Characteristics Associated With Serious Self-Harm Events in Children and Adolescents

This study was a secondary analysis of a previous, cross-sectional study designed to validate discharge diagnosis codes for neuropsychiatric events (NPEs) in children using encounter data between April 2016 and March 2020 at Monroe Carell Jr. Children’s Hospital at Vanderbilt in Nashville, TN, and Children's Hospital of Colorado in Aurora, CO.¹⁰ 

Children aged 5 to 17 years were included in the study if they were hospitalized with an International Classification of Diseases, 10th Revision (ICD-10), discharge diagnosis code diagnosis of an NPE. There were 12 NPE categories that were present among the included children: self-harm, mood disorders, psychosis/hallucination, altered mental status, ataxia/movement disorders, encephalitis, seizures, dizziness, headache, sleeping disorders, and vision changes (Supplemental Table 3).¹⁰  Electronic health records were reviewed for each encounter to collect self-reported demographic data (eg, age, sex, race, and ethnicity), concurrent discharge diagnoses, and disposition from the hospital (eg, discharge from the hospital, transfer to psychiatric facility, etc). Payer was defined by the payer for that specific encounter on the basis of hospital claims data and was categorized as government, private, or other. Disposition from hospital was also determined via chart review.

Psychiatric comorbidity was defined using the Child and Adolescent Mental Health Disorder Classification System (CAMHD-CS), which classifies child mental health disorders into 30 unique diagnosis groups across multiple coding systems and aligns with the Diagnostic and Statistical Manual of Mental Disorders (Fifth Edition).¹¹  CAMHD-CS was applied to all discharge ICD-10 diagnosis codes for each encounter to determine if a mental health disorder was present. All CAMHD-CS mental health disorder groups were included except for “suicide or self-injury,” which was included in the self-harm outcome. An additional category of “nonmental health diagnosis” was created for any discharge ICD-10 code that did not fall under the CAMHD-CS categorization system.

The outcome was a physician-confirmed serious self-harm-event that was present on admission and directly related to the hospitalization. A serious self-harm event was defined as any acute suicidal ideation or intentional self-harm event. All cases were adjudicated by a physician to confirm the presence of an intentional self-harm event that was present on admission and related to the hospitalization. To be considered a confirmed case by physician review, the description of the self-harm event (as listed in history of present illness, physical examination, or assessment and plan), was required to match the Diagnostic and Statistical Manual of Mental Disorders, Fifth Edition, definition for psychiatric events or be consistent with a description of a prespecified neurologic event. Physician reviewers were blinded to neuropsychiatric, suicide, and self-harm diagnoses.

Baseline characteristics of the population were summarized by the presence of a serious self-harm event. Bayesian profile regression was used to create profiles of varying risk levels for a self-harm event using diagnosis codes, age, and sex. This approach identifies clusters by analyzing covariate patterns among observations with similar probabilities for the desired outcome.¹²  This Bayesian framework has advantages over traditional clustering approaches, such as latent-class analysis. First, it considers the uncertainty associated with cluster assignments, unlike the “hard” clustering that is used by traditional methods. Second, the number of profiles is automatically determined by the regression model, foregoing the need for goodness-of-fit metrics such as Bayesian information criterion or Akaike information criterion. Third, the model links clusters to an outcome of interest while also using that outcome to drive cluster membership. This means that our model cannot only observe which patterns are most prevalent among the cohort, but specifically the patterns among children with the self-harm outcome. Finally, this method allows for correlation among covariates.¹²  The Bayesian framework also reflects the complexity of real-world clinical scenarios where multiple inputs influence medical decision-making rather than independent risk factors.

The primary Bayesian profile regression model used the 29 psychiatric diagnosis indicators, 1 nonpsychiatric diagnosis indicator, sex, and age as the covariates for the model. Covariates that were considered for profiles were determined a priori through expert consensus between pediatric psychiatrists, general pediatricians, pediatric hospitalists, pediatric complex care physicians, and pediatric pharmacoepidemiologists. Age was categorized into 3 groups: 5 to 9 years, 10 to 13 years, and 14 to 17 years based on CDC reporting¹³  and previous studies which showed significant increases in suicide rates within these subgroups.^14,15 

Within each profile generated, we determined if each covariate had a significant impact on the profile on the basis of the posterior probability of occurrence within that profile. The posterior probability was compared with the baseline previous probability (eg, the probability of a serious self-harm event across all profiles). If the probability of a covariate being present or absent in the profile was significantly different than the baseline probability of presence in the entire cohort, then the presence or absence was considered significant. If the profile probability was not significantly different than the baseline probability, then the covariate was deemed an insignificant contributor for that profile and ignored.

To compare profiles, odd ratios (ORs) with 95% credible intervals (CIs) were reported from the posterior means of each profile compared with a reference profile using bootstrap method.

The reference profile was chosen to be the profile whose risk of a self-harm event was the same as the overall risk of a self-harm event among the entire cohort, with the idea that this profile represents the average probability of a self-harm event over the entire population included in the study.

The PReMiuM package was used to conduct the Bayesian profile regressions, and all analyses were performed using R version 4.1.2.¹⁶  The study was approved by the Vanderbilt University Medical Center institutional review board.

There were 1098 children hospitalized with NPE and included in the study. The median age was 14 (interquartile range [IQR] 11–16, Table 1). A total of 406 (37%) children had a serious self-harm event outcome. Children who had an outcome event were more likely to be older (P < .001), transferred to a psychiatric inpatient facility (P < .001), and self-identify as “white” or “other” race/ethnicity (P = .03, Table 1). There were no differences by sex (P = .12).

Our primary analysis identified 4 distinct profiles of psychiatric comorbidity with varying levels of risk for a serious self-harm event. Profiles were notable for both conditions included, as well as excluded, from the profile. Four distinct profiles with varying risk for a serious self-harm event were identified: (1) low risk, (2) moderate risk, which had an identical level of risk to that of the baseline cohort, (3) high risk, and (4) very high risk. Compared with the average outcome risk of the entire study population, the low-risk profile (n = 494, 45%) had the lowest median risk of 0.04 (IQR 0.03–0.04; OR 0.08, 95% CI 0.04–0.15, Fig 1). The low-risk profile was composed of children aged 5 to 9 years that had a nonmental health diagnosis and an absence of mood disorders, behavioral disorders, psychotic disorders, developmental disorders, and trauma or substance-related disorders (Fig 2). The moderate-risk group had the same risk as that of the baseline risk for the entire cohort (37%) and served as the reference group (n = 88, 8%), with a median risk of 0.30 (IQR 0.27–0.33, Fig 1). This profile was characterized by the inclusion of several mood disorders and behavioral disorders in the absence of depressive disorders. The high-risk profile (n = 393, 36%) had a median risk of 0.69 (IQR 0.67–0.71; OR 5.09, 95% CI 3.11–8.38). This profile was composed of female adolescents aged 14 to 17 years with depression and anxiety in conjunction with substance- and trauma-related disorders. Personality and eating disorders were also significant for this profile. Importantly, the high-risk profile did not include behavioral and developmental disorders (Fig 2). The very high-risk profile (n = 123, 11%) had the highest median risk of 0.79 (IQR 0.73–0.79; OR 7.21, 95% CI 3.69–15.20, Fig 1) and was characterized by male children aged 10 to 13 years. This profile was similar to the high-risk profile because it also included anxiety and depressive disorders. The very high-risk profile differed from the high-risk profile through its inclusion of bipolar disorder, attention-deficit/hyperactivity disorder (ADHD), and trauma-related and developmental disorders like autism spectrum disorder or intellectual disability, along with conduct disorders. Both the high- and very high-risk profiles were also characterized by the absence of a concurrent nonmental health diagnosis (Fig 2).

In this study of >1000 children and adolescent hospitalized with an acute NPE, we identified 4 distinct profiles of psychiatric comorbidity on the basis of risk of an acute self-harm event (median risk ranging from 4% to 76%) using Bayesian profile regression. Our findings reveal unique and distinct patterns of comorbidity that are also distinguishable between age and sex. These profiles can be used to inform clinical decision-making by allowing the provider to better assess overall risk of self-harm.

Children aged 5 to <11 years account for 43% of emergency department (ED) visits for suicidal ideation or suicide attempt but only 2% of hospitalizations.¹⁷  The lowest risk profile in our analysis was characterized by children aged 5 to 9 years who had no major mood, behavioral, or developmental disorders present. Even though they constituted the lowest risk profile, it should be noted that our study population included individuals hospitalized for an NPE and therefore likely represents a population with a greater risk of self-harm events to that of the general population. In our analysis, the very high-risk profile was distinguished from the high-risk profile by the inclusion of younger male adolescents with ADHD, bipolar disorder, autism spectrum disorder, and other developmental disorders. A study of suicide decedents aged 5 to 11 years revealed that suicide deaths were more prevalent among boys; a mental health diagnosis was identified in 31%, with the most common being ADHD, depression, and other unspecified cooccurring disorders.¹⁸  The very high-risk group also reflects a concerning rise in death by suicide among those aged 10 to 13, who have seen rates nearly triple from 2007 to 2017.¹³  Despite demonstrating a greater level of risk for a self-harm event, studies into suicidal behavior in younger children and adolescents continue to receive limited attention.¹⁸  The dichotomy of younger children and adolescents being represented in both the lowest and highest risk profiles highlights the importance of being able to distinguish between high- and low-risk profiles.

Adolescents make up the majority of the hospitalizations for suicidal behavior.¹⁷  The high-risk profile was the second most prevalent profile, with nearly 40% of children falling into this cluster. This cluster was defined primarily by the combination of anxiety and depression in female adolescents aged 14 to 17. The characterization of this profile is supported by previous studies. Previous research has indicated that anxiety and depressive disorders are highly comorbid during youth and that this comorbidity can lead to more severe anxious and depressive symptoms.^19,20 

Although incidence of anxiety and depressive disorders between male and female children is relatively the same before adolescence, female adolescents are twice as likely to be diagnosed with either disorder during adolescence.²¹  Girls also have higher rates of suicidal ideation and attempts after puberty.^22,23  Another notable inclusion in the high-risk group was eating disorders. A recent study demonstrated that ED visits for suicide attempt in adolescent girls were 51% higher in February to March 2021 compared with February to March 2019.²⁴  Another analysis across a similar time period demonstrated that the only mental health conditions for which ED visits significantly increased during this period were eating and tic disorders.²⁵  These studies and others also noted the cooccurrence of eating disorders with other mental health conditions, namely anxiety and depressive disorders.^25–27  The very high-risk group was characterized by males aged 10 to 13, but it is important to note that ED visits for self-inflicted injury have increased nearly 20% among females aged 10 to 14 from 2009 to 2015.¹⁴  The gap in suicide rates among male and female young adolescents has been narrowing.¹⁵  Although the pandemic worsened the mental health crisis, our results indicate that this concerning trend among adolescent girls existed prepandemic, as well. It is important to highlight that these profiles are not static, and changes in risk profile characteristics may occur over time.

Depressive disorders, anxiety disorders, and trauma- or stressor-related disorders were the only 3 psychiatric diagnosis groups that were present among the high- and very high-risk profile. Depressive disorders are the most frequent psychiatric condition in children and are present in 50% to 65% of adolescents who have a suicide attempt.²⁸  The only distinguishable difference between the moderate-risk profile and the 2 highest-risk profiles was that the moderate-risk profile had a significant absence of depressive disorders, suggesting that these disorders play a major role in driving the risk of suicidality. Trauma- and stressor-related disorders have consistently been linked with suicidal ideation in children of all ages, although our analyses suggest that risk for self-harm is mediated by cooccurring anxiety and depression.^29–31 

Our study highlights that a framework using risk profile patterns is better aligned with the real-world complexity and prevalence of psychiatric comorbidities in self-harm events compared with traditional risk factor analyses.⁴  For example, if a male child presents with a combination of ADHD, anxiety, and bipolar disorder, just knowing the risk of self-harm associated with each disorder independent of other comorbidities may not provide an accurate assessment of risk because these disorders do not interact in a simple additive manner. This effect is demonstrated in children who have an anxiety disorder with comorbid depressive disorder because they are more likely to have more severe symptoms than a child who has an anxiety disorder without a depressive disorder.¹⁹  Bayesian profile regression also provides advantages compared with other clustering techniques such as latent-class analysis because Bayesian profiles are mutually informed by the model outcome and our model covariates were allowed to correlate with each other.^12,16 

Our findings are preliminary because the profiles were developed using data from 1000 children with neuropsychiatric complaints at 2 academic children’s hospitals. Future studies should focus on validating these risk profiles in a larger, more heterogeneous population of children and adolescents. Once validated, incorporating these objective risk profiles into prognostic scoring tools and/or clinical decision support applications within the electronic health record could inform disposition and subsequent management decisions, as well as effective resource allocation in the clinical setting.

Our results should be taken in context with several limitations. Although psychiatric diagnoses play an important role in suicidal ideation, there are several factors that could not be included in our analysis such as medication use, previous suicidal behavior,⁴  and family and social support.³²  Factors such as race and insurance were not included in our analysis because these are more reflective of underlying social disadvantage and structural racism. Additionally, previous studies have revealed that race can be replaced by other explanatory variables in decision-making models.³³  We focused on acute self-harm events that led to hospitalizations and were unable to assess self-harm events in settings that did not result in health care encounters. Despite these limitations, we believe that our profiles demonstrate support for a more novel framework that is more comprehensive and can improve clinical decision-making in the future.

Our study identified 4 distinct profiles of psychiatric comorbidity in children and adolescents with varying levels of risk for a serious self-harm event. These clinical profiles are based on a more comprehensive and flexible framework that complements previous assessments and can help identify children at elevated risk for a self-harm event.

ADHD

attention-deficit/hyperactivity disorder

CAMHD-CS

Child and Adolescent Mental Health Disorder Classification System

CI

credible interval

ED

emergency department

ICD-10

International Classification of Diseases, 10th Revision

IQR

interquartile range

NPE

neuropsychiatric event

OR

odds ratio