As the cause of more than 800,000 deaths every year, suicide has become a global public health priority. It is the second leading cause of death in young people aged 15 to 29 years, and for every death, it is estimated that there are 20 other suicide attempts [1]. In Spain, suicide has been the main cause of unnatural death since 2012 [2].

According to the World Health Organization [3], suicidal behavior refers to a range of behaviors that includes thinking about suicide (or ideation), planning to commit suicide, attempting suicide, and suicide itself.

Until a few decades ago, research efforts have been focused on curbing suicide deaths by trying to predict their occurrence. This predictive approach consisted of semistructured risk assessment using lists of risk factors and sometimes included suicide risk questionnaires or scales to express risk as low, moderate, or high [4,5]. Because suicide deaths are statistically a rare event, it has been difficult to develop sensitive tools with sufficient predictive value [6]. Recent reviews [5,7-10] of these models advocate a shift from models based on suicide prediction to those that emphasize assessment and management of the risk of suicide by identifying variables related to suicide behavior and stratifying risk in terms of severity and temporality.

Suicidal behavior has been consistently found to be associated with emotional states such as depression and hopelessness [11,12]. Bryan and Rudd [9] collected different variables that have been empirically demonstrated to be essential for risk assessment: predisposition to suicidal behavior (ie, psychiatric diagnoses, previous suicidal behavior), identifiable precipitants or stressors (ie, significant loss, relationship instability), a patient’s symptomatic presentation (eg, anhedonia, low self-esteem, sadness, dyssomnia, fatigue), presence of hopelessness, nature of suicidal thinking (eg, ideation, suicidal plan, lethality of means, explicit suicidal intent), impulsivity and self-control, and protective factors (eg, social support, life satisfaction). Emotional dysregulation seems to be also an important predictor of suicidal outcomes [13,14].

To improve accuracy in risk evaluations, ecological momentary assessment has been used to study suicidal behavior [15], which involves repeated sampling of people's behavior in real time in their natural environments, now typically collected via smartphones. This approach attempts to minimize recall bias and maximize ecological validity. Recent research using mobile phone–based momentary ecological assessments showed that suicidal ideation varied over short periods of time [16], indicating that real-time assessments and ecological validity could be a crucial approach for suicide prevention.

With more than 3.8 billion users around the globe [17], social media has transformed the world. People express their thoughts and emotions through social media [18]. These new forms of social interaction have been linked to suicidal behavior, nevertheless, recent studies [19] have highlighted the potential for social media to offer assistance in suicide prevention.

Twitter currently has 340 million users [20] who, in microblogging format, communicate what they are thinking or doing at a particular moment publicly with a limited number of characters. People express suicidal tendencies on Twitter [21], and although there is a support mechanism among users, this system is not automatic or in real time.

According to Christensen [22], web-based interventions for suicide prevention have focused mainly on three directions: (1) web-based screening for suicidality, (2) web-based therapeutic interventions for suicide prevention, and (3) real-time identification of individuals at risk, either by people or by computer language processing systems.

The use of social media in the real-time detection of mental health has already been proven. Specifically, Twitter has been proven useful in predicting depression [23-25], postpartum depression [24], and even posttraumatic stress disorder [26,27]. Machine learning algorithms have been used to assess the risk of suicide and identify suicidal individuals. Automatic machine learning classification systems that are able to effectively differentiate people who are at risk of suicide from those who are not [28-30] and identify temporal patterns in posts before suicide [31] have been developed. Reviews on the subject [22,32] yield similar results: social media is an empirically tested tool for suicide detection, but further validation is needed.

Recent studies have incorporated human coders in order to create language classification systems or validate machine learning results from natural language processing systems [33]. Nevertheless, the need to incorporate mental health experts in suicide assessments for the improvement of accuracy has been noted [34]. Furthermore, the detection of the risk of suicide for social media users located in Spain has yet to be explored.

Our objective was to analyze the risk of suicide among Twitter users who post in Spanish, by assessing the emotional content of their posts and other variables that have been identified as being related to suicidal behavior, such as perception of defeat, helplessness, and social support.

A total of 2509 tweets were obtained, of which 2018 were deemed not relevant by 3 evaluators, and 275 were deemed not relevant by 2 of 3 evaluators. There were 216 tweets that were found to be relevant by most evaluators, with 68 tweets considered to be relevant by all evaluators, exhibiting moderate reliability (Fleiss κ=0.41).

Tweets mainly conveyed sadness and defeat, with there being no desire to escape, no support, and no feelings of helplessness (Table 1). The median overall risk of suicide was 1.50 (IQR 1.00) on a scale from 0 to 4. The median severity of risk was 1.00 (IQR 1.16); 96.9% (186/192) of tweets did not indicate the presence of daily recurring thoughts of suicide, and the median intensity of suicidal thoughts was 4.50 (IQR 3.00) on a scale from 0 to 10.

There were direct correlations between severity of the risk of suicide at the time of generating the tweet and sadness, general risk, and intensity of suicide thoughts, as well as inverse correlations with extroversion, rational trait, and joy (Table 2).

The severity of risk of suicide at the time of generating the tweet was higher in people who expressed feelings of defeat, rejection, desire to escape, feelings of helplessness, lack of social support, and daily recurrent thoughts (Table 3).

The linear regression model (R²=0.750; adjusted R²=0.710) showed that the intensity of autolytic thoughts, fear, and valence were predictors of the severity of the risk of suicide at the time (Table 4). Both the intensity of autolytic thoughts and valence had positive coefficients, and fear had a negative coefficient. This indicated that when intensity was higher, valence was more positive, and when fear was lower, the severity of suicidal risk was higher. The model explained 75% of the variance.

Suicide prevention is a crucial field that needs to be developed to stop preventable deaths worldwide. The findings of our study reveal that social media can be used to help to identify individuals at risk. These findings suggest that it is possible to identify suicidal behavior through Spanish tweets, and it is possible to identify these posts, not only by using a suicide-related lexicon but also, by filtering tweets based on their emotional content. Tweets that show sadness, defeat, and perceived lack of social support suggest that there is a risk of suicide. These variables have been commonly associated with symptoms of depression and hopelessness [12,47].

One of the main challenges in suicide prevention is identifying not only the people at risk of experiencing suicidal behavior at some point in their lives but those who are at risk at a particular moment, in our case, while they are writing the tweet.

Suicidal ideation and its risk factors can fluctuate over short periods of time [16], which demonstrates the importance of differentiating general suicide risk (or suicide status, such as in individuals with long-term risk factors) from real-time suicide risk. In this exploratory study, we emphasized assessment of suicide risk at the moment of writing the post. Our results provide some clues about the phenomenon of suicidal behavior.

In our study, although the variable desire to escape from the situation was not identifiable in most posts showing potential risk, it was related to an increase in the severity of the risk at the time of writing the tweet in posts that expressed a desire to escape.This outcome could suggest that the variable of the desire to run from a suffering situation may be only relevant in situations with an increased risk of suicide at that time. We found the same pattern for feelings of helplessness at the time, which was only identifiable in high-risk tweets. These results are consistent with recent conceptualizations of acute suicidal behavior that suggest that this feeling of entrapment—“in which the escape from an unbearable life situation is perceived as both urgent and impossible” [48]—is linked with imminent suicidal behavior. Nevertheless, further research is needed.

Individuals exhibiting high levels of negative urgency and emotion reactivity might be more likely to develop suicidal ideation and resort to self-harm while experiencing negative affective states [14,49]. In our sample, tweets with higher risk (at the time of writing the tweets) were identified by higher sadness, higher general risk, and higher intensity of suicidal thoughts on a daily basis. In addition, they similarly showed feelings of defeat and rejection, as well as the perception of a lack of social support. These results are consistent with those in literature, with lack of social support or isolation being an especially well-established risk factor [3].

Although the role of impulsivity in suicidal behavior has not been clearly defined yet [50,51], it appears that a considerable proportion of suicide attempts are related to impulsive behavior [52,53]. In our study, we also included personality trait variables. Our results suggest that people at risk at the time of writing their tweet might show less extroversion and less rational personality traits in their posts, which could be associated with greater impulsiveness. Further research would shed more light on this subject.

Although insomnia or sleep problems variables have been identified as risk factors in suicide assessments, our findings suggest that this variable is not identifiable or relevant to social media posts. One study [54] notes that only nightmares are associated with suicidality.

We obtained preliminary data that might help us to predict increased real-time suicide risk. The interpersonal theory of suicide [55] posits that the simultaneous occurrence of 2 psychological states, a perceived burden to others, and a frustrated belonging or social isolation, as well as hopelessness regarding the potential of these states to change, results in the desire for suicide. Our findings are consistent with this theory, showing that sadness, feeling of defeat, or perceived lack of support are related to high-risk tweets. However, according to this theory, the simultaneous presence of suicidal desire and a high tolerance for pain and fear of death would be necessary to produce lethal or near-lethal behavior. The interpersonal theory of suicide also posits that high risk occurs when tolerance for pain and fear increases [49,56]. Our results seem to also suggest that real risk appears when ideation intensity increases and the fear of suicide decreases, making the resulting emotional valence less unpleasant. In other words, the detection of a decrease in fear and, therefore, a less aversive emotional state could predict an increased risk of suicidal behavior at the time on social media. If these variables effectively and consistently prove their ability to predict an increase in risk at the time, we could generate real-time machine learning systems that would detect predictor emotional states, such as the decrease in fear, to prevent potential deaths.

Our study has strengths but also limitations. The cross-sectional design of this study provided limited data about suicidal phenomenon on social networks. In the future, it would be interesting to be able to design a study that screens variation in emotional states and suicide risk over time as this may provide more relevant information on how the risk of suicide varies.

Twitter is considered to be one of the most popular social media platforms, with the greatest immediacy in posting, but because there is a character limit, some variables are barely detectable or measurable. More complex feelings or constructs are difficult to evaluate. Future research could include the evaluation of conversation threads or tweets in their context (for example, by taking into account tweets within the previous 24 hours), which would allow more accurate evaluations of relevant variables in suicide risk assessments.

The interrater reliability was moderate, differences between research and the clinical professional profile of the experts might have had an impact on reliability. In future investigations, reliability will be assessed based on expert profiles, and intensive specific training will be conducted.

One of the strengths of our study is that it incorporates experts in suicide risk assessment. In future research, it would be interesting to introduce validated scales to measure suicidality so that expert risk detection could be effectively validated. Moreover, it would be interesting for future research to use social media users with past suicide attempts to screen and validate our findings.

Our findings—identifying emotional content that might be relevant for real-time suicide prevention—can contribute to the development of new technology-based screening systems; however, more robust research is needed to establish whether social media screening can effectively reduce suicide outcomes and whether there is a way to ethically reach those individuals at risk.