This work may not be copied, distributed, displayed, published, reproduced, transmitted, modified, posted, sold, licensed, or used for commercial purposes. By downloading this file, you are agreeing to the publisher’s Terms & Conditions.
Risk Factors for Reattempt and Suicide Within 6 Months After an Attempt in the French ALGOS Cohort: A Survival Tree Analysis
Objective: Understanding the cumulative effect of several risk factors involved in suicidal behavior is crucial for the development of effective prevention plans. The objective of this study is to provide clinicians with a simple predictive model of the risk of suicide attempts and suicide within 6 months after suicide attempt.
Methods: A prospective observational cohort of 972 subjects, included from January 26, 2010, to February 28, 2013, was used to perform a survival tree analysis with all sociodemographic and clinical variables available at inclusion. The results of the decision tree were then used to define a simple predictive algorithm for clinicians.
Results: The results of survival tree analysis highlighted 3 subgroups of patients with an increased risk of suicide attempt or death by suicide within 6 months after suicide attempt: patients with alcohol use disorder and a previous suicide attempt with acute alcohol use (risk ratio [RR] = 2.92; 95% CI, 2.08 to 4.10), patients with anxiety disorders (RR = 0.98; 95% CI, 0.69 to 1.39), and patients with a history of more than 2 suicide attempts in the past 3 years (RR = 2.11; 95% CI, 1.25 to 3.54). The good prognosis group comprised all other patients.
Conclusions: By using a data-driven method, this study identified 4 clinical factors interacting together to reduce or increase the risk of recidivism. These combinations of risk factors allow for a better evaluation of a subject’s suicide risk in clinical practice.
Trial Registration: ClinicalTrials.gov identifier: NCT01123174
J Clin Psychiatry 2021;82(1):20m13589
To cite: Demesmaeker A, Chazard E, Vaiva G, et al. Risk factors for reattempt and suicide within 6 months after an attempt in the French ALGOS cohort: a survival tree analysis. J Clin Psychiatry. 2021;82(1):20m13589.
To share: https://doi.org/10.4088/JCP.20m13589
© Copyright 2021 Physicians Postgraduate Press, Inc.
aUniversity of Lille, Inserm, CHU Lille, U1172 – LilNCog—Lille Neuroscience & Cognition, Lille, France
bUniversity of Lille, CHU Lille, ULR 2694, CERIM, Public Health Department, Lille, France
cCentre National de Ressources et de Résilience (CN2R), Lille, France
*Corresponding author: Alice Demesmaeker, MD, MSc, Hôpital Fontan, CHU de Lille, F-59037, Lille Cedex, France (firstname.lastname@example.org).
Articles are selected for credit designation based on an assessment of the educational needs of CME participants, with the purpose of providing readers with a curriculum of CME articles on a variety of topics throughout each volume. Activities are planned using a process that links identified needs with desired results.
To obtain credit, read the article, correctly answer the questions in the Posttest, and complete the Evaluation. A $10 processing fee will apply.
After studying this article, you should be able to:
- Evaluate patients who have attempted suicide for clinical factors that interact to affect the risk of recidivism
The CME Institute of Physicians Postgraduate Press, Inc., is accredited by the Accreditation Council for Continuing Medical Education to provide continuing medical education for physicians.
The CME Institute of Physicians Postgraduate Press, Inc., designates this journal-based CME activity for a maximum of 1 AMA PRA Category 1 Credit™. Physicians should claim only the credit commensurate with the extent of their participation in the activity.
Note: The American Nurses Credentialing Center (ANCC) and the American Academy of Physician Assistants (AAPA) accept certificates of participation for educational activities certified for AMA PRA Category 1 Credit™ from organizations accredited by the ACCME.
Release, Expiration, and Review Dates
This educational activity was published in January 2021 and is eligible for AMA PRA Category 1 Credit™ through February 28, 2023. The latest review of this material was January 2021.
All individuals in a position to influence the content of this activity were asked to complete a statement regarding all relevant personal financial relationships between themselves or their spouse/partner and any commercial interest. The CME Institute has resolved any conflicts of interest that were identified. In the past year, Marlene P. Freeman, MD, Editor in Chief, has received research funding from JayMac and Sage; has been a member of the advisory boards for Otsuka, Alkermes, and Sunovion; and has been a member of the Independent Data Safety and Monitoring Committee for Janssen. No member of the CME
Institute staff reported any relevant personal financial relationships. Faculty financial disclosure appears at the end of the article.
With approximately 800,000 deaths by suicide throughout the world every year, including 10,000 in France, preventing suicide is a major public health issue.1 In 2013, the World Health Assembly adopted the first-ever Mental Health Action Plan of the World Health Organization (WHO).2 This plan aimed to reduce the rates of suicide by 10% from 2013 to 2020. Nonetheless, worldwide suicide rates have been declining over recent decades except in the United States.3 For example, the standardized mortality rate in France has decreased from 16.86 per 100,000 inhabitants in 2011 to 13.21 per 100,000 inhabitants in 2016 according to the national French death register “CépiDc.”4
The understanding of risk factors involved in suicidal behavior is crucial for the development of effective prevention plans. Interestingly, one of the most robust risk factors for death by suicide is a history of previous suicide attempts,5–7 as a substantial number of patients who attempt suicide ultimately die by suicide. According to a recent systematic review and meta-analysis,8 the pooled estimated incidence rate of subsequent fatal self-harm after an index attempt was relatively high at 1 year (1.6%) and even higher at 10 years (4.2%). Moreover, a recent exhaustive study9 of 2 national French registers including 45 million inhabitants highlighted that 1 year after suicide attempt, 2.6% of the subjects had died, with 34.4% of those deaths by suicide. According to that study, the period of greatest risk of death by suicide corresponds to the first 6 months after suicide attempt. In addition, if a previous suicide attempt is a risk factor for subsequent suicide, suicide attempt is also a strong predictor for suicide reattempt. For instance, after 1 year, nonfatal repetition rates are approximately 15%.10 Similar results were found in France, with a recurrence of suicide attempt estimated at 12% within the first year after nonfatal self-harm, with the vast majority (75%) of recurrence of suicide attempt occurring within the first 6 months.9
Beyond the risk related to a previous suicide attempt, several other factors seem to be involved in death by suicide in patients with a history of self-harm. A cohort study7 defined from Danish registers showed that individuals who jumped from a high place or in front of traffic form a high-risk group for suicide. Another Danish cohort study11 highlighted that age over 35 years as well as hanging, strangling, or suffocation as the method of the index attempt and receiving psychiatric hospitalization for that attempt were also risk factors for future suicide.11 Finally, an exploration of Swedish registers12 demonstrated that suffering from severe mental illness such as bipolar or psychotic disorder and a previous attempt by hanging were the most important risk factors for subsequent suicide.
To summarize, repetition of suicide attempt, the use of a violent method for a previous suicide attempt (jumping from a high place or in front of traffic, hanging, strangling, and suffocation), and a severe mental illness are the main factors involved in death by suicide according to different national registers. Moreover, the first 6 months after suicide attempt seem to be a particularly high-risk period for reattempt and subsequent suicide.
Notably, it has been suggested that a subject’s suicide risk was not limited to one event but would rather correspond to the interplay of different factors occurring throughout life from pre-birth (eg, low birth weight and genetic predisposition) to adulthood (socioeconomic factors, severe mental illness, substance abuse).1 In the same way, it has recently been indicated that the lifetime cumulative effect of several risk factors should be considered instead of individual risk factors for suicide.3
While the current scientific literature can determine the extent to which an individual or environmental factor increases the risk of suicide attempt or death by suicide through the use of conventional statistical methods, the published results encounter many drawbacks. The methods employed rarely produce simple tools that can be used by clinicians and hardly explore the interactions between variables.3 Conversely, flexible data-mining methods derived from computer science, such as survival trees, have been developed to detect interactions among variables that best explain a time-dependent variable of interest by means of recursive partitioning.13 The objective of this study is to use such data-miming methods to provide clinicians with a simple predictive model of the risk of suicide attempt and suicide within 6 months after suicide attempt.
A prospective observational cohort of 972 subjects from the ALGOS study was used for the analysis. The ALGOS study was a multicentric, prospective, single-blind, randomized, and controlled clinical trial with 2 parallel groups conducted from January 26, 2010, to February 28, 2013. Participants in the intervention group received a brief contact intervention for 6 months,14 and control participants received no intervention. Both groups were included in the present study, which was conducted in 23 French emergency services. After subjects were chosen for inclusion by an emergency physician, the ALGOS algorithm combines brief contact interventions such as the delivery of a crisis card (with emergency department phone number) for the first suicide attempts and a telephone contact between the 10th and 21st day after suicide attempt for those with previous attempts. Handwritten postcards were also sent at months 2, 3, 4, and 5 to patients with previous suicide attempts who were not available for a phone call or experiencing a suicidal crisis. These interventions were conducted by a team of trained psychologists from the University Hospital of Lille. Patients included were men and women of legal age (ie, at least 18 years old) who survived an suicide attempt in the 7 days prior to inclusion, regardless of the method used for the suicide attempt. Patients who had no suicidal intent, were homeless, were under guardianship, or were multirepeaters (more than 4 suicide attempts in the last 3 years) were excluded from the study. Multirepeaters were excluded from this trial since a lower impact of the brief contact interventions has been demonstrated for these subjects.15 Further details are available in the initial trial report.16 All participants in the ALGOS study provided signed consent.
This study received authorization from AFFSAPS (number NCT01123174) and was approved by the Committee for the Protection of Persons in the North-West Region (CPP North-West decision 09/63). The ALGOS study was registered in ClinicalTrials.gov (identifier: NCT01123174).
At inclusion, data on sociodemographic characteristics (age, sex, family and work status) and number of previous suicide attempts and method of suicide attempt (drug overdose, acute alcohol use), as well as responses to a questionnaire based on Beck and colleagues’ Scale for Suicide Ideation,17 were collected for all subjects. This questionnaire explored the circumstances related to the suicide attempt (expression of suicidal ideations, suicidal project, testamentary precautions, suicide note), predisposing factors (family history of mental illness, chronic pain, chronic medical condition, social isolation), and precipitating factors (self-medication, poor adherence to treatment, negative life events). All participants were also assessed via a screening question or questions from the Mini-International Neuropsychiatric Interview (MINI)18 corresponding to the DSM-IV and ICD-10 criteria for major depressive disorder, anxiety disorder (generalized anxiety disorder or panic disorder), alcohol use disorder (AUD), and eating disorder.
At 6 months, the number of suicidal recurrences and the date of the first recurrence were ascertained for all participants through a standardized telephone interview by trained psychologists.
Descriptive analyses. Descriptive statistics were calculated for the variables of interest. Continuous variables are presented as means and standard deviations (SDs). The 95% confidence intervals (CIs) were calculated using the central limit theorem. Discrete variables are expressed as frequencies and percentages.
Automated risk prediction. To explain the risk of suicide reattempt within 6 months after an index suicide attempt, we performed a survival tree analysis with data for all of the variables available at inclusion (age, sex, family and work status, randomization group [ALGOS or control group], number of previous suicide attempts, suicidal method used for the index suicide attempt, acute alcohol use during index suicide attempt) and clinical data. Survival trees are predictive data mining and decision support tools that allowed us to obtain homogeneous classes of individuals with regard to the time-dependent variable of interest, based on recursive partitioning, by handling interactions between covariates.13 Then, survival curves estimated by the Kaplan-Meier method were drawn corresponding to the leaves of the tree. Tree performance was assessed through a receiver operating characteristic curve and area-under-the-curve (AUC) estimation.
Predictive algorithm for clinicians. Finally, the results of the decision tree were used to define a simple predictive algorithm for clinicians. Kaplan-Meier curves were also drawn, and the AUC was computed. The hazard ratio and its 95% CI were computed using a Cox model.
Only complete cases were analyzed. Tests were 2-sided, and P values <c.05 were considered significant. The Rpart package from R software version 3.6.1 was used for the analysis.19
A total of 972 participants were included (Table 1). Regarding the number of previous suicide attempts, more than half of the patients were first-time attempters (53.3%). In addition, a high proportion of patients suffered from major depressive disorder and anxiety disorder (42.6% and 47.2%, respectively) at inclusion.
For 880 participants (90.5%), the complete status at 6 months was known. Of these patients, 117 (13.3%) had repeated suicide attempts, 11 (1.1%) of whom completed suicide. On those who completed suicide, 6 were female, 8 were non–first-time attempters, 9 were randomized to the control group, and most were suffering from major depressive disorder (7 participants) or anxiety disorder (9 participants). Ninety-two patients (9.5%) were lost to follow-up at 6 months (47 women and 45 men). Fifty-two (56.5%) of them were between 26 and 50 years old, and 53 (57.6%) were first attempters.
Automated Risk Prediction
Of the 878 (90.3%) patients with complete cases (ie, the root node of the tree) that were used for the analysis, 115 (13.1%) had repeated suicide attempts (see Figure 1). The classification and regression trees (CART) algorithm highlighted 5 leaves (ie, defining subgroups), among which 3 presented an increased risk of suicide attempt or suicide.
The first subgroup (Figure 1), leaf 1) was composed of 361 individuals without AUD, with fewer than 2 previous suicide attempts in the past 3 years, and without anxiety disorder. Twenty-three of these participants repeated suicide attempt during the 6 months. The relative event rate (RER) compared to the root was 0.49, and the risk ratio (RR) (95% CI) was 0.36 (0.23 to 0.55).
The next subgroup (Figure 1), leaf 2) was composed of 325 individuals without AUD, with a history of 1 or 2 previous suicide attempts in the past 3 years, and anxiety disorder. Forty-two of those 325 subjects repeated suicide attempt during the 6 months. Suffering from an anxiety disorder (generalized anxiety or panic disorder) nonsignificantly increased the risk of suicide attempt and suicide for these patients: the RER was 1, and the RR (95% CI) was 0.98 (0.69 to 1.39).
In the third subgroup (Figure 1), leaf 3), composed of 46 individuals without AUD and with a history of more than 2 suicide attempts in the past 3 years, 12 individuals repeated suicide attempt. Furthermore, recurrence of suicide attempt was associated with an increased risk of suicide reattempt and suicide: the RER was 2, and the RR (95% CI) was 2.11 (1.25 to 3.54).
The fourth subgroup (Figure 1), leaf 4) was composed of 19 individuals with AUD and without acute alcohol use during their last suicide attempt. None of them repeated suicide attempt. Therefore, the RER was 0.28, and the RR was close to 0.
Finally, the last subgroup (Figure 1), Leaf 5) was composed of individuals with AUD and acute alcohol use (AAU). Among those 127 patients, 38 (29.9%) repeated suicide attempt or died by suicide. The RER was 2.2 in regard to the root, and the RR (95% CI) was 2.92 (2.08 to 4.10).
The survival curves of each terminal leaf representing the time before the first recurrence of suicide attempt within 6 months, estimated by the Kaplan-Meier method, are rendered in Figure 2.
Predictive Algorithm for Clinicians
Based on the previous results, we separated the patients into 2 groups of good or poor prognosis to provide clinicians with a simple predictive algorithm. The poor prognosis group is composed of participants with alcohol use disorder who had acute alcohol use during their last suicide attempt, with anxiety disorder, or with a history of more than 2 suicide attempts in the past 3 years. The good prognosis group comprised all other patients. The survival curves with the 95% CI corresponding to the good and bad prognosis groups are rendered in Figure 3. The poor prognosis group appears to be a risk factor for suicide reattempt (HR = 1.78; 95% CI, 1.42 to 2.24; P < .01). This predictive algorithm obtains an AUC of 0.63. It renders a specificity of 0.47 (95% CI, 0.42 to 0.50), a sensitivity of 0.8 (95% CI, 0.73 to 0.87), a positive predictive value of 0.19 (95% CI, 0.15 to 0.22), and a negative predictive value of 0.94 (95% CI, 0.92 to 0.96).
Through this study, a very simple algorithm to predict suicide reattempt or suicide was proposed. This algorithm was based on 4 simple clinical factors and still had good predictive power. Indeed, 2 groups of good or poor prognosis of suicide reattempt were identified according to the sociodemographic or clinical data collected after their previous suicide attempt. The poor prognosis group is composed of participants suffering from AUD with AAU during their last suicide attempt, those with anxiety disorders such as generalized anxiety or panic disorder, and those with a history of more than 2 suicide attempts in the past 3 years. This predictive algorithm can be easily used by clinicians assessing patients who have attempted suicide to provide the most appropriate medical care.
The results of our survival tree analysis highlighted 3 subgroups of patients with an increased risk of suicide attempt or death by suicide within 6 months after suicide attempt. The first subgroup was composed of patients with an alcohol use disorder and a previous suicide attempt with acute alcohol use. Moreover, subjects with AUD for which previous suicide attempt was associated with an AAU had an even greater risk of suicide reattempt (RER = 2.2), while those without AAU had a lower risk (RER = 0.28). Thus, AAU during the last suicide attempt appears to be a determining factor in the future prognosis. The second subgroup included patients with a history of more than 2 previous suicide attempts in the past 3 years. Finally, the last subgroup was composed of patients with generalized anxiety or panic disorder, without AUD and with less than 2 previous suicide attempts in the past 3 years. Therefore, the increased risk of suicide attempt or death by suicide is best explained by one of these conditions: AUD associated with acute alcohol use during the last suicide attempt, an anxiety disorder, and the recurrence of suicide attempt.
Notably, we used a survival tree analysis to explore the possible interaction between a great number of sociodemographic or clinical variables. Although this method is rarely used in medical research, it allows us to identify the most relevant factors in a dataset without establishing a specific hypothesis and to study the association of different factors regarding a time variable of interest (ie, the risk of suicide/reattempt after suicide attempt). In contrast, traditional methods to analyze survival data are based on the Cox proportional hazard regression model to test a specific link between the covariates and the response. In this type of model, any interaction between variables must be specified beforehand.
Very few studies have focused on the association among risk factors using data-driven methods. Among decision trees, a regression tree was used in a study20 aiming to build a dynamic clinical decision-support system (CDSS) for suicide prevention in 2,802 suicide attempters. Compared to our results, that study identified impulsivity aspects and some interactions of factors that pose a particular risk of recidivism, such as eating disorders for women or a history of familial suicidal behavior and employment status for men. Another study21 that used a decision tree based on the chi-square automatic interaction detection (CHAID) algorithm in 2,754 middle and high schools also found different explanatory factors, including the severity of depression, which seems to be the strongest variable predicting suicide attempt, interacting with greater delinquency and lower family intimacy to increase the risk of suicide attempt. In addition, a classification tree developed to predict a high risk of suicide attempt in 6,686 Chinese high school students indicated interactions among depression, anxiety, social support, gender, self-esteem, family cohesion, and adaptability as predictors of high suicide attempt risk (eg, female adolescents with low social support and low depression scores were at high risk of suicide attempt).22 Therefore, anxiety seems to be the only risk factor found in all 3 studies. Finally, among 218 patients discharged from psychiatric hospitalization after an suicide attempt, a recursive partitioning classification tree highlighted that previous suicide attempt characteristics (such as definite plans or extensive preparation) posed a high risk of reattempt.23 In fact, studies investigating the interactions of factors to assess suicidal risk are rare. However, these emerging methods based on computer science have shown particularly good results to evaluate high-risk groups for suicide in a study comparing different statistical or computer-based methods.24
Studies that used more conventional methods (such as logistic regressions or Cox models) and that focused on few risk factors that are hypothesis-driven have often shown that certain factors are powerful in predicting suicide. According to our results, having an AUD or anxiety disorder (generalized anxiety or panic disorder) appears to be a determining factor in the prognosis after suicide attempt. Regarding psychiatric disorders, a meta-analysis25 obtained similar results. with an increased risk of suicide in patients with opioid use or women with alcohol use disorder. In contrast, a more recent study26 found the highest rates of suicide in subjects with substance use disorder but also the lowest rates in those with anxiety disorders.
Our findings also indicate an interaction between AUD and a previous suicide attempt with acute alcohol use, thus increasing the risk for these patients. Nevertheless, a recent review27 on risk factors for suicide attempt and suicide in patients with substance use disorder did not show any interaction with acute alcohol use during last suicide attempt but did show an association with environmental factors (marital and interpersonal relationship disruption, occupational and financial stressors, recent heavy substance use) as well as a history of previous suicide attempt. Similar results were found in the Australian Treatment Outcome Study28 among subjects suffering from heroin dependence, with high odds ratios for sociodemographic factors (female sex, younger age, less education) and acute substance use (polydrug use, benzodiazepine use, recent heroin overdose).28
Alcohol use disorder is a well-known risk factor for suicidal behavior.29 Regarding AUD, a Korean study30 confirmed an increased risk of suicide attempt with a logistic regression for students who used alcohol daily (odds ratio = 8.00). Moreover, in a recent study31 using logistic regression, an increased risk of suicide attempt was associated alcohol-related deaths following self-harm in addition to unemployment, sickness, or disabled status, among other factors; interestingly, suicide attempt was also associated with alcohol use during previous self-harm. Conversely, our study also found a significant reduction in the risk of recurrence in patients suffering from AUD who had not consumed alcohol during their previous suicidal act. It could be suggested that a reduction or cessation of alcohol consumption or the medical management of the AUD may decrease the risk of suicidal behavior.
Additionally, statistical methods such as multivariate Cox regression models used to predict the risk of recidivism have also shown an increased risk in patients who have made at least 1 attempt in their lifetime.32 Nevertheless, our method suggested an increased risk of reattempt for patients with more than 2 suicide attempts in the past 3 years, thus indicating the need to pay more attention to these patients.
Strengths and Limitations
One of the strengths of our study is that among the large number of variables studied, a small number of factors are ultimately involved in suicide reattempts and suicide and are brought to light by the CART method. The terminal nodes are clinically pertinent and based on simple combinations of medical risk factors. Moreover, the survival curve estimated by the Kaplan-Meier method corresponding to each terminal node showed very different prognoses depending on the subgroup to which they belonged. Another strength is the simple and intuitive nature of the CART algorithm, which allows for easy reading and captures much of the relevant covariate structure of the data.
One of the limitations of this study is the relatively high number of patients lost to follow-up at 6 months (9.5%). Nonetheless, the ALGOS study was composed of a sufficient number of patients, and most of the patients lost to follow-up were younger than 60 years old and were first attempters and thus were at lower risk for suicide attempt and death by suicide. In addition, baseline assessment investigated only 4 psychiatric disorders from the MINI. However, focusing on these common disorders appears to be convenient in clinical practice, especially during a first psychiatric interview in an emergency department setting. Finally, alcohol use disorder is common in France and may affect a large proportion of the subjects in our study. This fact could limit the extrapolation of our results to other parts of the world.
In this study, we suggested that the medical history and the method used for the previous suicide attempt were the most important risk factors for suicidal behavior. Indeed, by using a data-driven method, we found that alcohol use disorder and acute alcohol consumption during the last suicide attempt were associated with an increase in the risk of recidivism, including anxiety disorder and recurrence of suicide attempt. The absence of these factors was related to a decrease in this risk, and the interaction of these various factors contributed to reducing or increasing the risk of recidivism. To better understand the risk factors associated with suicide, we need to consider what happens with combinations of risk factors rather than examining them one by one or by using classical methods such as logistic regression. Suicide prevention should therefore also take place at several levels with better management of psychiatric disorders and increased secondary prevention of suicidal patients, with a focus on those with an alcohol use disorder.
Submitted: July 13, 2020; accepted October 8, 2020.
Published online: February 18, 2021.
Disclosure of off-label usage: The authors have determined that, to the best of their knowledge, no investigational information about pharmaceutical agents or device therapies that is outside US Food and Drug Administration–approved labeling has been presented in this activity.
Financial disclosure: Drs Demesmaeker, Chazard, Vaiva, and Amad have no personal affiliations or financial relationships with any commercial interest to disclose relative to the article.
Funding/support: The authors of this study received no funds for this research. The authors acknowledge the support of the French World Health Organization Collaborating Center in Mental Health and the French “Groupement d’Etude et de Prevention du Suicide” (GEPS).
Role of the sponsor: The providers of support had no role in the conduct of this research and on the composition of this report.
- Flexible data-mining methods derived from computer science, such as survival trees, have been developed to detect interactions among risk factors of suicide reattempt, allowing for identification of homogeneous subgroups of individuals with simple groupings of clinical features.
- Three subgroups of patients presented with an increased risk of suicide reattempt within 6 months after suicide attempt: patients with alcohol use disorder and a previous suicide attempt with acute alcohol use, patients with anxiety disorders, and patients with a history of more than 2 suicide attempts in the past 3 years.
- Risk factor combinations highlighted by a data-driven method of analyzing risk factors allow for a better evaluation of a subject’s suicide risk in clinical practice.