Nearly one million people die by suicide each year around the world. This means that suicide accounts for more annual deaths than things like war, AIDS, and car accidents. Although medical and technological advances have dramatically reduced the rates of many diseases over the past 100 years, the rate of suicide has not declined appreciably over the last century.
A LACK OF SCIENTIFIC PROGRESS
To reduce suicide, it is crucial to identify people at risk and quickly connect them with an effective intervention. Despite hundreds of studies involving millions of people over the past 50 years, our meta-analytic work has shown that science’s ability to predict suicidal behaviors has never improved. Similar to 50 years ago, current science is only slightly better than random guessing when it comes to predicting suicidal behaviors.
BARRIERS TO PROGRESS
Our work has identified several key barriers that have substantially limited progress on this front for several decades:
1. Very long studies. The average study that tries to predict suicidal behaviors is nearly 10 years long. The major problem with this is that most of what determines whether or not a suicidal behavior will occur happens in the minutes, hours, or days before the behavior. But less than 1% of existing studies have tried to predict suicidal behaviors over time periods of even a month or less (and none over time periods of less than a week). This is primarily because it is extremely difficult to conduct such short studies using traditional research methods. The result is that science has not been able to tap into the most important time period for suicidal behaviors — the moments before the behaviors occur.
2. Continued use of old (and ineffective) predictors. Since the beginning of suicide prediction research, science has stuck to a very small number of predictors, with 5 basic categories accounting for 80% of predictors and 9 basic categories accounting for over 95% of predictors. For example, a huge proportion of studies include depression (or something closely related to depression) as a predictor. It is impossible to advance the ability to predict suicide if the field continues to repeat the same studies. The inherent limitations of this repetitiveness are compounded by the fact that these popular factors (e.g., depression) are extremely poor predictors of future suicidal behaviors.
3. Predictors tested as constant factors. Most studies have treated predictors as constant phenomena. For example, many studies measure how depressed someone is and then test whether this depression level predicts suicidal behavior 10 years later. But psychological science has shown us that most predictors are not constant. Over the course of just a few minutes, someone may go from feeling fine to feeling extremely depressed. Such rapid changes may be strong predictors of suicidal behaviors, but very few studies have been able to measure such changes.
4. Predictors tested in isolation. It is extremely unlikely that a single factor can accurately predict future suicidal behaviors; accurate prediction will likely require a combination of several predictors. For example, a given person who feels hopeless is unlikely to attempt suicide in the near future. However, a white male who lives near a bridge, has few friends, and shows a rapid elevation in hopelessness in the hours after the breakup of a long-term relationship may be highly likely to attempt suicide in the near future. Once again, due to the constraints of traditional research methods, very few studies have tested such complex prediction models.
5. Limited scope and reach. Even if science had the ability to accurately predict suicidal behaviors, it currently has no platform through which to immediately identify those at high risk on a large scale. Most research to date has aimed to one day provide healthcare professionals with tools for accurate prediction. Although this is a laudable goal (and we aim to do the same), most individuals who engage in suicidal behaviors are not currently in contact with a healthcare professional. It is accordingly necessary to greatly expand the scope and reach of suicide prediction efforts.
A PROMISING POTENTIAL SOLUTION
Zenti provides a potential path through the barriers to progress noted above. First, Zenti gives us the ability to collect rich data on individuals in the moments before suicidal behavior occurs. This capability finally gives us a window into the most important time period for suicidal behaviors. Second, through its unique machine learning technology, Zenti allows us to develop a wide range of novel but theoretically-informed predictors that are likely to be far more accurate than traditional predictors.Third, Zenti enables us to examine the waxing and waning of predictors throughout a person’s social media history. This allows us to monitor rapid changes in potential risk factors with unprecedented accuracy and resolution. Fourth, we are able to employ Zenti’s flexible technology to develop and combine a large number and wide variety of predictors. Fifth, Zenti gives us the ability to screen over 100 million social media communications per day in real time to detect individuals at risk for suicidal behaviors. This represents a scope and reach large enough to have an impact on national and international rates of these behaviors. In short, with Zenti’s technology, we could revolutionize the ability to predict suicidal behaviors and immediately translate this ability into real world progress.
BRIEF PROJECT DESCRIPTIONS
Identification of suicide decedents. For our initial project, we will use Zenti’s technology to formulate several theoretically-informed predictors (i.e., “classes” that we train using a supervised machine learning algorithm interface) and use these predictors to identify suicide decedents within a massive backlog of social media communications collected over the past few years. This project will allow us to develop and refine powerful tools for suicide prediction that will serve as a crucial foundation for subsequent projects.
Prediction of future suicidal behaviors. Drawing on the algorithms developed during our initial project, we will conduct a series of studies designed to further refine our ability to accurately predict future suicide ideation, plans, gestures, attempts, and deaths. Within each study we will recruit and follow thousands of at-risk individuals and repeatedly assess their suicidal behaviors. Each study will provide an opportunity to improve our algorithms, further reducing false positives and false negatives. By the end of this project, we hope to have generated algorithms that produce near-perfect accuracy.
Our ultimate goal is to identify thousands of people at high risk for suicidal behaviors each day in real time, and to immediately connect them with a brief, game-like treatment app that dramatically reduces their risk of engaging in these behaviors. We believe that Zenti’s technology will allow us to accomplish the first half of this goal, and our own recently developed treatment apps will allow us to accomplish the second half of this goal. If we are successful, we will ignite the first meaningful worldwide reduction in suicidal behavior.