A cohort study of predictors of short-term nonfatal suicidal and self-harm events among individuals with mental health disorders treated in the emergency department.

A cohort study of predictors of short-term nonfatal suicidal and self-harm events among individuals with mental health disorders treated in the emergency department - PMC Skip to main content An official website of the United States government Here's how you know Here's how you know Search Log in Primary site navigation Search Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice J Psychiatr Res . Author manuscript; available in PMC: 2026 Jun 1. Published in final edited form as: J Psychiatr Res. 2025 Apr 22;186:458–468. doi: 10.1016/j.jpsychires.2025.04.035 A cohort study of predictors of short-term nonfatal suicidal and self-harm events among individuals with mental health disorders treated in the emergency department Steven C Marcus a , Sara Wiesel Cullen a , Timothy Schmutte b , Ming Xie c , Tony Liu d , Lyle H Ungar e , Nicholas C Cardamone c , Nathaniel J Williams f , Mark Olfson g Author information Article notes Copyright and License information a School of Social Policy & Practice, University of Pennsylvania, Philadelphia, PA b Department of Psychiatry, Yale University School of Medicine, New Haven, CT c Department of Psychiatry, Perelman School of Medicine, University of Pennsylvania, Philadelphia, PA d Department of Computer Science, Mount Holyoke College, South Hadley, MA e Department of Computer and Information Science, University of Pennsylvania, Philadelphia, PA f School of Social Work, Boise State University, Boise, ID g Department of Psychiatry, Vagelos College of Physicians and Surgeons, Columbia University and the New York State Psychiatric Institute, New York, NY, US ✉ Corresponding author. Sara Wiesel Cullen, PhD, MSW. School of Social Policy & Practice, University of Pennsylvania, Philadelphia, PA. swiesel@upenn.edu Issue date 2025 Jun. PMC Copyright notice PMCID: PMC13150847  NIHMSID: NIHMS2141438  PMID: 40318538 The publisher's version of this article is available at J Psychiatr Res Abstract Background. Patients presenting to emergency departments (EDs) for mental health problems have an elevated short-term risk of repeat ED visits, subsequent hospitalization, and suicide. Objective. Use health records to identify predictors of nonfatal suicidal or self-harm events following emergency department visits of individuals with mental health disorders. Methods. Electronic health record data from 2015–2022 were used to identify ED visits with mental health diagnoses for individuals 10 and older; extract 408 potential predictors including demographic, historical and baseline clinical characteristics from structured and unstructured data; and subsequent suicidal and self-harm events. We constructed a series of progressively additive logistic regression and gradient tree boosting (GTB) models to evaluate how groups of clinical features influenced likelihood of a nonfatal event 180-days after an ED visit. Results. Records identified 2,445,597 ED episodes and 176,000 subsequent suicidal or self-harm events within 180-days. Individuals experiencing an event relative to those without an event were less likely to be >65 (3.6% vs 16.7%; h =0.46), more likely to be male (55.7% vs 40.8%; h =0.30) and covered by Medicaid (61.5% vs 42.2%, h =0.39). The final model with 408 clinical features resulted in an AUC of 0.851 (logistic regression) and 0.863 (GTB). Diagnoses of bulimia nervosa (OR=1.84, p<.0001) and cutting (OR=2.62, p<.0001) were most highly associated with any subsequent event and suicidal self-harm, respectively. Conclusion. Machine learning algorithms effectively predicted nonfatal suicidal events within six months following ED visits among individuals with mental health disorders highlighting the importance of suicide symptom focused assessment and prevention efforts during routine emergency mental healthcare, particularly for patients with bulimia nervosa. Keywords: suicide prevention, emergency department, mental health INTRODUCTION 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, Franklin, Fox, et al., 2016 ). Although both suicidal self-harm and non-suicidal self-harm are associated with future suicide attempts ( Ribeiro, Franklin, Fox, 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, Franklin, Fox, et al., 2016 ; Whitlock & 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. METHODS This study adhered to the STROBE guidelines for cohort studies. Data Sources This study used EHR data from the Optum Labs Data Warehouse (OLDW) which contains longitudinal health information representing a range of ages and geographical regions across the U.S. for over 200 million commercial, Medicaid and Medicare Advantage enrollees ( Optum Labs, 2023 ). We received access to a deidentified clinical repository with EHR data, including medications prescribed, diagnoses, procedures, and clinical notes derived using Natural Language Processing (NLP; clinical narratives, patient-reported complaints, personal and family history, medication/treatment rationale, screening scores, and other assessments and discharge summaries). NLP is a computational linguistics technology that converts complex unstructured free-text clinical narratives into structured data to extract important clinical information from the unstructured text. Algorithms using NLP have been found to better predicted suicide and accidental death after hospital discharge than traditional structured data alone ( McCoy et al., 2016 ). A previously developed NLP algorithm, now integrated into the Optum Labs data warehouse, was utilized to identify and extract clinical concepts from unstructured free-text fields in the EHR. NLP algorithms automate the labor-intensive process of extraction and interpretation of narrative data ( Patra et al., 2024 ). The NLP pipeline includes multiple stages: tokenization to segment text into discrete units, syntactic parsing to identify phrase structures, and semantic processing to interpret clinical terms within context. Terms and attributes related to diagnoses, symptoms, and treatment are mapped through this process, leveraging both rule-based and machine-learning techniques to ensure accuracy. The algorithm produces structured output fields that classify terms as present, absent, or uncertain, based on contextual cues ( Green et al., 2019 ; Nunes et al., 2015 ). This refined algorithm facilitates the extraction of clinically relevant data from narrative text, supporting large-scale health services research and epidemiologic studies ( Nunes et al., 2016 ); extensive validation studies compared NLP-derived extractions against manually annotated gold-standard datasets, demonstrating high sensitivity and specificity for various clinical conditions. Including NLP data allowed us to capture mental health symptoms, substance use, and other clinical data related to suicide risk that may not be captured by ICD codes in claims data or by clinician-reported problem lists in the EHR ( Anderson et al., 2015 ; Arias et al., 2019 ; Davis et al., 2016). Sample From an initial sample of 5,553,428 ED visits for mental health disorders (excluding tobacco dependence) identified through EHR diagnosis data from September 30, 2015 to September 30, 2022, we excluded visits with a contiguous observation period prior to the index ED visit ( Figure 1 ). Additional exclusions included individuals under 10 years of age, a length of stay of ≥four days, no service use in the 180 days prior to index visit, episodes lacking NLP terms, and records with insufficient data defined as no data in their EHR for the 180 days post-index visit. The absence of data compromises our ability to discern whether patients truly had no subsequent care, were seen in a different health system, or were using different insurance that is not included in the Optum Labs database. We excluded children under 10 because suicidal symptoms are rare in this age group ( Nock et al., 2013 ), often present differently, and are assessed using different clinical approaches, making comparisons with older children less meaningful. We selected a length of stay of less than 4 days to focus on patients receiving acute emergency department (ED) care that reflect individuals who undergo assessment, stabilization, and disposition (either discharge or admission) within the standard ED workflow rather than patients with extended boarding in the ED. The first ED episode with a mental health disorder appearing in the data was counted as the index ED visit. If an individual had a claim on the same day or the day after discharge, it was grouped with and considered part of the index ED visit. Figure 1. Open in a new tab Study sample Outcomes Due to the relative infrequency of suicidal self-harm, both in general and in our dataset and to ensure sufficient power for our analyses, we combined suicidal self-harm with the more frequent self-harm and suicidal ideation events, similar to other studies that examine a broad continuum of suicidal outcomes ( Haroz et al., 2021 ). Diagnoses were derived from EHR-reported diagnoses and chart notes which reflect standard clinical documentation practices in EDs. Prior research has demonstrated that clinical mental health diagnoses, especially psychotic disorders, are of acceptable quality (Davis et al., 2016). The dependent variable, herein referred to as a suicidal or self-harm event, is any ED or inpatient visit with a diagnosis of nonfatal suicide attempt, self-harm, or suicidal ideation (ICD-10: X71–X83, T36–T50 with a 6th character of 2, T51–T65 with a 6th character of 2, T71 with a 6th character of 2, T14.91) ( Hedegaard et al., 2018 ), and R45.851 within 180 days of ED discharge. Predictors We identified a set of clinical characteristics as potential predictors of nonfatal suicidal and self-harm events based on a comprehensive review of prior research, which has identified specific domains and features associated with such outcomes. Diagnostic and NLP EHR data were used to derive a broad range of features, broken down to focus on the first two days of the episode, representative of the information available to clinicians when making care decisions, and the 180-days preceding the index visit to provide historical clinical context. Features ascertained during first two days of index hospitalization Demographic features account for substantial variation in suicide mortality ( Ivey-Stephenson et al., 2017 ). Categorical demographic characteristics were transformed into 10 binary variables based on age (10–17, 18–64, 65+), sex (M/F), race (white, black, Asian, other), ethnicity (Hispanic/non-Hispanic), and insurance (commercial/Medicaid/Medicare/uninsured/other), excluding a reference group for each variable. Mental health features or disorders , which are significant predictors of suicidal behavior ( Gili et al., 2019 ; Nock et al., 2010 ), were derived from diagnosis and NLP data within the EHR. After reviewing the diagnoses contained within the EHR records, 28 broad mental health diagnostic groups and 116 diagnostic sub-categories ( Supplemental Figure 1 ) were established, guided by the Clinical Classifications Software Refined (CCSR), a tool used to group mental health diagnoses into clinically meaningful categories. Within NLP data, 43 broad groups and 146 sub-categories were created by two clinical experts who grouped the text phrases with diagnostically related terms. The study team then coded the sentiments associated with each NLP diagnostic terms, excluding terms associated with a negative or indeterminate sentiment. All diagnoses and NLP features were binary variables to indicate the presence or absence of relevant mental health conditions. Physical health features , representing chronic conditions which have been associated with suicidal behavior ( Ahmedani et al., 2017 ; Scott et al., 2010 ), were derived from a review of the literature on chronic or acute health conditions associated with suicide risk including brain injury, HIV/AIDS, sleep disorders, migraine, epilepsy, congestive heart failure, COPD, back pain, and cancer. Specific ICD codes and their ICD-10 equivalents for these nine conditions ( Ahmedani et al., 2017 ) were used to create binary indicators for each diagnostic group. Another variable was created to capture a count of Elixhauser Comorbidity Index conditions, excluding duplicates with previously included diagnostic groups. Additional binary indicators were manually coded into 20 clinically homogenous groups representing the full spectrum of physical health conditions. Features ascertained in the 180-days preceding index hospitalization Prior mental health features. Prior mental and physical health disorders have been identified as predictive of future suicidal behavior ( Barak-Corren et al., 2017 ) and are thus included in the models. A set of historical mental health features using the same definitions as described above were created to identify the 28 broad mental health diagnoses and the 43 broad NLP mental health terms in the 180-days prior to admission. Prior physical health features. Similar procedures were used to identify the 10 physical health diagnoses (including Elixhauser conditions) and 20 NLP physical health groups. Prior mental health treatment features. Receipt of prior mental health care prior has been associated with effectiveness of ED interventions to prevent future suicidal behavior ( Cullen et al., 2024 ), as well as with suicidal behavior ( Bond et al., 2022 ). Four features characterized the history of prior service use associated with mental health care that might have been a precursor to the current ED visit. First, a variable representing the count of distinct generic medications prescribed known to have suicidal or depressive side effects ( Qato et al., 2018 ) was created ( Supplemental Figure 2 ). Second, we created a variable counting the number of days an individual experienced an ED visit with a mental health disorder (same definition as index event). Third, a variable counting the number of days an individual had a suicidal or self-harm episode (same definition as outcome) was created. Fourth, we quantified the number of days the patient received treatment in any setting (ED, inpatient or outpatient) for suicidal symptomology (diagnosis and NLP terms related to suicide, suicidal symptoms and suicidal behavior). Analysis Summary of analytic plan We compared two groups of patients: those who experienced a suicidal or self-harm event following an emergency department (ED) discharge and those who did not. First, we evaluated their demographic characteristics and prior mental health service use. Next, to predict the occurrence of consequent events, we developed several models that started with basic demographic features and progressively added more detailed clinical information, including diagnoses and treatment history. These models helped us assess the importance of different factors as predictors of these events. We also examined the performance of the models across various patient groups to ensure they worked well for all patients. To improve the predictions, we applied advanced statistical techniques that allowed us to better understand how different factors interact and contribute to the risk of suicidal or self-harm events. Sample characteristics Patient episodes involving a suicidal or self-harm event following ED discharge were compared to episodes without these events. This comparison was based on demographic characteristics, analyzed using chi-square tests, and prior mental health service use, analyzed using t-tests. Cohen’s h effect sizes calculated the differences in proportions across the groups for dichotomous variables while Cohen’s d was used to calculate difference in means across the groups for continuous variables. Suicidal or self-harm events following ED visit To investigate suicidal and self-harm events following an ED visit for mental health, we constructed a series of seven progressively additive logistic regression and gradient tree boosting (GTB) models. The primary objective was to assess how incorporating different feature sets impacted the model’s predictive performance, evaluated using the Area Under the Receiver Operating Characteristic Curve (AUC-ROC). The baseline model utilized only demographic features. Each subsequent model incrementally added feature sets to examine the impact of diagnoses and NLP terms describing clinical care that was received at the ED visit and in the 180-days preceding it. We used a 2/3 training set and 1/3 test set split to generate and evaluate all models, ensuring that each patient’s episodes were entirely contained within only one set. This approach helped maintain the integrity of performance metrics and ensured that the results accurately reflected the model’s generalizability. The baseline model (Model 1) was augmented by adding, in turn, diagnoses and NLP mental health features at index ED visit defined broadly and narrowly (Models 2–3). Then diagnosis and NLP physical health feature sets at index ED visit were added in while controlling for the narrowly defined mental health feature sets (Model 4). Subsequent models progressively added feature sets describing prior mental health features (Model 5), prior physical health features (Model 6), and prior mental health treatment features (Model 7). This approach allowed for the systematic evaluation of the contribution of each set of features, in the presence of already added features, to the model’s predictive power. Finally, to assess the specific contribution of NLP to the prediction models, we reran the final model using only structured EHR data (Model 7 without NLP) to allow for the evaluation of how well the model performed without NLP, knowing the challenges accompanying collecting NLP data. Each model’s performance was assessed by examining the AUC, a measure of how well the model distinguished between individuals who did and did not experience suicidal or self-harm events following an ED visit. For each model the increase in AUC is presented relative to the prior model to observe the incremental impact of adding each feature set. We also examined sensitivity, specificity, and positive predictive value (PPV) at optimal Youden cut points, which selects a cut point that provides the best trade-off between sensitivity and specificity ( Youden, 1950 ). We then trained a similar set of GTB models to capture complex interactions and non-linear relationships between the features to potentially improve model performance. Finally, we re-ran Model 7 (no NLP) with a larger sample of individuals who did not meet the exclusion criterion of needing any NLP term during the index episode to ensure that the results were comparable and not significantly impacted by this exclusion, responsible for substantial sample attrition. Stratifying the effects of demographics Using the final model with all features (Model 7), we evaluated how the AUC varied in models stratified by patient demographic characteristics and examined rates of suicidal/self-harm outcomes, along with the sensitivity and specificity for the models, within these demographic strata. Rank of predictors within feature sets To identify the most predictive sets of features for suicidal and self-harm events, we ranked features overall and within predefined clinical feature sets. First, odds ratios were calculated for all individual features with the magnitude used to rank these features. We then grouped the features into clinically relevant feature sets and determined the median rank of features within each group. This allowed us to rank entire feature sets. To complement this, we also used Shapley values, which quantify each feature’s contribution to the model’s predictions. We ranked the sum of the Shapley values for features within each set and, again, calculated the median rank for each group. Ranking by both odds ratios and Shapley values are presented because they reflect different aspects of the prediction model. Odds ratios are informative for rare but high-risk features, while Shapley values account for the frequency of the feature and its importance ( Liu & Ungar, 2021 ). Identifying predictive features In order to identify the most influential features in the final model, we calculated the odds ratios associated with suicidal and self-harm events following an ED visit for included variables. We examined the relative influence of these features on four separate outcomes representing the continuum of suicidal and self-harm events: any suicidal or self-harm event, suicidal self-harm, self-harm, and suicidal ideation. To account for spurious findings associated with infrequent events, we removed 42 features which occurred in the data less than 150 times. All analyses were conducted in scikit-learn, using the logistic regression and gradient tree boosting models ( Pedregosa et al., 2011 ). RESULTS The final sample size included 2,445,597 ED episodes, representing 1,341,253 individuals. There were a total of 176,000 suicidal and self-harm events in the 180 days post-discharge, representing 7.2% of the total sample. Demographic and prior treatment features Table 1 summarizes the demographic characteristics of individuals with and without any suicidal or self-harm event within 180 days post-discharge. Although differences in the demographic distribution were statistically significant between those with and without a follow-up event, because of the large sample size not all the statistically significant differences were clinically meaningful. Using a Cohen’s h effect size of ≥0.30 to denote clinical significance, important demographic characteristics included age greater than 65, sex, and Medicaid insurance status; notably, individuals experiencing a suicidal or self-harm event were less likely to be over 65 (3.6% vs. 16.7%; h =0.46), but more likely to be male (55.7% vs. 40.8%; h =0.30) and covered by Medicaid (61.5% vs. 42.2%; h =0.39) compared to those without an event. Table 1. Demographic and prior treatment features of individuals with and without any suicidal or self-harm event within 180-days after discharge from the ED Any suicidal or self-harm event in 180 days post discharge No suicidal or self-harm event in 180 days post discharge Effect size N % N % 176,000 7.2 2,269,597 92.8 Demographic features Age Group age 10–17 13534 7.7 124919 5.5 0.09 age 18–64 156215 88.7 1765522 77.8 0.30 age 65+ 6251 3.6 379156 16.7 0.46 Sex Female 77941 44.3 1343004 59.2 0.30 Male 98059 55.7 926593 40.8 0.30 Race White 125834 71.5 1683437 74.2 0.06 African American 38498 21.9 416030 18.3 0.09 Asian 705 0.4 14364 0.6 -0.03 Other/Unknown 10963 6.2 155766 6.9 -0.03 Hispanic Ethnicity 10990 6.2 153290 6.8 0.02 Insurance type Commercial 88348 50.2 1379207 60.8 0.21 Medicaid 108199 61.5 957082 42.2 0.39 Medicare 39174 22.3 581386 25.6 0.08 Medicare non-elderly 33758 19.2 259370 11.4 0.22 Uninsured 21393 12.2 269500 11.9 0.01 Other insurance 31417 17.9 385536 17.0 0.02 Mean Standard deviation Mean Standard deviation Effect size Mental health treatment preceding index event Mean count of ED days with MH/DSH/SI from DX 2.74 3.74 0.82 1.8 0.96 Mean count of acute ED/IP days with DSH/SI from DX 1.33 2.59 0.09 0.49 1.48 Mean count of ED/IP/OP days with DSH/SI from DX, NLP 7.38 10.55 2.68 5.34 0.80 Open in a new tab There were clinically meaningful differences with respect to service use in the 180-days preceding the index event. As compared to those without any suicidal or self-harm events, those with events had a higher mean number of mental health-related ED visits (2.74 vs 0.82; d =0.96); mean count of suicidal or self-harm events (1.33 vs 0.09; d=1.48); and mean count of days with suicidal symptomology in inpatient, outpatient or ED settings (7.38 vs 2.68; d =0.80). Feature set models Logistic regression and GTB models were progressively constructed, adding sets of variables, with each model’s AUC used to assess model performance. The baseline model, which included only 10 demographic variables, had an AUC of 0.677 ( Table 2 ). Adding the 28 mental health diagnosis variables increased the AUC to 0.777. The subsequent addition of the 146 NLP mental health variables produced only a slight increase in AUC to 0.806. The final model which added 224 variables to characterize current mental/physical health at the index visit and the period prior to admission (also prior mental health treatment), produced an AUC of 0.851. Overall, the incremental addition of features resulted in diminishing returns, with the AUC leveling off as it approached its maximum value in Model 7. The final logistic model achieved a sensitivity of 75.9%, a specificity of 78.2%, and a positive predictive value (PPV) of 21.3% at the Youden cutoff value. The model without any NLP data included 178 variables and returned an AUC (0.838) only 1.5% lower than the full 408-variable model. Table 2. Additive value of variable sets on predicting suicidal or self-harm events Feature Set (n=# of variables) Model 1 Model 2 Model 3 Model 4 Model 5 Model 6 Model 7 Model 7 (no NLP) Demographic features (n=10) x x x x x x x x Clinical feature sets during index ED visit Mental health features during Index ED Visit MH diagnoses [broad] (n=28) x MH diagnoses [narrow] (n=116) x x x x x x MH NLP Terms [broad] (n=43) MH NLP Terms [narrow] (n=146) x x x x x Physical health features during Index ED Visit PH Diagnoses** (n=10) x x x x x PH NLP Terms (n=20) x x x x Clinical feature sets 180-days prior to Index ED visit Mental health features in 180-days prior to ED MH diagnoses (broad)* (n=28) x x x x MH NLP Terms (broad)* (n=43) x x x Physical health features in 180-days prior to ED PH Diagnoses (n=10) x x x PH NLP Terms (n=20) x x Mental health treatment features in 180-days prior to ED Visit Med rx w suicidal or depressive side effects (n=2) x x Count of ED visits with MH/DSH/SI from DX (n=1) x x Count of acute ED/IP visits with DSH/SI from DX (n=1) x x Count of ED/IP/OP days with DSH/SI from DX or NLP (n=1) x Count of ED/IP/OP days with DSH/SI from DX (n=1) x Number of variables in model 10 38 272 302 373 403 408 178 AUC (logistic) 0.677 0.777 0.811 0.814 0.847 0.848 0.851 0.838 AUC % increase 14.77 4.38 0.37 4.05 0.12 0.35 −1.53 AUC (GTB) 0.689 0.786 0.818 0.82 0.861 0.862 0.863 0.853 AUC (GTB) % increase 14.08 4.07 0.24 5.00 0.12 0.12 −1.16 Open in a new tab The logistic regression models showed a steady increase in AUC with each additional feature set, highlighting the importance of comprehensive data inclusion. The GTB models consistently, but only slightly, improved model performance across all models by capturing complex interactions and non-linear relationships between the features. The final GTB model achieved an AUC of 0.863. Finally, when we ran the final model on the larger sample which did not exclude individuals with no NLP term during the index episode, the AUC was similar at 0.833. Examining the effects of demographics The prevalence of events was highest in youngest (ages 10–17) and lowest in the oldest (≥65 years) age groups ( Supplemental Figure 3 ). The AUC of the final predictive model was lower in the younger age group (0.771) than in the non-elderly adult (18–64 years, AUC=0.845) and older (≥65 years, AUC=0.862) age groups. The AUC was also lower among Hispanic (0.816) relative to non-Hispanic (0.856) patients. Ranking of predictors within feature sets Table 3 presents the three features most predictive of any suicidal and self-harm events after the index visit, within each clinical feature set and the final 366-feature model. The clinical feature sets with the best performance, indicated by the lowest median OR rank, were demographics, mental health treatment in the 180 days prior, and mental health diagnosis during the index visit. Within these sets, the top-ranking features were, respectively, Medicaid insurance, count of days with a suicidal or self-harm visit in prior 180-days, and bulimia nervosa. When calculating ranking using Shapley values, the top-ranking sets were mental health NLP, demographics, and mental health diagnosis during the index visit. Table 3. Top 3 predictors of any suicidal or self-harm event by clinical feature set Predictors N Feature set rank Feature Ranking ** (Overall rank out of 366 features) Odds ratios * Shapley values ** Highest ranked Second highest ranked Third highest ranked Demographics 10 1 2 Medicaid Insurance (10) Ages 10–17 (11) Medicare Insurance (20) Clinical Feature Sets During Index Visit Mental health features Mental health (Diagnosis) 76 3 3 Bulimia nervosa (1) Suicidal ideation (2) Inhalant-related conditions (4) Mental health (NLP) 144 7 1 Suicidal Ideation (3) Schizophrenia (7) Dependent Disorder (14) Physical health features Physical health (Diagnosis) 10 9 8 Epilepsy (55) HIV/AIDS (131) Sleep disorders (203) Physical health (NLP) 20 10 4 Infectious symptoms (107) NLPGRP_PH0_020-sensory (152) Skin and soft tissue disorders (210) Clinical Feature Sets 180-Days Prior to Index ED Visit Mental health features Mental health (Diagnosis) 28 4 10 Suicide and Self-Harm (96) Disruptive, impulse-control and conduct disorders (121) Depressive disorders (130) Mental health (NLP) 43 5 7 Stressor-Symptoms (115) Mood-Disorder (120) Lives-Alone (125) Physical health features Physical health (Diagnosis) 10 8 9 Count of ELIXHUASER (95) HIV/AIDS (141) Brain injury (163) Physical health (NLP) 20 6 6 Oncological conditions (153) Pain symptoms (155) Urinary symptoms (156) Mental health treatment features MH treatment variables 5 2 5 Count of days with an acute suicidal event (12) Count of days with an ED event (76) Count of prescriptions for medications with suicidal risk *** (82) Open in a new tab * Based on median of odds ratio ** Based on sum of Shapley values *** Medication prescribed within 60 days preceding index date Significant predictors of suicidal or self-harm events Table 4 identifies the 25 variables that had the most significant impact on likelihood of acute events in the 180-days following an ED visit for each of the outcomes on the continuum of suicidal events. The features most highly associated with any suicidal or self-harm event were bulimia nervosa diagnosis (OR=1.84, p<.0001), suicidal ideation identified in diagnosis data (OR=1.76, p<.0001) and suicidal ideation identified in NLP data (OR=1.62, p<.0001). Supplemental Figure 4 expands this table with all 408 features ranked for any suicidal and self-harm events. The features most highly associated with suicidal self-harm were diagnoses of self-harm cutting (OR=2.62, p<.0001), self-harm asphyxiation (OR=1.98, p<.0001), and bulimia nervosa (OR=1.85, p<.0001). The features most highly associated with self-harm were diagnoses of self-harm cutting (OR=2.53, p<.0001), an inhalant-related condition (OR=2.31, p<.0001), and non-psychoactive substance abuse (OR=2.13, NS). The features most highly associated with suicidal ideation were a suicidal ideation diagnosis (OR=1.85, p<.0001), suicidal ideational in NLP (OR=1.71, p<.0001), and schizophrenia in NLP (OR=1.55, p<.0001). Table 4. Top 25 clinical features with the highest odds ratios associated with any suicidal or self-harm event, suicidal self-harm, self-harm, and suicidal ideation 180-days after ED discharge Any suicidal or self-harm event a (7.2%) Suicidal self-harm b (1.00%) Self-harm c (0.84%) Suicidal ideation d (5.36%) Clinical Feature Odds Ratio Clinical Feature Odds Ratio Clinical Feature Odds Ratio Clinical Feature Odds Ratio Bulimia nervosa diagnosis (n=285) 1.84 ** Self-harm cutting diagnosis (n=2385) 2.62 ** Self-harm cutting diagnosis (n=2385) 2.53 ** Suicidal ideation diagnosis (n=51818) 1.85 ** Suicidal ideation diagnosis (n=51,818) 1.76 ** Self-harm asphyxiation diagnosis (n=195) 1.98 ** Inhalant-related condition diagnosis (n=314) 2.31 ** Suicidal ideation NLP (n=161749) 1.71 ** Suicidal Ideation NLP (n=161,749) 1.62 ** Bulimia nervosa diagnosis (n=285) 1.85 ** Non-psychoactive substance abuse diagnosis (n=159) 2.13 Schizophrenia NLP (n=57797) 1.55 ** Inhalant-related condition diagnosis (n=314) 1.55 ** Suicidal ideation NLP (n=161749) 1.83 ** Pica NLP (n=795) 2.03 ** Childhood trauma/stress disorder diagnosis (n=171) 1.46 Gender dysphoria/transsexualism diagnosis (n=487) 1.55 ** Self-harm poisoning diagnosis (n=7641) 1.81 ** Self-harm asphyxiation diagnosis (n=195) 1.97 * Dependent disorder NLP (n=255) 1.46 * Cutting, self-harm diagnosis (n=2,385) 1.50 ** Anxiolytic Use Disorder NLP (n=1133) 1.76 ** Poisoning diagnosis (n=5533) 1.91 ** Bulimia nervosa diagnosis (n=285) 1.43 Schizophrenia NLP (n=57,797) 1.47 ** Borderline Personality NLP (n=8893) 1.58 ** Bulimia nervosa diagnosis (n=285) 1.86 * Schizoaffective disorder diagnosis (n=15934) 1.40 ** Schizoaffective diagnosis (n=15,934) 1.42 ** Inhalant-related conditions diagnosis (n=314) 1.58 Other undetermined suicide and self-harm diagnosis (n=446) 1.84 Neurotic symptoms NLP (n=994) 1.39 ** Anorexia diagnosis (n=321) 1.41 * Age 10–17 1.55 ** Self-harm poisoning diagnosis (n=7641) 1.83 ** Male gender (n=307421) 1.39 ** Medicaid Insurance (n=319,847) 1.39 ** Gender dysphoria/transsexualism (n=487) 1.53 ** Age 10–17 (n=41067) 1.67 ** Medicaid insurance (n=319487) 1.38 ** Age 10–17 (n=41,067) 1.38 ** Suicidal ideation diagnosis (n=51818) 1.53 ** Impulse control disorder NLP (n=1366) 1.57 ** Anorexia diagnosis (n=321) 1.36 Acute suicidal event diagnosis in past 180 days (n=128,526) 1.36 ** Dissociative/conversion disorder diagnosis (n=3335) 1.52 ** Other eating disorder diagnosis (n=917) 1.48 Homeless NLP (n=33127) 1.33 ** Non-psychoactive substance abuse diagnosis (n=159) 1.36 Self-harm thoughts NLP (n=3111) 1.52 ** Other specified-substance use condition diagnosis (n=40728) 1.42 ** Somatization diagnosis (n=240) 1.32 Dependent personality disorder NLP (n=255) 1.36 Violent thoughts NLP (n=413) 1.50 * Dependent Disorder NLP (n=255) 1.40 Gender dysphoria/ transsexualism diagnosis (n=487) 1.31 Borderline Personality Disorder NLP (n=8,893) 1.35 ** Other disruptive, impulse-control disorder diagnosis (n=969) 1.50 ** Hyperventilation NLP (n=3822) 1.38 * Medicare Insurance (n=186720) 1.30 ** Dissociative/conversion disorder diagnosis (n=3,335) 1.34 Pica NLP (n=795) 1.49 Overdose NLP (n=38923) 1.38 ** Substance-induced anxiety condition diagnosis (n=616) 1.29 * Poisoning, undetermined diagnosis (n=5,533) 1.34 Anorexia diagnosis (n=321) 1.48 Dissociative/conversion disorder diagnosis (n=3335) 1.36 * Age 10–17 (n=41067) 1.28 ** Self-harm Thoughts NLP (n=3,111) 1.31 Poisoning diagnosis (n=5533) 1.47 ** Poor defined mental disorder diagnosis (n=6079) 1.35 ** Borderline Personality NLP (n=8893) 1.28 ** Poor defined mental disorder diagnosis (n=6,079) 1.29 Borderline personality disorder diagnosis (n=5433) 1.46 ** Opioid-related conditions diagnosis (n=28712) 1.34 ** Bipolar Disorder NLP (n=113809) 1.26 ** Medicare Insurance (n=186,720) 1.28 Medicaid insurance (n=319847) 1.43 ** Induced Psychosis NLP (n=1521) 1.33 Schizophrenia disorder diagnosis (n=27056) 1.26 ** Stimulant-related condition diagnosis (n=40,387) 1.27 ** Non-psychoactive substance abuse (n=159) 1.42 Other disruptive, impulse-control disorder diagnosis (n=969) 1.30 Stimulant-related condition diagnosis (n=40387) 1.26 ** Impulse Control Disorder NLP (n=1,366) 1.27 ** Bipolar symptoms diagnosis (n=66640) 1.40 ** Medicaid Insurance (n=319847) 1.29 ** Other personality disorder diagnosis (n=4368) 1.26 ** Other eating disorder diagnosis (n=917) 1.27 * Caucasian race (n=541992) 1.40 ** Bipolar Disorder NLP (n=113809) 1.29 ** Poor defined mental disorder diagnosis (n=6079) 1.24 ** Female (n=425,899) 1.26 ** Other eating disorder diagnosis (n=917) 1.31 Mental Status Change NLP (n=41821) 1.26 ** Dissociative/conversion disorder diagnosis (n=3335) 1.24 ** Schizophrenia diagnosis (n=27,056) 1.25 ** Somatization symptoms NLP (n=6612) 1.30 ** Somatization Symptoms NLP (n=6612) 1.26 * 050-Depressed Mood NLP (n=352982) 1.24 ** Open in a new tab Excludes features with a frequency count of less than 150 a AUC=0.851 b AUC=0.866 c AUC=0.751 d AUC=0.847 * p-value is significant at <.01 ** p-value is significant at <.001 DISCUSSION This study examined factors associated with suicidal and self-harm events within 180 days of discharge from the ED, focusing on demographic, service use, and mental health data. Key findings include that patient age, sex, and Medicaid status were clinically meaningful predictors, with younger individuals, males, and those with Medicaid more likely to experience these events. Higher rates of prior mental health-related ED visits and suicidal symptoms were also associated with these events. Our predictive models that included demographic, clinical, and NLP-derived features showed strong performance. Overall, the study highlights the importance of using a wide range of data to improve predictions and clinical understanding of suicidal and self-harm events. These findings complement recent work examining the performance of machine learning models in predicting suicide death following ED visits for mental health diagnoses ( Simon et al., 2024 ) by using these same methods to predict nonfatal outcomes, which are themselves independent predictors of future suicidal behavior ( Ribeiro, Franklin, Got, et al., 2016 ; Whitlock & Knox, 2007 ). Together, this research addresses the continuum of suicide-related behavior occurring after a critical point in help-seeking (ED episode) and harnesses readily available data to inform clinical decision-making. Our prediction model using a wide range of variables derived from structured and free text in electronic health records had an overall area under the curve (AUC) of 0.851 for nonfatal suicidal and self-harm events during the 6 months following a mental health ED visit. An AUC of 0.8 to 0.9 is generally considered excellent ( Hosmer & Lemeshow, 2000 ). The mental health features from the index ED visit substantially increased the AUC above and beyond what was achieved with demographic characteristics alone. Bulimia nervosa diagnosis, self-harm cutting diagnoses and suicidal ideation during the ED visit were strong independent predictors of suicidal and self-harm events over the following six months. The classification performance of our final model compares favorably with previous efforts to predict nonfatal suicidal events after ED visits with only structured EHR data. One study, limited to mental health ED patients aged 8–18 years, reported an AUC of 0.775 for a broad continuum of suicidal behavior, including suicide ideation and attempts at three months ( Haroz et al., 2021 ). Another analysis of structured EHR data of patients presenting to psychiatric ED services of a major academic medical center had an AUC of 0.65 for six-month suicide attempts ( Nock et al., 2022 ). Our final model also outperformed prediction of nonfatal suicide attempts with standardized assessment measures. For example, a UK study of four risk scales administered to patients presenting to EDs with self-harm reported six-month AUCs for repeat self-harm ranging from 0.49 (SAD PERSONS scale) to 0.71 (Re-ACT Self-Harm Rule) ( Steeg et al., 2018 ). Another study found that the Columbia Suicide Severity Rating Scale achieved an AUC of 0.589 for nonfatal suicide attempts during the 12-months following adult ED visits with suicidal symptoms ( Brown et al., 2020 ). Still, this research adds to the recent literature using machine learning techniques to improve the understanding and prediction of suicidal thoughts and behaviors ( Somé et al., 2024 ) which is relevant to and has been studied by a global audience of emergency and healthcare providers ( Huang et al., 2022 ; Nock et al., 2008 ). The final model was developed for suicidal and self-harm event prediction of mental health patients presenting to EDs where clinicians may not be experienced at or staffed for suicide risk assessment. Approximately one in five (20.7%) US EDs do not have mental health specialists consistently available to evaluate their patients ( Cullen et al., 2020 ). In our analysis, mental health ED patients were at high risk of suicidal or self-harm events following their visits. During the 180-day follow-up period, approximately 7.2% had one or more event resulting in ED or hospital admission. As expected, however, this event rate is lower than that of patients with specific indications of suicide risk at the index ED visit. For example, one-fifth (20%) of a Swedish cohort of adult ED patients with deliberate self-harm had repeat episodes over the following 6 months ( Bilén et al., 2011 ), and 9.2% of a Danish cohort (ages ≥10) had a subsequent suicide attempt following an index ED suicide attempt visit ( Fedyszyn et al., 2016 ). Nevertheless, our finding that roughly one in fourteen patients presenting for ED mental healthcare had suicidal or self-harm related ED visits or hospital admissions over the following six months underscores the importance of suicide symptom focused assessment and prevention efforts during routine emergency mental healthcare. An ED diagnosis of bulimia nervosa, while relatively uncommon, emerged as the single strongest predictor of any suicidal or self-harm event. Suicidal behavior has long been recognized as a clinical concern in people with eating disorders. Individuals with bulimia nervosa have higher rates of suicide attempts and self-harm compared to those with restrictive anorexia nervosa ( Cucchi et al., 2016 ; Mandelli et al., 2019 ). In people with bulimia nervosa, emotional dysregulation and affective intensity may play mediating roles in the development of suicidal symptoms and self-harm behaviors ( Gordon et al., 2016 ). The strong correlation between bulimia nervosa and subsequent suicidal and self-harm events highlights the need for careful suicide risk assessment of patients with bulimia nervosa. Self-harm related cutting was the most significant predictor of subsequent suicidal and other self-harm in our study, notable because even though self-harm behaviors are often not considered suicidal gestures in and of themselves, other studies have still found them strongly predictive of future suicidality, including suicide attempts ( Whitlock & Knox, 2007 ). Among individuals with repeated self-harm episodes, one-third switched method of harm between episodes, further suggesting that self-harm may be an unpredictable behavior ( Owens et al., 2015 ) that should trigger suicide assessment. Suicidal ideation coded at the time of ED visit, as expected, was also a leading predictor of suicide-related events during follow-up. Suicide ideation noted in text, more than three times as common as coded suicidal ideation, was associated with similar adjusted odds of suicide-related events. The substantially higher prevalence of written notations than coding of suicidal symptoms reinforces prior research that ICD codes for suicidal behavior are frequently incomplete, capturing only a fraction of those with suicidal symptoms ( Stanley et al., 2018 ; Zhong et al., 2019 ). Despite the importance of this text feature, we found that including features derived from free text only modestly increased the overall prediction of nonfatal suicidal events above and beyond prediction based only on structured features. Limitations. Although this dataset included a diverse range of publicly insured, commercially insured, and uninsured patients, the performance of prediction models may not generalize to other patient populations, other healthcare systems or other countries that do not have electronic health records. Models also do not necessarily apply to the prediction of death by suicide, which is a distinct phenomenon with predictors distinct from the more common suicidal ideation and nonfatal self-harm ( Klonsky et al., 2016 ), included to study a wide range of suicidal behavior. Further, given that more than half of index (first) suicide attempts in the US are fatal ( Bostwick et al., 2016 ), the results of this study of nonfatal events following mental health emergency department visits are not broadly applicable to the clinical or public health prevention of suicide. Future research needs to identify ways to assess risk in the community with this population prior to this first attempt. However, since most nonfatal events are followed by another often more serious suicidal event in the subsequent year ( Bostwick et al., 2016 ; Nock et al., 2008 ), people who have nonfatal events remain a high-risk group in need of intervention. Since some individuals who make suicide attempts do not have mental disorders ( Oquendo et al., 2024 ), they may not be captured in this analysis of ED mental health visits. Our analysis of outcomes also did not include those who engaged in suicidal or self-harm behavior that did not result in contact with the healthcare system as they would be undetected in the electronic health records. The strength of associations of NLP-based variables may vary based on the specific processes for encoding this data. Although NLP data were incorporated to include additional information beyond what providers document directly in claims, we acknowledge that because clinician’s notes may be brief and focused only on the presenting condition, the EHR may not contain a comprehensive assessment of a patient’s clinical functioning. Uncertainty also exists concerning the validity and reliability of clinical diagnoses. Finally, the performance of our models was not compared with clinical measures of suicide risk assessment. Future research could also consider comparing GTB with other machine learning and non-machine learning analyses as there are limitations to all statistical techniques. We further acknowledge some common limitations with machine learning models include data dependency and overfitting, challenges to model interpretability, and its atheoretical nature. A prediction algorithm using structured and free text data from EHR demonstrated accurate prediction of suicide-related events over a six-month period following mental health ED visits. Taking account of mental health characteristics during the six months before the ED visit substantially increased the accuracy of the model predicting nonfatal suicidal and self-harm events over the six months after the visit. In the future, such a data-driven approach could help frontline clinicians improve their identification of patients at elevated risk and inform selection of timely mental health interventions ( Doupnik et al., 2020 ). A key next step in the development of such a technology involves creating an EHR-driven clinical decision support tool that generates brief, easy-to-read, and readily interpretable reports for ED clinicians to review in real-time predicting both nonfatal and fatal suicidal events. Supplementary Material Supplementary Files NIHMS2141438-supplement-Supplementary_Files.docx (119.9KB, docx) Funding/support: This research was funded by the National Institute of Mental Health, National Institutes of Health, grant R01MH126895. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health. Role of the sponsor: The sponsors had no role in the design, analysis, interpretation, or publication of this study. Footnotes Conflict of interest statement : The authors have no conflicts of interest to report. REFERENCES Ahmedani B, Westphal J, Autio K, Elsiss F, Peterson E, Beck A, Waitzfelder B, Rossom R, Owen-Smith A, Lynch F, Lu C, Frank C, Prabhakar D, Braciszewski J, Miller-Matero L, Yeh H, Hu Y, Doshi R, Waring S, & Simon G (2019). Variation in patterns of health care before suicide: A population case-control study. Preventive Medicine, 127, 105796. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ahmedani BK, Peterson EL, Hu Y, Rossom RC, Lynch F, Lu CY, Waitzfelder BE, Owen-Smith AA, Hubley S, Prabhakar D, Williams LK, Zeld N, Mutter E, Beck A, Tolsma D, & Simon GE (2017). Major Physical Health Conditions and Risk of Suicide. American Journal of Preventive Medicine, 53(3), 308–315. 10.1016/j.amepre.2017.04.001 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Anderson H, Pace W, Brandt E, Nielsen R, Allen R, Libby A, West D, & Valuck R (2015). Monitoring suicidal patients in primary care using electronic health records. J Am Board Fam Med, 28(1), 65–71. 10.3122/jabfm.2015.01.140181 [ DOI ] [ PubMed ] [ Google Scholar ] Arias S, Boudreaux E, Chen E, Miller I, Camargo CJ, Jones R, & Uebelacker L (2019). Which chart elements accurately identify emergency department visits for suicidal ideation or behavior? Archives of Suicide Research, 23(3), 382–390. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Barak-Corren Y, Castro V, Javitt S, Hoffnagle A, Dai Y, Perlis R, Nock M, Smoller J, & Reis B (2017). Predicting suicidal behavior from longitudinal electronic health records. American Journal of Psychiatry, 174(2), 154–162. 10.1176/appi.ajp.2016.16010077 [ DOI ] [ PubMed ] [ Google Scholar ] Bilén K, Ottosson C, Castrén M, Ponzer S, Ursing C, Ranta P, Ekdahl K, & Pettersson H (2011). Deliberate self-harm patients in the emergency department: factors associated with repeated self-harm among 1524 patients. Emerg Med J, 28, 1019–1025. [ DOI ] [ PubMed ] [ Google Scholar ] Bond A, Bandel S, Rodriguez T, Anestis J, & Anestis M (2022). Mental Health Treatment Seeking and History of Suicidal Thoughts Among Suicide Decedents by Mechanism, 2003–2018. JAMA Network Open, 5(3), e222101. 10.1001/jamanetworkopen.2022.2101 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bostwick J, Pabbati C, Geske J, & McKean A (2016). Suicide Attempt as a Risk Factor for Completed Suicide: Even More Lethal Than We Knew. Am J Psychiatry, 173(11), 1094–1100. 10.1176/appi.ajp.2016.15070854 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Britton P, Ilgen M, Rudd M, & Conner K (2012). Warning signs for suicide within a week of healthcare contact in Veteran decedents. Psychiatry Research, 200(2–3), 395–399. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Brown L, Boudreaux E, Arias S, IW M, May A, Camargo CJ, Bryan C, & Armey M (2020). C- SSRS performance in emergency department patients at high risk for suicide. 2020;50(6):. . Suicide Life Threat Behav, 50(6), 1097–1104. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Bruffaerts R, Kessler R, Demyttenaere K, Bonnewyn A, & Nock M (2015). Examination of the population attributable risk of different risk factor domains for suicidal thoughts and behaviors. Journal of Affective Disorders, 187, 66–72. 10.1016/j.jad.2015.07.042 [ DOI ] [ PubMed ] [ Google Scholar ] Burke T, Ammerman B, & Jacobucci R (2019). The use of machine learning in the study of suicidal and non-suicidal self-injurious thoughts and behaviors: A systematic review. Journal of Affective Disorders, 245, 869–884. 10.1016/j.jad.2018.11.073 [ DOI ] [ PubMed ] [ Google Scholar ] Burke T, Ammerman B, & Jacobucci R (2019). The use of machine learning in the study of suicidal and non-suicidal self-injurious thoughts and behaviors: A systematic review. Journal of Affective Disorders, 245(Feb 15), 869–884. 10.1016/j.jad.2018.11.073 [ DOI ] [ PubMed ] [ Google Scholar ] Burstein B, Agostino H, & Greenfield B (2019). Suicidal attempts and ideation among children and adolescents in US emergency departments, 2007–2015. JAMA Pediatrics, 173(6), 598–600. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Chang G, Weiss A, Orav E, & Rauch S (2014). Predictors of frequent emergency department use among patients with psychiatric illness. Gen Hosp Psychiatry, 36(6), 716–720. 10.1016/j.genhosppsych.2014.09.010 [ DOI ] [ PubMed ] [ Google Scholar ] Choi S, Lee W, Yoon J-H, Won J-U, & Kim D (2018). Ten-year prediction of suicide death using Cox regression and machine learning in a nationwide retrospective cohort study in South Korea. Journal of Affective Disorders, 231, 8–14. 10.1016/j.jad.2018.01.019 [ DOI ] [ PubMed ] [ Google Scholar ] Connery H, Taghian N, Kim J, Griffin M, Rockett I, Weiss R, & McHugh R (2019). Suicidal motivations reported by opioid overdose survivors: A cross-sectional study of adults with opioid use disorder. Drug and Alcohol Dependence, 205, 107612. 10.1016/j.drugalcdep.2019.107612 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Crump C, Sundquist K, Sundquist J, & Winkleby M (2014). Sociodemographic, psychiatric and somatic risk factors for suicide: A Swedish national cohort study. Psychological Medicine, 44(2), 279–289. 10.1017/S0033291713000810 [ DOI ] [ PubMed ] [ Google Scholar ] Cucchi A, Ryan D, Konstantakopoulos G, Stroumpa S, Kaçar A, Renshaw S, Landau S, & Kravariti E (2016). Lifetime prevalence of non- suicidal self-injury in patients with eating disorders: a systematic review and meta-analysis. Psychol Med, 46, 1345–1358. [ DOI ] [ PubMed ] [ Google Scholar ] Cullen S, Diana A, Olfson M, Xie M, & Marcus S (2020). Impact of around the clock mental health staffing may improve the emergency department management of patients who deliberately self-harm. Psychiatric Serv, 9, 913–919. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Cullen S, Olfson M, Xie M, & Marcus S (2024). Impact of Emergency Department Safety Planning on 30-Day Mental Health Service Use. Psychiatric Services, 75(6), 528–533. 10.1176/appi.ps.20230187 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Davis K, Sudlow C, & Hotopf M (2016). Can mental health diagnoses in administrative data be used for research? A systematic review of the accuracy of routinely collected diagnoses. BMC Psychiatry, 16, 2. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Davis K, Sudlow C, & Hotopf M (2016). Can mental health diagnoses in administrative data be used for research? A systematic review of the accuracy of routinely collected diagnoses. BMC Psychiatry, 16(1), 263. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] DelPozo-Banos M, John A, Petkov N, Berridge D, Southern K, LLoyd K, Jones C, Spencer S, & Travieso C (2018). Using neural networks with routine health records to identify suicide risk: Feasibility study JMIR Mental Health, 5(2), e10144. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Doupnik S, Rudd B, Schmutte T, Worsely D, McCarthy E, Eggan E, Bridge J, & Marcus S (2020). Association of suicide prevention interventions with subsequent suicide attempts, linkage to follow-up care, and depression symptoms for acute care settings. JAMA Psychiatry, 77(10), 1021–1030. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Fedyszyn I, Erlangsen A, Hjorthoj C, Madsen T, & Nordentoft M (2016). Repeated suicide attempts and suicide among individuals with a first emergency department Contact for attempted suicide: a prospective, nationwide, Danish register-based study J Clin Psychiatry, 77(6), 832–840. [ DOI ] [ PubMed ] [ Google Scholar ] Galfalvy H, Oquendo M, & Mann J (2008). Evaluation of clinical prognostic models for suicide attempts after a major depressive episode. Acta Psychiatrica Scandinavica, 117(4), 244–252. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Gili M, Castellvi P, Vives M, de la Torre-Luque A, Almenara J, Blasco M, Cebria A, & Gabilondo A (2019). Mental disorders as risk factors for suicidal behavior in young people: A meta-analysis and systematic review of longitudinal studies. Journal of Affective Disorders, 245, 152–162. [ DOI ] [ PubMed ] [ Google Scholar ] Gordon K, Simonich H, Wonderlich S, Dhankikar S, Crosby R, Cao L, Kwan M, Mitchell J, & Engel S (2016). Emotion dysregulation and affective intensity mediate the relationship between childhood abuse and suicide-related behaviors among women with bulimia nervosa. Sui Life Threaten Behav, 46(1), 79–87. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Green C, Perrin N, Hazlehurst B, Janoff S, DeVeaugh-Geiss A, Carrell D, Grijalva C, & Liang C (2019). Identifying and classifying opioid-related overdoses: A validation study. Pharmacoepidemiological Drug Safety, 1127–1137. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Guidotti R, Monreale A, Ruggieri S, Turini F, Giannotti F, & Pedreschi D (2018). A survey of methods for explaining black box models. ACM Computing Surveys (CSUR), 51(5), 1–42. [ Google Scholar ] Haroz E, Kitche C, Nestadt P, Wilcox H, DeVylder J, & Kharrazi H (2021). Comparing the predictive value of screening to the use of electronic health record data for detecting future suicidal thoughts and behavior in an urban pediatric emergency department: A preliminary analysis. Sui Life Threaten Behav, 51, 1189–1202. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Hedegaard H, Schoenbaum M, Claassen C, Crosby A, Holland K, & Proescholdbell S (2018). Issues in developing a surveillance case definition for nonfatal suicide attempt and intentional self-harm using International Classification of Diseases, Tenth Revision, Clinical Modification (ICD-10-CM) coded data (National Health Statistics Reports, Issue. [ PubMed ] [ Google Scholar ] Hoertel N, Franco S, Wall M, Oquendo M, Kerridge B, Limosin F, & Blanco C (2015). Mental disorders and risk of suicide attempt: A national prospective study. Molecular Psychiatry, 20(6), 718–726. 10.1038/mp.2015.19 [ DOI ] [ PubMed ] [ Google Scholar ] Hosmer D, & Lemeshow S (2000). Applied Logistic Regression. In (2nd ed., pp. 160–164). John Wiley and Sons. [ Google Scholar ] Huang Y, Zhu C, Feng Y, Ji Y, Song J, Wang K, & Yu F (2022). Comparison of three machine learning models to predict suicidal ideation and depression among Chinese adolescents: A cross-sectional study. Journal of Affective Disorders, 15(December), 221–228. [ DOI ] [ PubMed ] [ Google Scholar ] Ivey-Stephenson A, Crosby A, Jack S, Haileyesus T, & Kresnow-Sedacca M-J (2017). Suicide trends among and within urbanization levels by sex, race/ethnicity, age group, and mechanism of death—United States, 2001–2015. MMWR Surveillance Summaries, 66(18), 1–16. 10.15585/mmwr.ss6618a1 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Karmakar C, Luo W, Tran T, Berk M, & Venkatesh S (2016). Predicting risk of suicide attempt using history of physical illnesses from electronic medical records. JMIR Mental Health, 3(3), e19. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kessler R, Chiu W, Demler O, & Walters E (2005). Prevalence, severity, and comorbidity of 12-month DSM-IV disorders in the National Comorbidity Survey Replication. Archives of General Psychiatry, 62(6), 617–627. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Kessler R, Warner C, Ivany C, Petukhova M, Rose S, Bromet E, Brown M. r., Cai T, & Colpe L (2015). Predicting suicides after psychiatric hospitalization in US Army soldiers: The Army Study to Assess Risk and Resilience in Servicemembers (Army STARRS). JAMA Psychiatry, 72(1), 49–57. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Klonsky E, May A, & Saffer B (2016). Suicide, suicide attempts, and suicide ideation12. Ann Rev Psychol, 12, 307–330. [ DOI ] [ PubMed ] [ Google Scholar ] Liu T, & Ungar L (2021). Towards Cotenable and Causal Shapley Feature Explanations In AAAI 2021 Workshop: Trustworthy AI for Healthcare, [ Google Scholar ] Mandelli L, Arminio A, Atti A, & Ronchi D (2019). Suicide attempts in eating disorder subtypes: a meta-analysis of the literature employing DSM-IV, DSM-5, or ICD-10 diagnostic criteria. Psychol Med, 49, 1237–1249. [ DOI ] [ PubMed ] [ Google Scholar ] McCarthy J, Bossarte R, Katz I, Thompson C, Kemp J, Hannemann C, Nielson C, & Schoenbaum M (2015). Predictive modeling and concentration of the risk of suicide: Implications for preventive interventions in the US Department of Veterans Affairs. American Journal of Public Health, 105(9), 1935–1942. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] McCoy TJ, Castro V, Roberson A, Snapper L, & Perlis R (2016). Improving prediction of suicide and accidental death after discharge from general hospitals with natural language processing JAMA Psychiatry, 73, 1064–1071. 10.1001/jamapsychiatry.2016.2172 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nock M, Borges G, Bromet E, Alonso J, Angermeyer M, Beautrais A, Bruffaerts R, Chiu W, de Girolamo G, Gluzman S, de Graaf R, Gureje O, Haro J, Huang Y, Karam E, Kessler R, Lepine J, Levinson D, Medina-Mora M,…Williams D (2008). Cross-national prevalence and risk factors for suicidal ideation, plans and attempts. Br J Psychiatry, 192(2), 98–105. 10.1192/bjp.bp.107.040113 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nock M, Green J, Hwang I, McLaughlin K, Sampson N, Zaslavsky A, & Kessler R (2013). Prevalence, correlates, and treatment of lifetime suicidal behavior among adolescents: Results from the National Comorbidity Survey Replication Adolescent Supplement (NCS-A). JAMA Psych, 70(3), 300–310. 10.1001/2013.jamapsychiatry.55. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nock M, Hwang I, Sampson N, & Kessler R (2010). Mental disorders, comorbidity and suicidal behavior: Results from the National Comorbidity Survey Replication. Molecular Psychiatry, 15(8), 868–876. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nock M, Millner A, Ross E, Kennedy C, Al-Suwaidi M, Barak-Corren Y, Castro V, Castro-Ramirez F, Lauricella T, Murman N, Petukhova M, Bird S, Reis B, Smoller J, & Kessler R (2022). Prediction of suicide attempts using clinician assessment, patient self-report, and electronic health records. JAMA Network Open, 5(1). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Nordentoft M, Mortensen P, & Pedersen C (2011). Absolute risk of suicide after first hospital contact in mental disorder. Archives of General Psychiatry, 68(10), 1058–1064. 10.1001/archgenpsychiatry.2011.113 [ DOI ] [ PubMed ] [ Google Scholar ] Nordin N, Zainol Z, Noor M, & Chan F (2022). Suicidal behaviour prediction models using machine learning techniques: A systematic review. Artificial Intelligence in Medicine, 132. 10.1016/j.artmed.2022.102395 [ DOI ] [ PubMed ] [ Google Scholar ] Nunes A, Mortimer K, Loughlin J, Wang F, & Dore D (2015, August 22–26, 2015). Refinement of a Generalized Natural Language Processing Algorithm for the Identification of Clinical Terms from Free-Text Clinical Notes 31st International Conference on Pharmacoepidemiology and Therapeutic Risk Management, Boston, MA. [ Google Scholar ] Nunes A, Yang J, L R, Engel S, Kurtyka K, & Tunceli K (2016). Assessing occurrence of hypoglycemia and its severity from electronic health records of patients with type 2 diabetes mellitus. Diabetes Research and Clinical Practice, 121, 192–203. [ DOI ] [ PubMed ] [ Google Scholar ] Olfson M, Gao Y, Xie M, Cullen S, & Marcus S (2021). Suicide Risk Among Adults With Mental Health Emergency Department Visits With and Without Suicidal Symptoms. J Clin Psychiatry, 82(6), 20m13833. 10.4088/JCP.20m13833 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Optum Labs. (2023). Optum Labs and Optum Labs Data Warehouse (OLDW) Descriptions and Citation [Reproduced with permission from Optum Labs; ]. [ Google Scholar ] Oquendo M, Wall M, Wang S, Olfson M, & Blanco C (2024). Lifetime suicide attempts in healthy volunteers: results from a national epidemiological study. JAMA Psychiatry 81(6), 572–578. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Owens D, Kelley R, Munyombwe T, Bergen H, Hawton K, Cooper J, & Ness J (2015). Switching methods of self-harm at repeat episodes: Findings from a multicentre cohort study. J Affect Disord, 180, 44–51. [ DOI ] [ PubMed ] [ Google Scholar ] Owens P, Fingar K, Heslin K, Mutter R, & Booth C (2017). Emergency department visits related to suicidal ideation, 2006–2013. Rockville, MD: [ PubMed ] [ Google Scholar ] Patra B, Lepow L, Kumar P, Vekaria V, Sharma M, Adekkanattu P, Fennessy B, Hynes G, Landi I, Sanchez-Ruiz J, Ryu E, Biernacka J, Nadkarni G, Talati A, Weissman M, Olfson M, Mann J, Zhang Y, Charney A, & Pathak J (2024). Extracting social support and social isolation information from clinical psychiatry notes: comparing a rule-based natural language processing system and a large language model Journal of the American Medical Informatics Association, ocae260. 10.1093/jamia/ocae260 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Pedregosa F, Varoquaux G, Gramfort A, Michel V, Thirion B, Grisel O, Blondel M, Prettenhofer P, Weiss R, Dubourg V, Vanderplas J, Passos A, Cournapeau D, Brucher M, Perrot M, & Duchesnay E (2011). Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 12, 2825–2830. [ Google Scholar ] Penzenstadler L, Gentil L, Grenier G, Khazaal Y, & Fleury M (2020). Risk factors of hospitalization for any medical condition among patients with prior emergency department visits for mental health conditions. BMC Psychiatry, 20, 431–445. 10.1186/s12888-020-02835-2 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Peyre H, Hoertel N, Stordeur C, Lebeau G, Blanco C, McMahon K, Basmaci R, Lemogne C, Limosin F, & Delorme R (2017). Contributing factors and mental health outcomes of first suicide attempt during childhood and adolescence: Results from a nationally representative study. The Journal of Clinical Psychiatry, 78(06), e622–e630. 10.4088/JCP.16m10876 [ DOI ] [ PubMed ] [ Google Scholar ] Pigoni A, Delvecchio G, Turtulici N, Madonna D, Pietrini P, Cecchetti L, & Brambilla P (2024). Machine learning and the prediction of suicide in psychiatric populations: a systematic review. Translational Psychiatry, 14(1), 140. 10.1038/s41398-024-02852-9 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Plemmons G, Hall M, Doupnik S, Gay J, Brown C, Browning W, Casey R, Freundlich K, Johnson D, Lind C, Rehm K, Thomas S, & Williams D (2018). Hospitalization for suicide ideation or attempt: 2008–2015 Pediatrics, 141(6), e2017–2426. 10.1542/peds.2017-2426 [ DOI ] [ PubMed ] [ Google Scholar ] Qato D, Ozenberger K, & Olfson M (2018). Prevalence of prescription medications with depression as a potential adverse effect among adults in the United States. JAMA, 319(22), 2289–2298. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ribeiro J, Franklin J, Fox K, Bentley K, Kleiman E, Chang B, & Nock M (2016). Self-injurious thoughts and behaviors as risk factors for future suicide ideation, attempts, and death: a meta-analysis of longitudinal studies. Psychological Medicine, 46(2), 225–236. 10.1017/S0033291715001804 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Ribeiro J, Franklin J, Got K, & Bentley K (2016). Self-injurious thoughts and behaviors as risk factors for future suicide ideation, attempts, and death: a meta-analysis. Psychol Med, 42(2), 225–236. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Scott K, Hwang I, Chiu W-TA, Kessler R, Sampson N, Angermeyer M, Beautrais A, Borges G, Bruffaerts R, & de Graaf R (2010). Chronic Physical Conditions and Their Association With First Onset of Suicidal Behavior in the World Mental Health Surveys Psychosomatic Medicine, 72(7), 712–719. 10.1097/PSY.0b013e3181e3333d [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Simon G, Johnson E, Lawrence J, Rossom R, Ahmedani B, Lynch F, Beck A, Waitzfelder B, Ziebell R, Penfold R, & Shortreed S (2018). Predicting suicide attempts and suicide deaths following outpatient visits using electronic health records. American Journal of Psychiatry, 175(10), 951–960. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Simon G, Johnson E, Shortreed S, Ziebell R, Rossom R, Ahmedani B, Coleman K, Beck A, Lynch F, & Daida Y (2024). Predicting suicide death after emergency department visits with mental health or self-harm diagnoses. General Hospital Psychiatry, 87, 13–19. 10.1016/j.genhosppsych.2024.01.009 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Somé N, Noormohammadpour P, & Lange S (2024). The use of machine learning on administrative and survey data to predict suicidal thoughts and behaviors: a systematic review. Frontiers in Psychiatry, 15(March 4), 1291362. 10.3389/fpsyt.2024.1291362 [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Stanley B, Currier G, Chesin M, Chaudhury S, Jager-Hyman S, Galfalvy H, & Brown G (2018). Suicidal behavior and non-suicidal self-injury in emergency departments underestimated by administrative claims data. Crisis, 39(5), 318–325. 10.1027/0227-5910/a000499 [ DOI ] [ PubMed ] [ Google Scholar ] Steeg S, Quinlivan L, Nowland R, Carroll R, Casey D, Clements C, Cooper J, Davies L, Knipe D, Ness J, O’Connor R, Hawton K, Gunnell D, & Kapur N (2018). Accuracy of risk scales for predicting repeat self-harm and suicide: a multicentre, population-level cohort study using routine clinical data. BMC Psychiatry, 18(1). [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Tran T, Luo W, Phung D, Harvey R, Berk M, Kennedy R, & Venkatesh S (2014). Risk stratification using data from electronic medical records better predicts suicide risks than clinician assessments. BMC Psychiatry, 14(1), 76. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Walsh C, Ribeiro J, & Franklin J (2017). Predicting risk of suicide attempts over time through machine learning. Clinical Psychological Science, 5(3), 457–469. 10.1177/2167702617691560 [ DOI ] [ Google Scholar ] Werkmeister C, Cunningham R, Freeland A, Stanley J, Every-Palmer S, & Kuehl S (2023). Missed presentations, missed opportunities: A cross-sectional study of mental health presentation undercounting in the emergency department. Emergency Medicine Australasia, 35(2), 276–282. 10.1111/1742-6723.14114 [ DOI ] [ PubMed ] [ Google Scholar ] Whitlock J, & Knox K (2007). The relationship between self-injurious behavior and suicide in a young adult population. Arch Pediatr Adolesc Med, 161(7), 634–640. 10.1001/archpedi.161.7.634 [ DOI ] [ PubMed ] [ Google Scholar ] Wong J, Horwitz M, Zhoi L, & Too S (2019). Using machine learning to identify health outcomes from electronic health record data. Pharmacoepidemiol 2019;5:331–342. . Pharmacoepidemiol, 5, 331–342. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Youden W (1950). Index for rating diagnostic tests. Cancer, 3(1), 32–35. 10.1002/1097-0142(195001)3:1<32::AID-CNCR2820030106>3.0.CO;2-3 [ DOI ] [ PubMed ] [ Google Scholar ] Zhong Q, Mittal L, Nathan M, Brown K, Knudson González D, Cai T, Finan S, Gelaye B, Avillach P, Smoller J, Karlson E, Cai T, & Williams M (2019). Use of natural language processing in electronic medical records to identify pregnant women with suicidal behavior: towards a solution to the complex classification problem. Eur J Epidemiol, 34(2), 153–162. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Supplementary Materials Supplementary Files NIHMS2141438-supplement-Supplementary_Files.docx (119.9KB, docx) ACTIONS View on publisher site PDF (453.5 KB) Cite Collections Permalink RESOURCES Similar articles Cited by other articles Links to NCBI Databases Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top

Related documents

• Non-suicidal self-injury and other key predictors of suicide deaths among children and adolescents in England: a population-based case-control study.#1107509
• Half of alcohol, drug, and self-harm presentations cannot be identified in coded emergency department data: a diagnostic accuracy study of a large language model#658774
• Psychiatric emergency events and risk factors among patients with severe mental disorders: an 8-year retrospective analysis in an urban district of Beijing.#1001248
• Psychosocial factors of deliberate self-harm in Afghanistan: a hospital based, matched case-control study#135238
• Testing an integrated model of population mental health: study protocol for the epidemiological, two-arm Mannheim incidence and inception cohort study (MAKO).#418404