Individuals presenting to emergency departments (EDs) for mental health problems have an elevated short-term risk of repeat ED visits (Chang et al., 2014), subsequent hospitalization (Penzenstadler et al., 2020), suicide (Ahmedani et al., 2019), and suicidal behavior following an ED visit (Nock et al., 2010) regardless of whether they present with suicidal symptoms (Olfson et al., 2021). Increasing numbers of people present to EDs with mental health disorders (Burstein et al., 2019; Owens et al., 2017; Plemmons et al., 2018) requiring clinicians to make difficult level of care disposition decisions with limited clinical context.

Suicidal symptoms and behaviors occur along a continuum, which range from the most critical yet rare events of suicide death and attempts to suicidal ideation and self-harm, which occur more frequently (Klonsky et al., 2016). Many studies of suicide risk prediction examine a broad range of suicide-related behaviors, including suicidal self-harm, self-harm, suicidal ideation, and suicide attempt (Burke et al., 2019; Somé et al., 2024), as these have each been observed to confer increased risk for subsequent suicidal behaviors (Ribeiro et al., 2016). Although both suicidal self-harm and non-suicidal self-harm are associated with future suicide attempts (Ribeiro et al., 2016), intent of self-harm behavior can be difficult to discern in emergency department settings and can pose diagnostic challenges. In addition, suicide attempts are often undercounted or recorded only as a physical injury (Werkmeister et al., 2023) and nonfatal overdoses are often coded as non-intentional overdoses even though they are often accompanied by suicidal intent (Connery et al., 2019)). Further, a cross-national study found that 60 % of transitions from ideation to plan or attempt occurred within the first year after onset of ideation across all countries (Nock et al., 2008). Thus, while it is important to distinguish between suicide attempts, self-harm and suicide ideation, timely intervention with all suicidal thoughts and behavior along the suicide continuum may have the potential to impact future risk.

Epidemiological data suggest suicidal behavior risk is driven by numerous small effects that cumulatively increase vulnerability (Bruffaerts et al., 2015). For example, the lifetime risk of suicide among women following inpatient treatment of depression is 3.8 %, but if that risk is considered in combination with a second risk factor, substance use disorder, the lifetime suicide risk increases to 7.1 % (Nordentoft et al., 2011). Another analysis found that lifetime suicide attempt risk increased monotonically from a relative risk of 3.7 for adults with one mental disorder to 29.0 for adults with six or more disorders (Nock et al., 2010). Another study found the effects of 18 separate mental disorders on suicide attempt risk were exerted almost entirely through a general psychopathology factor representing the shared effect across disorders (Hoertel et al., 2015). A large number of potential predictors and their frequent co-occurrence (Hoertel et al., 2015; Kessler et al., 2005; Peyre et al., 2017) support consideration of complex combinations of risk factors (Nock et al., 2010).

Past research in outpatient settings has demonstrated the general feasibility of using machine learning methods to successfully predict suicidal behavior (Barak-Corren et al., 2017; Kessler et al., 2015; McCarthy et al., 2015; Tran et al., 2014). By considering the timing of clinical encounters, their setting, associated diagnoses, and other clinical features, machine learning methods can capture the additive effects of multiple related risk processes (Britton et al., 2012; Crump et al., 2014) and identify underlying service patterns that precede suicide and nonfatal suicide attempts. Electronic health records (EHRs) are a rich data source containing structured and unstructured text of clinical data that can be used to identify patterns of clinical problems and service use to help quantify short-term risk and inform level of care ED decisions. Several prior machine learning models have used selected elements from EHRs (Barak-Corren et al., 2017; Choi et al., 2018; DelPozo-Banos et al., 2018; Karmakar et al., 2016; Simon et al., 2018; Tran et al., 2014; Walsh et al., 2017), but focused on relatively long-time horizons (i.e. >1 year) (Choi et al., 2018; Galfalvy et al., 2008; Kessler et al., 2015) or exclusively on fatal outcomes (McCoy et al., 2016; Simon et al., 2024).

In order to complement and build on prior research (Simon et al., 2024) to provide interpretable insights for clinicians to aid in decision-making regarding the level of care for patients with mental health disorders presenting to the ED, we focused on nonfatal acute suicidal events which are more common than fatal events and key indicators of heightened suicidality risk (Ribeiro et al., 2016; Whitlock and Knox, 2007). Recent advances in artificial intelligence (Guidotti et al., 2018) provide methods for extracting information from complex machine learning models (Wong et al., 2019), which have the potential to improve estimates of future suicidal behavior and changes in risk over time (Huang et al., 2022; Nordin et al., 2022), as machine learning techniques demonstrate greater prediction accuracy over traditional statistical methods (Burke et al., 2019; Pigoni et al., 2024).

This cohort study uses an easily accessible and common source of healthcare system data to model the risk of future suicidal and self-harm events in a large sample of more than two million ED episodes. Using longitudinal structured and unstructured EHR data, we developed a series of models to identify predictors of short-term risk for nonfatal suicidal and self-harm events within 180 days following an ED visit among patients with mental health disorders. This study employed both traditional and machine learning methods to explore the importance of individual clinical features and feature sets as predictors of patient-level outcome risk. By focusing on a critical point of help-seeking, when individuals present to the ED with mental health complaints, the study objective is to harness readily available patient data to inform clinical decision-making regarding treatment planning and guide timely interventions to prevent future suicidal behavior.