A Study Of Influencing Factors And Prediction Models For Attempted Suicide And Its Recurrent Suicidal Behavior

BackgroundSuicide has been prioritized as an urgent public health issue of global importance.According to the World Health Organization statistics,suicide is the 17th leading cause of death worldwide,inflicting more than 700,000 premature deaths every year.Attempted suicide is even more common,which is approximately 20 times more frequent than suicide death.Furthermore,attempted suicide is considered the most important risk factor for future suicide death.The research on the influencing factors of overall attempted suicide could provide guidance on the formulation of universal prevention strategies and measures for suicide.Further research on the potential subtypes of attempted suicide,and the characteristics and influencing factors of different subtypes is conducive to the adoption of targeted prevention and intervention measures,and to a more reasonable allocation of health resources.Currently,a few previous studies have categorized suicide attempters without the priori classification.However,these studies identified an inconsistent number of subtypes,and even no such studies were reported in China.Typologies of attempted suicide under Chinese unique circumstances are of great worth for exploration.The popularity of machine learning(ML),characterized as data-driven approaches,has brought new opportunities and challenges to the establishment of accurate and reliable suicide prediction models.Can ML algorithms contribute to better performance for suicide prediction than traditional regression methods?This is still controversial and needs more studies to validate.In addition,most prior studies on predicting attempted suicide selected only one single type of ML algorithm,despite the ensemble method being an essential strategy to improve prediction performance.To the best of the author’s knowledge,suicide prediction models based on the Stacking ensemble algorithm have not been reported.Suicide attempters are prone to commit suicidal behavior again,including non-fatal repeated attempted suicide and fatal suicide death.Evidence from developed countries indicated that people with prior suicide attempts had a 30 to 100 times higher risk of future suicidal behavior compared to the general population.As the characteristics of attempted suicide in China are quite different from those in developed countries,how about the recurrent suicidal behavior among suicide attempters in China?To date,a few studies on recurrent suicidal behavior after suicide attempt have been reported in China,most of which were limited by the small sample size,short follow-up,high rate of lost to follow-up,and lack of exploration on influencing factors of recurrent suicidal behavior.Therefore,a comprehensive,large-scale,and long-term cohort study of suicide attempters is needed to study recurrent suicidal behavior in China.It is of great significance to develop an accurate and reliable recurrent suicidal behavior prediction model for suicide prevention among suicide attempters.It is also able to provide a reference for the formulation of early prevention strategies and measures for recurrent suicidal behavior.However,the prediction models for recurrent suicidal behavior among suicide attempters have never been reported,and the value of variables accessed at index suicide attempt to predict subsequent suicidal behavior remains to be proved.ObjectivesBased on the background,we conducted the study on attempted suicide and its primary outcome,recurrent suicidal behavior.The objectives of the study were as follows.1.To explore influencing factors of attempted suicide from a wide range of dimensions,to explore the interaction effects of these factors,to identify subtypes of attempted suicide,and to analyze characteristics,influencing factors,and outcomes of different subtypes.2.To develop the risk prediction models of attempted suicide using conventional regression methods and several ML algorithms,and to evaluate and compare their prediction performance.3.To report the rate of recurrent suicidal behavior,and to explore influencing factors and their interactions of recurrent suicidal behavior among suicide attempters.4.To develop,to validate,and to evaluate the risk prediction model of recurrent suicidal behavior based on the random survival forests(RSF)algorithm among suicide attempters.Methods1.A 1:1 matched case-control study design was used to study influencing factors and prediction models of attempted suicide.(1)Participants Fifteen disease surveillance counties in Shandong Province were randomly selected as study sites for data collection.Suicide attempters in these study sites were consecutively recruited as cases.A suicide attempt in the study refers to self-harmed actions with the purpose of death,but it does not result in death.Information of suicide attempters was obtained from the emergency and hospitalization records of large hospitals in the study sites.For each suicide attempter,a community control was selected matched on the same age(±3 years),sex,and residential village.Altogether 1151 suicide attempters and 1151 community controls were enrolled in the study.(2)Data collection Data was collected through one-on-one and face-to-face structured interviews.Variables consisted of demographic characteristics,family and social characteristics,psychological characteristics,psychiatric characteristics,stress,and characteristics of the index suicide attempt.(3)Statistical analysis The characteristics of suicide attempters and community controls were described and compared.Stepwise conditional logistic regressions were performed to explore influencing factors of attempted suicide.Multiplicative and additive interactions of identified influencing factors were tested respectively.Latent class analysis was used to identify subtypes of attempted suicide,and then described the characteristics and analyzed the influencing factors of different subtypes.Logistic regression methods and four ML algorithms,namely,elastic network(EN),artificial neural network(ANN),gradient boosting decision tree(GBDT),and Stacking ensemble algorithms,were used to build prediction models of attempted suicide.Model performance was evaluated by ROC curves,AUC values,accuracy,and so on.Predictor importance was also presented.2.A prospective cohort study design was used to study influencing factors and prediction models of recurrent suicidal behavior.(1)Participants The 1151 suicide attempters were then followed up to collect the outcome of attempted suicide.Ultimately,1103 participants were successfully followed up.(2)Data collection The primary outcome was recurrent suicidal behavior,consisting of non-fatal recurrence(recurrent attempted suicide)and fatal recurrence(suicide).The outcome was ascertained by face-to-face interviews using the recurrent suicidal behavior scale in two follow-up surveys.(3)Statistical analysis The characteristics of suicide attempters with and without recurrent suicidal behavior were described and compared.Incidence density of recurrent suicidal behavior was reported.Cumulative risk curves were estimated by the Kaplan-Meier method,and cumulative incidence rates were calculated.Stepwise Cox proportional hazard models were performed to explore influencing factors of recurrence.Multiplicative and additive interactions of identified influencing factors were tested respectively.Subgroup analyses stratified by sex and age were also conducted.Influencing factors of fatal and non-fatal recurrent suicidal behavior were further analyzed too.Based on the factors identified,the random survival forests(RSF)algorithm was used to develop a prediction model for recurrent suicidal behavior.The concordance index(C index),time-dependent ROC curves,AUC values.Brier scores,and calibration curves were used to evaluate model performance.Predictor importance was also presented.Statistical analyses were carried out using SPSS 24.0,R 4.1.1,and Mplus 7 software.Tests of statistical significance were two-sided,and statistical significance was set at a=0.05.Results1.Influencing factors of attempted suicideAmong 1151 suicide attempters and 1151 paired community controls enrolled in the study,414(36.0%)were males and the mean age was 37 years.Conditional logistic regression showed that education levels(OR=0.80,95%CI:0.75-0.84)and social support(OR=0.91,95%CI:0.88-0.95)were negatively associated with the risk of attempted suicide,mental disorder(OR=2.12,95%CI:1.15-3.91),the mental stimulation of negative life events(OR=1.03,95%CI:1.01-1.04),and depression(OR=1.12,95%CI:1.10-1.14)were positively associated with the risk of attempted suicide.2.Interaction effects between influencing factors of attempted suicideInteraction analyses indicated that negative multiplicative interaction of the three groups of influencing factors had statistical significance,including the interaction between mental stimulation of negative life events(NLEs)and mental disorder(OR of interaction term=0.973,95%CI:0.952-0.996),the interaction between mental stimulation of NLEs and social support(OR=0.997,95%CI:0.996-0.999),and the interaction between mental stimulation of NLEs and depression(OR=0.999,95%CI:0.998-0.999).Positive additive interaction of four groups of influencing factors had statistical significance,including the interaction between depression and educational level(Synergy index,S=1.075,95%CI:1.021-1.142),the interaction between depression and social support(S=1.050,95%CI:1.032-1.086),the interaction between depression and mental stimulation of NLEs(S=1.028,95%CI:1.018-1.050),and the interaction between depression and mental disorder(S=1.067,95%CI:1.010-1.215).3.Latent class analysis of attempted suicideThe latent class analysis identified two subtypes of attempted suicide,715(62.1%)suicide attempters fell into the impulsive subtype and 436(37.9%)suicide attempters fell into the well-planned subtype.The characteristics of the two subtypes were different:impulsive suicide attempters were characterized by lower suicide intent,higher impulsiveness,and were more likely to use pesticide method(P values<0.05);well-planned suicide attempters had advancing age,higher levels of depression and anxiety,and were more likely to use methods other than pesticide and to have histories of suicide attempt(P values<0.05).Furthermore,the influencing factors of the two subtypes were different:they shared education levels.depression,and social support as common factors.The mental stimulation of NLEs and dysfunctional impulsivity were unique influencing factors for impulsive suicide attempters,and family history of suicide and mental disorder were unique influencing factors for well-planned suicide attempters.4.Prediction models of attempted suicidePrediction models of attempted suicide,developed by logistic regression and four ML algorithms(EN,ANN,GBDT,and Stacking algorithms),yielded good predictive performance in the test set,with AUC values larger than 84%and accuracy values larger than 75%.Four ML models achieved slightly better performance than the logistic model,and the Stacking ensemble model was optimal to predict attempted suicide,which reported the largest AUC value(90.2%).Depression,the mental stimulation of NLEs,and social support were the top 3 important predictors for attempted suicide across all prediction models.5.Rate of recurrent suicidal behaviorOf 1151 suicide attempters at baseline,1103 were followed up successfully.During the study period,altogether 49 participants had recurrent suicidal behavior.The incidence density of recurrent suicidal behavior was 6.03 per 1000 person-years.The 1-,3-,5-,and 8-year cumulative incidence rates of recurrent suicidal behavior were 0.91%,2.45%,3.49%,and 4.83%,respectively.The incidence density of recurrence was higher in males than that in females(P=0.017),which was higher in the high age group than that in the low age group(P=0.002).Well-planned suicide attempters reported higher incidence density than impulsive ones(P=0.005).Non-fatal recurrence represented a higher incidence density than fatal recurrence(P=0.007).The highest incidence density was found within 2 years after the index attempted suicide.6.Influencing factors of recurrent suicidal behaviorAge(HR=1.04,95%CI:1.02-1.06),marital status(HR=2.28,95%CI:1.22-4.26),working status(HR=1.90,95%CI:1.06-3.40),mental disorder(HR=2.42,95%CI:1.25-4.69),and anxiety(HR=1.03,95%CI:1.01-1.06)were influencing factors of recurrent suicidal behavior.No interaction effects were found in these factors.In addition,the influencing factors were different between different sex and age groups.The factors of fatal and non-fatal recurrence were also different.7.Prediction model of recurrent suicidal behaviorThe prediction model of recurrent suicidal behavior among suicide attempters developed by the RSF algorithm achieved good predictive performance.The model represented good discrimination and calibration performance,as supported by the C index of 0.826,time-dependent AUC values larger than 79%,Brier scores close to 0,and calibration curves close to diagonal.Furthermore,mental disorder emerged as the most important predictor,with a contribution degree of 22.24%.Conclusions1.Education levels,mental disorder,the mental stimulation of NLEs,depression,and social support were influencing factors of attempted suicide.The negative multiplicative interaction between mental stimulation of NLEs and mental disorder,that between mental stimulation of NLEs and depression,and that between mental stimulation of NLEs and social support had statistical significance.The positive additive interaction between depression and educational level,that between depression and mental disorder,that between depression and mental stimulation of NLEs,and that between depression and social support had statistical significance.2.There were two distinct subtypes of attempted suicide,that is the impulsive subtype and the well-planned subtype,representing different patterns of suicidal behavior.Well-planned suicide attempters were more likely to have recurrent suicidal behavior.This suggested that targeted prevention strategies and measures should be formulated for different subtypes.3.ML algorithms achieved better performance than the logistic model to predict attempted suicide,and the Stacking ensemble model had optimal prediction performance.Depression,the mental stimulation of NLEs,and social support were the top 3 important predictors for attempted suicide across all prediction models.This was a helpful exploration toward the development of prediction models which could be widely used in practice,providing some enlightenment on the method and predictors selected.4.Suicide attempters had a higher risk of suicidal behavior than the general population.Age,marital status,working status,mental disorder,and anxiety were influencing factors of recurrent suicidal behavior.This enhanced the understanding of recurrent suicidal behavior,and emphasized the importance of follow-up monitoring and intervention measures after a suicide attempt.5.Prediction model of recurrent suicidal behavior using RSF algorithm among suicide attempters achieved good predictive performance.It also demonstrated that variables accessed at index suicide attempt were valuable for the prediction of subsequent suicidal behavior.This provides the possibility for accurate prediction and personalized intervention of suicidal behavior.Innovations1.Without the priori classification,this study identified the potential subtypes of attempted suicide with different suicidal patterns,and further compared their characteristics,influencing factors,and outcomes.It is of great significance to understanding the heterogeneity and structure of attempted suicide.2.This study developed prediction models on attempted suicide by applying several algorithms with a wide range of characteristics.The Stacking ensemble algorithm displayed the optimal performance,which improved the prediction performance by integrating the results of various machine learning models.This could provide a reference in methods for the establishment of a suicidal behavior prediction model in the future.3.Based on the prospective cohort of suicide attempters with a long follow-up time,this study explored the risk and influencing factors of recurrent suicidal behavior,and built a prediction model on recurrent suicidal behavior using RSF.The model achieved good predictive performance.The results could provide reference to the prevention of recurrence among suicide attempters.The prediction model of recurrent suicidal behavior has not been reported in the current publicly reported literature.