ArticlePublisher preview available

The Differences Between Suicide Ideators and Suicide Attempters: Simple, Complicated, or Complex?

Authors:
Xieyining Irene Huang at Florida State University
Xieyining Irene Huang
Jessica D. Ribeiro
Jessica D. Ribeiro
Jessica D. Ribeiro
  • This person is not on ResearchGate, or hasn't claimed this research yet.
Joseph C. Franklin
Joseph C. Franklin
Joseph C. Franklin
  • This person is not on ResearchGate, or hasn't claimed this research yet.
Read publisher preview
Download citation
To read the full-text of this research, you can request a copy directly from the authors.

Abstract

Objective: Suicide ideators and suicide attempters might differ in 3 possible ways. First, they might differ in a simple way such that one or a small set of factors are both necessary and sufficient to distinguish between the 2 groups. Second, ideators and attempters might differ in a complicated way such that a specific combination of a large set of factors is necessary and sufficient for the distinction. Third, complex differences might exist: many possible combinations of a large set of factors may be sufficient to distinguish the 2 groups, but no combination may be necessary. This study empirically examined these possibilities. Method: Across 5 samples (total N = 3,869), univariate logistic regressions were conducted to test for simple differences. To test for complicated and complex differences, machine learning (ML) methods were used to identify the optimized algorithm with all variables. Subsequently, the same methods were repeated after removing the top 5 most important or discriminative variables, and a randomly selected 10% subset of variables. Multiple logistic regressions were conducted with all variables. Results: Results were consistent across samples. Univariate logistic regressions on average yielded chance-level accuracy. ML algorithms with all variables showed good accuracy; substantial deviation from the optimized algorithms through the removal of variables did not result in significantly poorer performance. Multiple logistic regressions produced poor to fair accuracy. Conclusions: Differences between suicide ideators and attempters are complex. Findings suggest that their differences may be better understood on a psychological primitive level than a biopsychosocial factor level. (PsycINFO Database Record (c) 2020 APA, all rights reserved).
A preview of this full-text is provided by American Psychological Association.
Learn more
Content available from Journal of Consulting and Clinical Psychology
This content is subject to copyright. Terms and conditions apply.
The Differences Between Suicide Ideators and Suicide Attempters:
Simple, Complicated, or Complex?
Xieyining Huang, Jessica D. Ribeiro, and Joseph C. Franklin
Florida State University
Objective: Suicide ideators and suicide attempters might differ in 3 possible ways. First, they might differ
in a simple way such that one or a small set of factors are both necessary and sufficient to distinguish
between the 2 groups. Second, ideators and attempters might differ in a complicated way such that a
specific combination of a large set of factors is necessary and sufficient for the distinction. Third,
complex differences might exist: many possible combinations of a large set of factors may be sufficient
to distinguish the 2 groups, but no combination may be necessary. This study empirically examined these
possibilities. Method: Across 5 samples (total N⫽3,869), univariate logistic regressions were conducted
to test for simple differences. To test for complicated and complex differences, machine learning (ML)
methods were used to identify the optimized algorithm with all variables. Subsequently, the same
methods were repeated after removing the top 5 most important or discriminative variables, and a
randomly selected 10% subset of variables. Multiple logistic regressions were conducted with all
variables. Results: Results were consistent across samples. Univariate logistic regressions on average
yielded chance-level accuracy. ML algorithms with all variables showed good accuracy; substantial
deviation from the optimized algorithms through the removal of variables did not result in significantly
poorer performance. Multiple logistic regressions produced poor to fair accuracy. Conclusions: Differ-
ences between suicide ideators and attempters are complex. Findings suggest that their differences may
be better understood on a psychological primitive level than a biopsychosocial factor level.
What is the public health significance of this article?
The present findings indicate that there are indeterminate ways to accurately distinguish between
suicide ideators and suicide attempts: no particular factors or factor combinations are necessary.
Combined with other recent evidence of complex biopsychosocial factor contributions to suicidality,
these findings indicate that the factor-based approach to understanding, predicting, and preventing
suicidality may be inherently limited. Suicidality may be more effectively understood, predicted, and
prevented in terms of a different level of analysis—psychological primitives.
Keywords: suicide, suicidal ideation, suicide attempt, machine learning, complexity
Supplemental materials: http://dx.doi.org/10.1037/ccp0000498.supp
Suicide is the 10th leading cause of death in the United States
(Centers for Disease Control and Prevention, 2016). In 2017, it
was estimated that 10.6 million American adults seriously consid-
ered suicide and 1.4 million attempted suicide (Substance Abuse
and Mental Health Services Administration, 2018). Given that
many people think about suicide and relatively few attempt sui-
cide, researchers have proposed that accurate differentiation be-
tween ideators and attempters can aid suicide prediction and pre-
vention efforts. Questions surrounding this differentiation have
accordingly become of major theoretical and empirical interest
over the past 20 years.
There are three general possibilities for how two groups can
differ on a factor level. First, the groups could differ in a simple
way. It is important to note that simple does not equate to simplis-
tic. The former means uncomplicated and readily understood,
whereas the latter entails oversimplification. If suicide ideators and
attempters differ in a simple way, one factor or a small set of
factors (e.g., 2–10) are both necessary and sufficient to distinguish
between the two groups. By necessary, we mean that without that
factor (or specific small set of factors), it is impossible to distin-
This article was published Online First February 27, 2020.
XXieyining Huang, Jessica D. Ribeiro, and Joseph C. Franklin, De-
partment of Psychology, Florida State University.
Study 1 was supported in part by the Association for Psychological
Science Student Research Grant and the Florida State University Depart-
ment of Psychology Graduate Research Development Award. Study 2 was
supported in part by funding from the Military Suicide Research Consor-
tium (MSRC), an effort supported by the Office of the Assistant Secretary
of Defense for Health Affairs (Award W81XWH-10 –2-0181). Opinions,
interpretations, conclusions, and recommendations are those of the authors
and are not necessarily endorsed by the funding agencies.
Correspondence concerning this article should be addressed to Xieyining
Huang, Department of Psychology, Florida State University, 1107 West
Call Street, Tallahassee, FL 32304. E-mail: huang@psy.fsu.edu
This document is copyrighted by the American Psychological Association or one of its allied publishers.
This article is intended solely for the personal use of the individual user and is not to be disseminated broadly.
Journal of Consulting and Clinical Psychology
© 2020 American Psychological Association 2020, Vol. 88, No. 6, 554–569
ISSN: 0022-006X http://dx.doi.org/10.1037/ccp0000498
554
... Franklin (2019) has similarly argued that multiple distinct combinations of affect and environmental circumstances can lead to suicidal states when they match an individual's internal understanding of what it means to be suicidal. Empirically, these arguments for heterogeneous suicidal states within and between individuals are supported by the failure of computational modeling to identify a single variable or set of variables that reliably predicts suicidal behavior across populations and environments (Goddard et al., 2022;Huang et al., 2020aHuang et al., , 2020b, prompting researchers to argue that suicidal states are heterogeneous. The present results lend further support to this perspective. ...
Individual Differences in Affect During Suicidal and Nonsuicidal States Among Military Personnel and Veterans With Posttraumatic Stress Disorder (PTSD)
Article
Full-text available
  • Feb 2025
Objective: This study examined individual differences in affect during nonsuicidal and suicidal states among 116 military personnel and veterans diagnosed with posttraumatic stress disorder. Method: Current affect and suicidal ideation were assessed four times per day for 14 consecutive days via ecological momentary assessment using the Positive and Negative Affect Scale–Short Form and Scale for Suicidal Ideation (SSI) items. Positive and Negative Affect Scale–Short Form items were used to create four affect states based on the circumplex model of core affect: pleasant activation (e.g., excited), activated pleasure (e.g., proud), unpleasant activation (e.g., aroused), and activated displeasure (e.g., anxious). Mixed-effects regression modeling was used to assess variability in mean affect scores across active (SSI Item 4) and passive (SSI Item 5) suicidal states and variability in the correlation between affect and severity of suicidal ideation. Results: Mean pleasant activation, F(2,7) = 4.8, p = .043, and activated pleasure, F(2, 5) = 22.0, p = .003, were significantly lower during active suicidal versus nonsuicidal states. Mean unpleasant activation, F(2, 7) = 19.9, p = .001, and activated displeasure, F(2, 9) = 42.3, p < .001, were higher during active suicidal versus nonsuicidal states. Activated pleasure was less variable, χ²(1) = 6.3, p = .012, but unpleasant activation, χ²(1) = 8.0, p = .005, and activated displeasure were more variable, χ²(2) = 15.7, p < .001, during active suicidal versus nonsuicidal states. Severity of suicidal ideation was significantly correlated with all four affect states; all correlations varied significantly across participants. Conclusion: Reported affect during suicidal states varies across military personnel and veterans with posttraumatic stress disorder, suggesting suicidal states are heterogeneous. Differentiating affective arousal from affective valence can provide more nuanced understandings of suicide risk.
View
... The potential of ML models was illustrated by several prominent studies on the prediction of suicidal behavior, which reported substantially higher predictive performance of ML models compared to more traditional models such as logistic regression (e.g., Fox et al., 2019;Huang et al., 2020;Walsh et al., 2018). The authors of these studies consistently interpreted their findings as strong evidence for substantial interaction effects between common risk factors for suicidal behavior. ...
Beyond the Hype: A Simulation Study Evaluating the Predictive Performance of Machine Learning Models in Psychology
Preprint
Full-text available
  • Dec 2024
Although Machine learning (ML) models are gaining popularity in psychological research, the debate about their usefulness ranges from hype to disillusionment. The discrepancy between the hopes placed in ML and the psychometric reality is often attributed to the quality of psychological datasets, which tend to be small and subject to imprecise measurement. In this simulation study, we examined the data requirements necessary for complex ML algorithms to perform well. We compared the performance of Elastic Net Regressions and Gradient Boosting Machines (GBM) under various conditions: (a) sample size, (b) number of irrelevant predictors, (c) predictor reliability, (d) effect size, and (e) nature of the data generating model (i.e., linear vs. non-linear effects). Our study had two major goals: First, we investigated whether the models achieved the maximum predictive performance (i.e., recovery of the simulated effects). Second, we examined whether the respective models were able to accurately distinguish between relevant and irrelevant predictors. Our results showed that Elastic Net Regressions with pre-specified interaction terms outperformed GBM models in almost all conditions, but still only achieved the maximum possible predictive performance under optimal conditions (N = 1,000, perfectly reliable predictors, predominantly linear effects, and an exceptionally large effect size of R² = .80), which are arguably rarely met in psychological research. We stress that data quality fundamentally limits the performance of ML models in a similar way to more traditional regression analyses and discuss the validity of comparing increasingly complex models in psychological research.
View
... Consistent with this perspective, machine learning has been increasingly used to analyze the complex relationship among suicide and its many factors (15,16). While machine learning has proven useful for modeling nonlinear and interactive relationships among multiple risk and protective factors, machine learning models rarely replicate across samples, limiting their clinical utility in general practice settings (17)(18)(19). While controlled experiments like randomized clinical trials can offer insights into the causes of suicidal behavior, this approach is often expensive and infeasible to employ with sufficient speed and scale, implicating the potential value of alternative approaches. ...
Multi-environment prediction of suicidal beliefs
Article
Full-text available
  • Nov 2024
Suicide disproportionately affects the military and veteran population, yet the task of identifying those at an increased risk of suicidal behavior remains challenging. In the face of this complex issue, novel machine learning methods have been applied to study the relationship between suicide and potential risk factors, but are often not generalizable to new and unseen samples. Herein, we examine the problem of prediction on unknown environments, commonly known as environment-wise domain adaptation, as it relates to the prediction of suicidal beliefs, measured with items from the Suicide Cognitions Scale (SCS). We adapt several recently invariance-based models trained using a sample consisting of people without any prior suicidal ideation (SI) to the prediction of suicidal beliefs of those with prior SI. In addition, we examine the possible causal relations regarding the SCS. Using a prospective sample of 2744 primary care patients with 17 risk and protective factors, we show that, to some extent, these methods are able to generalize to a new environment, namely, a sample with prior SI. Additionally, our results indicate suicidal ideation and suicidal behavior are likely to be causal children of SCS.
View
... The considerable dynamics of suicidal ideation and behaviour -in addition to other factors (6) -make it incredibly difficult to reliably assess a person's suicide risk (7,8). In addition, a wide variety of constellations of stress factors can lead to suicidal behaviour (9) and suicidal experience and behaviour is therefore motivated very differently in individual cases (10). Treatment approaches for suicidal patients must do justice to this 'individuality' of suicidal ideation and behaviour (11). . ...
Ambivalence model of suicidality [ABS-model]: an orientation model for the treatment of suicidal individuals
Article
Full-text available
  • Nov 2024
The existing models for understanding suicidal ideation and behavior do not provide satisfactory orientation for clinical-therapeutic work with suicidal clients. Based on the observation that ambivalence accompanies the entire suicidal process and building on the empirical knowledge about suicidal ambivalence, this article presents the ambivalence model of suicidality (ABS model), a new clinical working model that aims to provide a framework for risk assessment, case conceptualization and treatment planning in the treatment of suicidal individuals. The model divides the suicidal process into three phases (uncertainty phase, transition phase and action phase), describes the psychological state within the different phases, and identifies phase-specific therapeutic interventions. The ABS model is a descriptive model that can be used to structure and organize crisis intervention and psychotherapy with suicidal patients.
View
... 44,100 Similarly, sex assigned at birth is not depicted: although some find elevated attempts 8,121 and greater relative risk of suicide death in autistic females than males, 3 overall more autistic males than females die by suicide. 3,9,10 Despite advances in non-autistic populations, [28][29][30][31][32] developments in predicting and preventing suicide for autistic people are hampered by the fact that well-established constructs related to suicidality do not necessarily apply to autistic people. [33][34][35][36][37][38] Reduced fear of death appears to distinguish autistic 33,35 and non-autistic 26,39,40 individuals with historical suicide attempts, but 'perceived burdensomeness' has been linked not only with suicide ideation but also suicide attempts in autistic people. ...
Suicidality in autistic people: a mixed-methods approach to contributing factors
Preprint
Full-text available
  • Nov 2024
Suicide is a leading cause of death for autistic people. Efforts to understand the ideation-to-action trajectory through theories developed in non-autistic people have met with limited success. Autistic people’s perceptions of the factors underpinning suicidal experiences are a vital source of information which may advance conceptual understanding, assist risk assessment, and guide policy and clinical practice about suicide prevention. In a large autistic sample (n=1369), we explored which factors autistic people perceived as the most important contributing factors to their suicidal thoughts and feelings; whether these differed between autistic people of different genders and ages; and whether contributing factors could usefully differentiate between people with different lifetime experience with suicidal thoughts and behaviour.
View
... Die erhebliche Dynamik suizidalen Erlebens und Verhaltens verunmöglicht dabei -neben anderen Faktoren [Teismann et al., 2019] -eine sichere Einschätzung des Suizidrisikos einer Person [Franklin et al., 2017;Woodford et al., 2019]. Hinzu kommt, dass die unterschiedlichsten Konstellationen von Belastungsfaktoren in suizidales Verhalten münden können [Huang et al., 2020], suizidales Erleben und Verhalten somit im Einzelfall sehr unterschiedlich motiviert ist [May et al., 2020]. Dieser "Individualität" suizidalen Erlebens und Verhaltens müssen Behandlungsansätze suizidaler Patient:innen gerecht werden [Jobes, 2023]. ...
Ambivalenzmodell der Suizidalität (ABS-Modell): Ein Verstehens- und Orientierungsmodell für die Behandlung suizidaler Personen
Article
  • Aug 2024
b> Einleitung: Suizidales Erleben und Verhalten weist eine erhebliche Dynamik auf und ist sehr heterogen motiviert. Ein genaues Verständnis der individuellen Funktionsweise sui­zidalen Erlebens und Verhaltens stellt die Voraussetzung einer therapeutischen Aufarbeitung suizidaler Krisen dar. Methode: Ausgehend von der Feststellung, dass ambivalentes Erleben den gesamten suizidalen Prozess begleitet, wird im vorliegenden Artikel das Ambivalenzmodell der Suizidalität (ABS-Modell) als Verstehens- und Orientierungsmodell für die klinische Praxis vorgestellt. Ergebnisse: In dem Modell wird der suizidale Prozess in drei Phasen unterteilt (Unsicherheits-, Transitions- und Handlungsphase), es wird die psychische Befindlichkeit innerhalb der verschiedenen Phasen beschrieben und es werden phasenspezifische the­rapeutische Ansatzpunkte benannt. Diskussion/Schlussfolgerung: Beim ABS-Modell handelt es sich um ein deskriptives Modell, anhand dessen die Krisenintervention und Psychotherapie suizidaler Personen strukturiert und gestaltet werden kann.
View
View
View
Development of a Subtle Screening for Suicidal Ideation: Psychometric Characteristics and Implications for Family Therapists
Article
Full-text available
  • Jun 2024
  • Contemp Fam Ther
While 10 million U.S. adults experience suicidal thoughts and 1.7 million attempt suicide annually, candid, truthful endorsement of assessment items related to suicidal ideation (SI) can be inhibited by stigma, shame, and embarrassment and a fear of involuntary hospitalization. Suicidal ideation in, and suicide attempts by, family members increase the suicide risk among other members by several times, and so accurate detection of SI is crucial for couple and family therapists. To address concerns about stigma and false negatives in screening for SI, a 19-item subtle screening of suicidal ideation (SSSI) was developed from a pool of 32 “proxy” items tapping psychological pain, emotional intelligence, and negative alterations in mood and cognitions. A demographics form, a measure of suicidal ideation, measures of anxiety, depression, and traumatic stress, and versions of the Beck Hopelessness Scale and the Interpersonal Needs Questionnaire were also used for data collection. Principal components analysis and reliability, correlation, and multiple regression procedures on data from a non-random, diverse sample of adults (N = 306) provided evidence of excellent reliability (α = .93) and convergent and discriminant validity for the SSSI. The three-dimensional SSSI accounted for 54.9% of the variance in a direct measure of suicidal thoughts, and a Receiver Operator Characteristic curve identified a cut-off score of 35 with a sensitivity of .937 and specificity of .81, indicating the instrument successfully identifies those with and without suicidal thoughts. Clinical implications and future research are discussed.
View
View
Relationships Among Suicide Ideators, Attempters, and Multiple Attempters in a Young-Adult Sample
Article
Full-text available
  • Nov 1996
The relationships among suicide ideators, attempters, and multiple attempters were explored in 332 psychiatric patients referred specifically for suicidal ideation or behavior. Previous researchers have subsumed multiple attempters under the general category of attempters. However, comparisons across a range of variables, including Axis I diagnoses from the revised 3rd edition of the Diagnostic and Statistical Manual of Mental Disorders (American Psychiatric Association, 1987) depressive and anxiety symptoms, suicidal ideation, hopelessness, problem solving, and a range of personality features revealed that multiple attempters presented a more severe clinical picture and, accordingly, elevated suicide risk compared with attempters and ideators. Observed differences between groups were maintained when attempters with “questionable intent” (i.e., those making equivocal attempts) were excluded from the analyses.
View
Psychological primitives can make sense of biopsychosocial factor complexity in psychopathology
Article
Full-text available
  • Oct 2019
  • BMC MED
Background: Many agree that the biopsychosocial contributions to psychopathology are complex, yet it is unclear how we can make sense of this complexity. One approach is to reduce this complexity to a few necessary and sufficient biopsychosocial factors; although this approach is easy to understand, it has little explanatory power. Another approach is to fully embrace complexity, proposing that each instance of psychopathology is caused by a partially unique set of biopsychosocial factors; this approach has high explanatory power, but is impossible to comprehend. Due to deficits in either explanatory power or comprehensibility, both approaches limit our ability to make substantial advances in understanding, predicting, and preventing psychopathology. Thus, how can we make sense of biopsychosocial factor complexity? Main text: There is a third possible approach that can resolve this dilemma, with high explanatory power and high comprehensibility. This approach involves understanding, predicting, and preventing psychopathology in terms of a small set of psychological primitives rather than biopsychosocial factors. Psychological primitives are the fundamental and irreducible elements of the mind, mediating all biopsychosocial factor influences on psychopathology. All psychological phenomena emerge from these primitives. Over the past decade, this approach has been successfully applied within basic psychological science, most notably affective science. It explains the sum of the evidence in affective science and has generated several novel research directions. This approach is equally applicable to psychopathology. The primitive-based approach does not eliminate the role of biopsychosocial factors, but rather recasts them as indeterminate causal influences on psychological primitives. In doing so, it reframes research away from factor-based questions (e.g., which situations cause suicide?) and toward primitive-based questions (e.g., how are suicidality concepts formed, altered, activated, and implemented?). This is a valuable shift because factor-based questions have indeterminate answers (e.g., infinite situations could cause suicide) whereas primitive-based questions have determinate answers (e.g., there are specific processes that undergird all concepts). Conclusion: The primitive-based approach accounts for biopsychosocial complexity, ties clinical science more directly to basic psychological science, and could facilitate progress in understanding, predicting, and preventing psychopathology.
View
Model Complexity Improves the Prediction of Nonsuicidal Self-Injury
Article
Full-text available
  • Jun 2019
Objective: Efforts to predict nonsuicidal self-injury (NSSI; intentional self-injury enacted without suicidal intent) to date have resulted in near-chance accuracy. Incongruence between theoretical understanding of NSSI and the traditional statistical methods to predict these behaviors may explain this poor prediction. Whereas theoretical models of NSSI assume that the decision to engage in NSSI is relatively complex, statistical models used in NSSI prediction tend to involve simple models with only a few theoretically informed variables. The present study tested whether more complex statistical models would improve NSSI prediction. Method: Within a sample of 1,021 high-risk self-injurious and/or suicidal individuals, we examined the accuracy of three different model types, of increasing complexity, in predicting NSSI across 3, 14, and 28 days. Univariate logistic regressions of each predictor and multiple logistic regression with all predictors were conducted for each timepoint and compared with machine learning algorithms derived from all predictors. Results: Results demonstrated that model complexity was associated with predictive accuracy. Multiple logistic regression models (AUCs 0.70-0.72) outperformed univariate logistic models (average AUCs 0.56). Machine learning models that produced algorithms modeling complex associations across variables produced the strongest NSSI prediction across all time points (AUCs 0.87-0.90). These models outperformed all multiple logistic regression models, including those involving identical study variables. Machine learning algorithm performance remained strong even after the most important factor across algorithms was removed. Conclusions: Results parallel recent findings in suicide research and highlight the complexity that underlies NSSI. (PsycINFO Database Record (c) 2019 APA, all rights reserved).
View
Transition to suicide attempt from recent suicide ideation in U.S. Army soldiers: Results from the Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS)
Article
Full-text available
Background Most people with suicide ideation (SI) do not attempt suicide (SA). Understanding the transition from current/recent SI to SA is important for mental health care. Our objective was to identify characteristics that differentiate SA from 30‐day SI among representative U.S. Army soldiers. Methods Using a unique case–control design, soldiers recently hospitalized for SA (n = 132) and representative soldiers from the same four communities (n = 10,193) were administered the same questionnaire. We systematically identified variables that differentiated suicide attempters from the total population, then examined whether those same variables differentiated all 30‐day ideators (n = 257) from the total population and attempters from nonattempting 30‐day ideators. Results In univariable analyses, 20 of 23 predictors were associated with SA in the total population (0.05 level). The best multivariable model included eight significant predictors: interpersonal violence, relationship problems, major depressive disorder, posttraumatic stress disorder (PTSD), and substance use disorder (all having positive associations), as well as past 12‐month combat trauma, intermittent explosive disorder (IED), and any college education (all having negative associations). Six of these differentiated 30‐day ideators from the population. Three differentiated attempters from ideators: past 30‐day PTSD (OR = 6.7 [95% CI = 1.1–39.4]), past 30‐day IED (OR = 0.2 [95% CI = 0.1–0.5]), and any college education (OR = 0.1 [95% CI = 0.0–0.6]). The 5% of ideators with highest predicted risk in this final model included 20.9% of attempters, a four‐fold concentration of risk. Conclusions Prospective army research examining transition from SI to SA should consider PTSD, IED, and education. Combat exposure did not differentiate attempters from ideators. Many SA risk factors in the Army population are actually risk factors for SI.
View
Risk Factors for the Transition from Suicide Ideation to Suicide Attempt: Results from the Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS)
Article
Full-text available
  • Feb 2018
Prior research has shown that most known risk factors for suicide attempts in the general population actually predict suicide ideation rather than attempts among ideators. Yet clinical interest in predicting suicide attempts often involves the evaluation of risk among patients with ideation. We examined a number of characteristics of suicidal thoughts hypothesized to predict incident attempts in a retrospective analysis of lifetime ideators (N = 3,916) drawn from a large (N = 29,982), representative sample of United States Army soldiers. The most powerful predictors of first nonfatal lifetime suicide attempt in a multivariate model controlling for previously known predictors (e.g., demographics, mental disorders) were: recent onset of ideation, presence and recent onset of a suicide plan, low controllability of suicidal thoughts, extreme risk-taking or “tempting fate,” and failure to answer questions about the characteristics of one’s suicidal thoughts. A predictive model using these risk factors had strong accuracy (area under the curve [AUC] = .93), with 66.2% of all incident suicide attempts occurring among the 5% of soldiers with highest composite predicted risk. This high concentration of risk in this retrospective study suggests that a useful clinical decision support model could be constructed from prospective data to identify those with highest risk of subsequent suicide attempt.
View
View
Predicting Imminent Suicidal Thoughts and Nonfatal Attempts: The Role of Complexity
Article
  • May 2019
For decades, our ability to predict suicidal thoughts and behaviors (STBs) has been at near-chance levels. The objective of this study was to advance prediction by addressing two major methodological constraints pervasive in past research: (a) the reliance on long follow-ups and (b) the application of simple conceptualizations of risk. Participants were 1,021 high-risk suicidal and/or self-injuring individuals recruited worldwide. Assessments occurred at baseline and 3, 14, and 28 days after baseline using a range of implicit and self-report measures. Retention was high across all time points (> 90%). Risk algorithms were derived and compared with univariate analyses at each follow-up. Results indicated that short-term prediction alone did not improve prediction for attempts, even using commonly cited “warning signs”; however, a small set of factors did provide fair-to-good short-term prediction of ideation. Machine learning produced considerable improvements for both outcomes across follow-ups. Results underscore the importance of complexity in the conceptualization of STBs.
View
Virtual reality suicide: Development of a translational approach for studying suicide causes
Article
  • Dec 2018
  • BEHAV RES THER
Causal knowledge is crucial for understanding and preventing suicide. Unfortunately, we have little direct knowledge about suicide causes because we cannot conduct experiments that seek to make suicide more likely. In such situations, translational approaches can provide valuable, though tentative, information. We sought to establish a new translational approach by developing a laboratory approximation of suicide with new virtual reality (VR) technologies. Such an approach would allow researchers to tentatively investigate the causes of suicide by conducting experiments that introduce purported causes of suicide and observe their effects on VR suicide rates. Across three studies (total N = 498), results indicated that our two VR suicide scenarios (jumping from heights; shooting oneself) were safe; rated as unpleasant, realistic, and suicide-relevant; associated with several relevant predictors of VR suicide completion, including male sex, suicidal desire, suicidal capability, agitation, and prior suicidality; associated with reasons for not engaging in VR suicide that are similar to the reasons people give for not engaging in actual suicide; and produced 5% completion rates under neutral conditions and 25% completion rates after reward/avoid manipulations. We hope that future work further improves this approach and applies it to more directly test ideas about suicide causes and suicide prevention.
View
Predicting drug-resistant epilepsy — A machine learning approach based on administrative claims data
Article
  • Dec 2018
Patients with drug-resistant epilepsy (DRE) are at high risk of morbidity and mortality, yet their referral to specialist care is frequently delayed. The ability to identify patients at high risk of DRE at the time of treatment initiation, and to subsequently steer their treatment pathway toward more personalized interventions, has high clinical utility. Here, we aim to demonstrate the feasibility of developing algorithms for predicting DRE using machine learning methods. Longitudinal, intersected data sourced from US pharmacy, medical, and adjudicated hospital claims from 1,376,756 patients from 2006 to 2015 were analyzed; 292,892 met inclusion criteria for epilepsy, and 38,382 were classified as having DRE using a proxy measure for drug resistance. Patients were characterized using 1270 features reflecting demographics, comorbidities, medications, procedures, epilepsy status, and payer status. Data from 175,735 randomly selected patients were used to train three algorithms and from the remainder to assess the trained models' predictive power. A model with only age and sex was used as a benchmark. The best model, random forest, achieved an area under the receiver operating characteristic curve (95% confidence interval [CI]) of 0.764 (0.759, 0.770), compared with 0.657 (0.651, 0.663) for the benchmark model. Moreover, predicted probabilities for DRE were well-calibrated with the observed frequencies in the data. The model predicted drug resistance approximately 2 years before patients in the test dataset had failed two antiepileptic drugs (AEDs). Machine learning models constructed using claims data predicted which patients are likely to fail ≥3 AEDs and are at risk of developing DRE at the time of the first AED prescription. The use of such models can ensure that patients with predicted DRE receive specialist care with potentially more aggressive therapeutic interventions from diagnosis, to help reduce the serious sequelae of DRE.
View
Predicting suicide attempts in adolescents with longitudinal clinical data and machine learning
Article
  • Apr 2018
Background Adolescents have high rates of nonfatal suicide attempts, but clinically practical risk prediction remains a challenge. Screening can be time consuming to implement at scale, if it is done at all. Computational algorithms may predict suicide risk using only routinely collected clinical data. We used a machine learning approach validated on longitudinal clinical data in adults to address this challenge in adolescents. Methods This is a retrospective, longitudinal cohort study. Data were collected from the Vanderbilt Synthetic Derivative from January 1998 to December 2015 and included 974 adolescents with nonfatal suicide attempts and multiple control comparisons: 496 adolescents with other self‐injury (OSI), 7,059 adolescents with depressive symptoms, and 25,081 adolescent general hospital controls. Candidate predictors included diagnostic, demographic, medication, and socioeconomic factors. Outcome was determined by multiexpert review of electronic health records. Random forests were validated with optimism adjustment at multiple time points (from 1 week to 2 years). Recalibration was done via isotonic regression. Evaluation metrics included discrimination (AUC, sensitivity/specificity, precision/recall) and calibration (calibration plots, slope/intercept, Brier score). Results Computational models performed well and did not require face‐to‐face screening. Performance improved as suicide attempts became more imminent. Discrimination was good in comparison with OSI controls (AUC = 0.83 [0.82–0.84] at 720 days; AUC = 0.85 [0.84–0.87] at 7 days) and depressed controls (AUC = 0.87 [95% CI 0.85–0.90] at 720 days; 0.90 [0.85–0.94] at 7 days) and best in comparison with general hospital controls (AUC 0.94 [0.92–0.96] at 720 days; 0.97 [0.95–0.98] at 7 days). Random forests significantly outperformed logistic regression in every comparison. Recalibration improved performance as much as ninefold – clinical recommendations with poorly calibrated predictions can lead to decision errors. Conclusions Machine learning on longitudinal clinical data may provide a scalable approach to broaden screening for risk of nonfatal suicide attempts in adolescents.
View