporting for all included studies was eight (out of a maximum of 12), ranging from 5 to

| Statistical strength of clinical risk factors and risk categorisation
The overall association between the clinical risk factors and suicide was weak (OR = 1.37, 95% CI 1.30-1.46)with high heterogeneity (Q-value = 10,324, p < 0.01, I 2 = 93.5).Publication bias was observed towards variables reporting smaller effect sizes (Egger's regression = 0.70, p < 0.01).After adjusting the effect size using the Duval and Tweedie method, the adjusted OR was 1.64 (95% CI 1.54-1.75)with 83 imputed data points to the right of the mean (Figure 2. Funnel plot).Excluding factors that were significantly protective against suicide, the remaining risk factors were associated with an OR of 1.55 (95% CI 1.48-1.63)with high heterogeneity (Q-value = 5389, p < 0.01, I 2 = 89.1)and no evidence of publication bias (Egger's regression intercept = 0.09, p = 0.63).
Eight clinical risk factors were associated with a lower risk of suicide (OR ≤ 0.66, inverse of OR = 1.50) (Table 3. Strength of clinical risk factor matrix).In order of protective effect, these factors were Black ethnicity, Hispanic ethnicity, younger age, diagnosis of an adjustment disorder, absence of psychiatric diagnosis, family connection, and self-harm by taking an overdose or selfpoisoning.
Thirty factors were not associated, or were weakly associated, with later suicide (0.66 < OR < 1.5).These included current social stress, outpatient follow-up referral, being partnered, economic deprivation, employment, other mental disorder, single status, completion of psychosocial assessment, higher education level, recent bereavement, non-rural residence, other method of self-harm (typically other than cutting or overdose), psychiatric symptoms as self-harm precipitant, racial minority (not White, Black or Hispanic), psychotropic medication involvement in self-harm, self-harm by cutting, non-urban residence, historical psychiatric treatment, divorce or separation, recent medical or psychiatric treatment, self-harm involving alcohol, pensioner status, anxiety disorder, substance use disorder, feeling hopeless/worthless/helpless, alcohol use disorder, hospital admission after self-harm, personality disorder, unemployment, and current psychiatric treatment.In addition, four aggregated risk factors (social connectedness (aggregating employed, connected to family, married), any admission after suicide attempt, reduced social connectedness (aggregating living alone or homeless, divorced or separated, single, widowed, unemployed), and any psychiatric diagnosis) were neither strongly protective nor risk factors for suicide.
Seventeen factors were weakly to moderately associated with later suicide (1.5 ≤ OR < 2.5) listed in Note: [1] calculated based on categorical and continuous data so unable to calculate pooled outcome.[2] employed, connected to family, and married.[3] living alone or homeless, divorced or separated, single, widowed, and unemployed.[4] adjustment disorder, alcohol use disorder, anxiety disorder, bipolar disorder, depressive disorder, affective disorder, personality disorder, psychotic disorders, substance use disorders and other mental disorders.[5] affective disorder, bipolar disorder, and depressive disorder.[6] any admission, psychiatric admission, and medical or surgical admission.[7] any violent method included selfharm methods defined as violent in the primary research and self-harm specified to be by drowning, firearm, hanging/asphyxiation, and jumping.
increasing OR order.These factors included psychiatric admission after self-harm, bipolar disorder, suicidal ideation or intent at the time of self-harm, lethal method of self-harm, affective disorder, historical psychiatric admission, depressive disorder, physical comorbidity, widowhood, psychotropic medication use, psychotic disorder, older age, living alone, male gender, repetition of selfharm, age, and self-harm by gassing.The aggregated risk factor any mood disorder was also weakly to moderately associated with later suicide.Three factors moderately to strongly associated with later suicide (2.5 ≤ OR < 4) were being White, high-risk categorisation (using suicide risk scales or based on multiple risk factors), and self-harm by jumping from a height/into traffic.
Three separate factors strongly to very strongly associated with suicide (OR ≥ 4) were self-harm by hanging/ asphyxiation, self-harm by drowning, and self-harm using a violent method.The aggregated variable of selfharm using any violent method (including self-harm methods defined as violent in the primary research and self-harm specified to be by drowning, firearm, hanging/ asphyxiation and jumping) was also associated with an OR ≥ 4.0.Use of a firearm was the sole very strongly associated risk factor, with OR ≥ 10.
Fourteen studies reported 17 high-risk categorizations, nine of which were derived from nominal scales, primarily based on the Suicide Intent Scale. 99Six categorizations were derived using statistical methods, often logistic regression, while two were based on the cooccurrence of clinical risk factors.No study examined clinical judgement or machine learning to define a highrisk category.The meta-analytic estimate of high risk-categorisation was OR = 2.58 (95% CI 2.09-3.19,p < 0.01) with a median OR = 2.80 (range 0.68-6.25).
A subset of 11 papers reported Hazard Ratios (HR) (Supplementary Material 4. Pooled hazard ratio of clinical risk factors).Violent methods and use of a firearm were associated with the highest pooled HR.

| Publication bias and metaregression
Publication bias was assessed for 18 risk factors with at least ten samples (male sex, age, younger age, older age, self-harm by sharp/cutting, self-harm by overdose/ poison, self-harm by another method, self-harm by lethal method, repetition of self-harm, suicidal ideation or intent at the time of self-harm, alcohol use, psychotic disorders, substance use, historical psychiatric treatment, physical comorbidity, current social stress, high-risk categorisation).Some evidence of publication bias was found for male sex and psychotic disorders; adjustment using the trim and fill method did not significantly alter the effect size of the risk factor male, whereas the strength of the risk factor psychotic disorder was significantly affected (adjusted OR = 2.54, 95% CI 2.00-3.23)(see Supplementary Material 5. Publication bias).Length of follow up was associated with a higher OR for psychotic disorders and a lower OR for male sex and lethal method.Year of publication was associated with a lower OR associated with substance use disorders and past psychiatric treatment, and higher OR associated with male sex.A broad definition of self-harm was associated with a lower OR for other self-harm methods and psychotic disorders.Total strength of reporting scores were associated with a lower OR for young age and other self-harm methods and a higher OR associated with current social stress (see Supplementary Material 6. Meta-regression).

| DISCUSSION
We meta-analysed 62 risk or protective factors for suicide among individuals presenting with STB in 75 primary research studies.The main finding was a weak association between most clinical risk factors and suicide.This aligns with Franklin et al.'s study that found a similarly weak association among populations that were not selected by the presence of STB. 100 However, our study revealed an unexpected result: the strongest clinical risk factors for suicide were related to self-harm behaviour, specifically the use of violent methods such as drowning, hanging/asphyxiation, and firearms.These methods are known to be more lethal, 101 increasing the likelihood of fatal outcomes if repeated.
In addition to analysing individual risk factors, we assessed higher-risk categorizations based on suicide risk scales and other methods that combine multiple risk factors, resembling how clinicians consider suicide risk heuristically.Our analysis found that risk categorisation, or suicide risk modelling, was weaker in this setting compared to inpatients 102 and other mental health settings. 15nterestingly, the association between the risk factors of self-harm by jumping, hanging, drowning, firearms and the use of a violent method surpassed the statistical strength of risk categorizations based on risk scales or modelling using two or more risk factors.One possible explanation for the weakness in risk categorisation in this setting is that every subject, in each of the primary studies, had presented with STB, such that the presence or absence of STB, which is generally considered important in suicide risk assessment, could not be used as a basis for discrimination for future suicide.
Our study has several limitations.Firstly, the primary studies were limited to Western and high-income countries, reducing the generalisability of the findings to other regions such as Africa, Asia, and South America where most suicides occur. 1 Secondly, we found heterogeneity between studies for almost every variable, which could not be fully explained by the moderator variables we examined.This suggests that specific contextual factors influencing the relationship between risk factors and later suicide might not have been captured in the metaanalysis.Thirdly, important factors such as sexual orientation, 103 gender diversity, 104 and trauma 7,58 were not inlcuded in three or more sudies and could not be included in the meta-analysis.Fourthly, some risk factors were reported broadly in the primary research such that more specific potential risk factors might have been missed.For example, only three of the twelve effect size data for physical comorbidity specified the medical diagnoses, such that potentially strong associations with particular medical diagnoses (such with neoplastic disorder), might be missed.Fifthly, our meta-analysis focused on associations between clinical risk factors and suicide, but potential interactions between risk factors were not captured.Suicide is believed to result from a complex interaction of underlying factors, 105 which could be examined by a meta-analysis of individual subject-level data, rather than the count and effect size data use in this metaanalysis.Sixthly, no study examined risk factors over periods of less than a year.This raises the possibility that some associations with suicide might have greater (or lesser) statistical strength than we estimated over shorter periods of follow-up.Lastly, it is important to note that the observed associations should not be interpreted as causal relationships.While some factors may have a causal impact (for example, a tendency to use violent suicide methods) our study does not claim to elucidate the causes of suicide among people with STB.
The study suggests limited usefulness of risk categorisation and raises concerns about the effectiveness of suicide risk categorisation after STB.Although there is ongoing research into the utility of suicide risk categorisation after STB, 106 the high rates of suicide in this group of patients 4 and the weakness of statistical risk factors in this setting suggest that health services should prioritise treatment needs and safety planning for all patients presenting with STB rather than just concentrating resources on those considered to be at higher risk.
Clinical interventions might mitigate some of the clinical risk factors.The strong association between selfharm by firearm and suicide highlights the role of heath professionals in preventing firearm deaths.Professionals should know and use firearm notification laws in their jurisdiction and might otherwise encourage patients and carers to limit the ready access to firearms and ammunition.More broadly the presence of affective, depressive, and psychotic disorders as significant risk factors for suicide suggests the importance of treating these conditions, regardless of overall risk classification.People expressing suicidal ideation also carry additional risk, highlighting the potential benefits of psychological therapies in this setting.We found that measures of social connectedness were somewhat protective, while reduced social connectedness was associated with suicide risk, emphasising the importance of addressing social network for people with STB, potentially through peer support workers or allied health staff. 107Unfortunately, most risk factors for suicide were fixed or unmodifiable, including sex, age, repetition, and violent self-harm methods.The preponderance of these fixed risk factors underscores the challenge of reducing suicide in this context.However, we observed that three fixed risk factors (male sex, lethal method, diagnosis of psychotic disorder) showed a declining OR over more extended periods of follow-up, while other dynamic or modifiable factors (suicidal ideation or intent, alcohol use, current social stress, mood disorder) were unchanged over time.These findings challenge the traditional dichotomy between fixed and modifiable suicide risk factors, suggesting that even when addressing dynamic risk factors, the associated risk of suicide persists over time.
To conclude, this meta-analysis emphasises the urgent need for data from low-and middle-income countries, which account for the majority of suicides. 1Future research could also examine the predictive strength of clinical risk categorisation (into heuristically derived higher-and lower-risk groups recorded in the routine care of patients presenting with STB) using co-registered mortality data at follow-up. 15The weakness of high-risk categorizations should not discourage future research, as existing suicide risk scales have not fully incorporated the complete range of risk factors and the application of modern machine learning methods for suicide risk modelling remains untapped.

Funnel
Plot of Standard Error by Log odds ratioF I G U R E 2 Funnel plot of included studies.
Pooled odds ratio of clinical risk factors with 3 or more sample.
Strength of clinical risk factors matrix categorised by risk factor type.
T A B L E 3 a Indicates aggregated risk factor.