According to the United States Center for Disease Control, “suicide is a leading cause of death in the United States, with 47,646 deaths in 2021.” [1] Suicide risk, in terms of suicidal ideation and suicide attempt, is particularly elevated in people who identify as sexual and gender minority, including lesbian, gay, bisexual, and transgender (LGBT) identities [2–5].

One of the many challenges in research on suicidal thoughts and behaviors is that outcomes are relatively rare and multi-determined [6]. The causes of suicidal behaviors may consist of a combination of factors, which vary between individuals and groups over time. As Mościcki noted, risk may be distal or proximal in time to the event [7]. Distal factors, such as past trauma, may increase risk in conjunction with changes in proximal risk factor(s), such as re-traumatization. More proximal factors can include sudden life disruptions (e.g., job loss, relationship failures) [8] and dynamic factors in the symptomology of serious mental illnesses such as major depressive disorder, bipolar disorder, and schizophrenia [9]. Importantly, although sexual and gender minority populations have greater burdens of suicidal thoughts and behaviors [10], in this study, we maintain that LGBT status is not a direct risk factor, rather it is associated with other factors that increases suicide risk. Principal among these related risks is minority stress [11], which operates at several levels including interpersonally (e.g., family rejection); social adversities such as homelessness and victimization; fears of engaging in healthcare due to system- and provider-level issues such as discrimination and bias [12]. Minority stress may contribute to or interact with other proximal risk factors for suicide such as comorbid mental health and medical conditions (e.g., depression, chronic pain) [6, 13]. A better understanding requires novel analyses of variable interactions and identification of new factors [14].

A critical barrier to better understand suicide risk and its correlates among LGBT individuals is the overall lack of structured data on sexual orientation and gender identity (SO/GI) in healthcare administrative and clinical data systems, such as electronic health records (EHR). Despite recent progress in SO/GI data collection in some health systems [15], there remain large gaps. For example, in the US, the Centers for Disease Control and Prevention (CDC) National Syndromic Surveillance Program is a crucial source of data to monitor suicide attempts necessitating medical treatment [16]. However, because most EHR systems do not have structured SO/GI data fields (or if they do the fields are largely missing data) [17], LGBT people are not visible in data reported by emergency departments. Consequently, as EHR systems continue to develop and implement SO/GI fields, researchers have turned to machine learning innovations to develop health services research for LGBT patients [18–20].

In addition to limited administrative data on suicidal ideation and suicide attempt, information about suicide mortality of LGBT patients is even more scant. Because SO/GI data are not collected at the time of death [21], no vital statistics systems include data to learn about LGBT mortality. The CDC’s National Violent Death Reporting System (NVDRS) is one example, however, it only focuses on violent death and SO/GI data are missing for approximately 80% of decedents in the data repository [22]. Recent efforts using EHR data have attempted to estimate suicide mortality for LGBT individuals, but these efforts have used either solely diagnosis-based efforts to identify patients [4] or used idiosyncratic machine learning techniques [5]. As a nascent, heterogeneous field, inquiry into LGBT suicide mortality needs more study to help determine nuances in methodology and corroboration of findings.

EHR data also offer opportunities to examine many potential risk factors for suicide, including demographic and clinical variables [23]. Temporal information in longitudinal records provides additional information by allowing researchers to differentiate an event in the past from one more proximal to the event of interest. Moreover, risk factors for suicide are likely to interact with each other and do not necessarily have a linear relationship with the outcome. A type of deep neural network (DNN) model called transformer is one potential form of analysis to identify and explore sexual and gender minority statuses and suicide risk. DNN has the capability of modeling non-linear relationships and complex interactions between covariates [24]. Transformer, in particular, has demonstrated superior capability in modeling sequence data such as temporal history [25].

A trained DNN model is sometimes viewed as a “black box” as it can be difficult for humans to interpret, partly due to the large number of hyperparameters. Fortunately, in recent years, researchers including the present team have developed explainable artificial intelligence (AI) methods to enable the estimation of feature contribution and feature interaction [26, 27]. Using these methods, we can identify risk factors and assess interactions in a similar fashion as in traditional statistical analyses.

For the present analyses, we utilized EHR data from the US Department of Veteran Affairs (VA) Corporate Data Warehouse (CDW). Over 20 years ago, the VA, which operates the largest healthcare system in the U.S., adopted a comprehensive “paperless” EHR [28] that is integrated system-wide. Data from 2000 onward is stored in the CDW, making it readily accessible for clinical research. Thus the VA EHR provides a unique laboratory in which to conduct observational research from a host of structured and unstructured data [29]. We identified a cohort of patients had or had not died by suicide and their disclosed LGBT status were identified using an NLP algorithm [30]. The objective of the study is to assess the impact of disclosed LGBT status on suicide death using DNN model and explainable AI.

Among, the 8.8 million veterans, we found 200,834 (2.3%) veterans with disclosed LGBT status and 437,689 (5.0%) veterans with disclosed non-LGBT status from clinic notes. The rest had unknown LGBT status. Before matching, patients with disclosed LGBT status had lower all-cause mortality rates (13% LGBT, 15% non-LGBT, 23% LGBT-unknown) (Table 1). Veterans with disclosed LGBT status were more likely to die from suicide (2.2% of overall all-cause mortality) than those with non-LGBT status (1.8%) and LGBT-unknown (1.1%) veterans. These findings are consistent with previous research [36].

After matching, we identified 21,942 patients who died from suicide (cases) and 21,942 controls. Cases were younger (by 1.4 years), more likely to be male, single or divorced, and non-Hispanic White. Additionally, cases were less likely to have a religious affiliation and more likely to have unknown LGBT status (Table 2).

The DNN model achieved an AUC of 76.7% and sensitivity and specificity both of 70.0% in classifying suicide death. We obtained similar AUC for training (77.6%) and validation (76.3%), showing no evidence of outfitting. Using our explainable AI, we evaluated the impact of each feature on the predicted outcome. Known risk factors for suicide death, such as suicide ideation, mental illness, and drug abuse, were among the top features in our model (Figure 2).

Our primary feature of interest (disclosed LGBT status) was found to be protective. Specifically, both disclosed LGBT and non-LGBT status was associated with lower risk than unknown LGBT status (impact score = −0.462 and −0.459, respectively). Patients with disclosed LGBT status had slightly lower risk than those who with disclosed non-LGBT status (impact score difference = −0.003) based on our analysis.

Additionally, disclosed LGBT status interacted with other features in the context of suicide mortality risk. For example, black, female, married, or older LGBT veterans appear to be at greater risk for suicide death. On the other hand, LGBT veterans of various religions and married but separated appear to be at lower risk (Figure 3).

As stated in the introduction, suicide claims tens of thousands of lives in the United States. Individuals identified as LGBT have a particularly high risk [2–5]. This, however, should not be interpreted as that sexual and gender minorities are inherently more prone to suicide. Minority stress associated with discrimination and bias, for example, is a key factor.

Predicting suicide death is extremely challenging due to its rarity. Machine learning (ML) is well-suited for this task because it can handle a large number of predictors and capture non-linear associations and interactions between variables. While using ML has enhanced prediction accuracy, these models still exhibit very low positive predictive values (PPVs) [37]. Additionally, a known limitation of ML models is their low clinical interpretability [38]. Our DNN model, which uses a case-control study design, has achieved a PPV of 69%. We are also able to provide additional insight into the model using explainable AI.

Like most observational studies, this study does not intend to draw causal conclusions [39]. Randomized controlled trials, which are the gold standard for providing causal inference, are not suitable for the present analysis, as the contemporary scientific evidence suggests that LGBT status is not modifiable. This analysis provides informative insights into the association between disclosed LGBT status and the risk of suicide death beyond structured EHR data, with adjustments for a comprehensive list of clinical factors. Moreover, the explainable AI interpretation is consistent with previous research findings, reassuring our confidence in the model’s accuracy and interpretability. Among 203 input features, many with highest or lowest model impacts have evidence in the literature. For example, suicide ideation often precede suicide attempt and death [40]. Mental illnesses such as depression and drug abuse are recognized as major risk factors for suicidality [41]. The link between gun access and suicide has been repeatedly reported [42]. Alternatively, certain religious affiliations and marital status have been found to be protective [43, 44].

After adjusting for other risk factors, we found that known LGBT status (positive and negative) was associated with reduced suicide risk compared to unknown LGBT status. Patients who were self-reported as LGBT had slightly lower risk of suicide death than those who were self-reported as non-LGBT, and both groups had lower risk than the unknown LGBT group. In other words, we found that documentation of SO/GI status does not confer an increase in suicide death when compared to patients of lacking SO/GI documentation, independent of other factors. This finding may seem surprising, but it is plausible with a closer look.

Lynch et al, reported “The crude suicide rate among sexual minority veterans (82.5 per 100,000 person-years) was higher than the rate in the general veteran population (37.7 per 100,000 person-years).” In our analysis, the all-cause mortality rate is 13% and the suicide mortality rate is 0.3% for LGBT group, which is similar to those reported by Lynch et al. (13% and 0.4%). The suicide mortality rate per 100,000 person-years is, however, much lower at 39.1. This is partly because we used the enrollment date as the index date and Lynch used the first documentation date as the index date. If we use our first documentation date of LGBT within our observation period (2009/10 to 2018/12), the rate is 76.6 which is much similar to Lynch’s number (82.5). Another reason is that our original dataset included more LGBT Veterans (231,032 vs. 96,896), due to the differences in the respective NLP methods. Both NLP tools [30, 45] achieved a PPV of 86%, while our tool had higher sensitivity (88% vs. 80%), which led to a higher number of LGBT individuals. In preparation of this manuscript, we double checked the NLP PPV: we randomly sampled another 300 LGBT individuals from our dataset, performed human review, and found the PPV to be 86%.

Our analysis also adjusted for a wide range of covariates. Our results confirmed our hypothesis that disclosed LGBT status alone is distally associated with suicide death, and other risk factors have more direct impacts on the outcome. There is no direct contradiction with the previous reports. First, we did not simply report the prevalence of suicide death. Rather, we adjusted for important covariates such as suicide ideation, depression, drug abuse, marital status, and religion. Second, we did not assume that people without disclosed LGBT status are not LGBT. Third, LGBT documentation is increasing over time [46] and the social pressure associated with LGBT is decreasing. Consequently, this study can be viewed as a second-generation disparities study [47], in that it identifies several risk factors that account for group differences in prevalence of the outcome. These modifiable risk factors are a key point around which interventions can be developed and tested to reduce the elevated prevalence of suicidal risk among LGBT individuals.

Additionally, it is known that patient engagement is critical to suicide prevention. Known LGBT status suggests a higher level of engagement than unknown status, which may suggest an informed presence bias [48, 49], i.e., more engagement in the health systems increases the opportunities for SO/GI documentation and more engagement likely means greater service utilization and more opportunities for care and suicide prevention services. The latter is a particularly important point within the context of VA, which has constructed a large infrastructure for suicide prevention (e.g., creating a national network of Suicide Prevention Coordinators, predictive analytics, and patient media outreach) [50].

In looking at the interaction scores, one may have the misconception that “LGBT marriage” is a risk factor. Please note that a positive interaction score indicates that the risk is more or the benefit is less than the additive impacts. In this case, disclosed LGBT status and being married are both beneficial factors. The positive interaction score means that the combined benefit is less than the sum of the individual benefits. A similar example is diet and exercise interventions on reducing ageing-related cognitive decline that, while individually shown to have positive effects [51, 52], lack strong evidence of additive effects [53] i.e., there may be a ceiling effect of the benefits.