Examining the Nature of Suicide Risk over Time Using Machine Learning
Principal Investigator: 
Organization: 
Florida State University

It is commonly said among scientists and clinicians that suicide risk changes the closer an individual gets to attempting suicide. That is, different risk factors for suicide change in their importance as the individual gets closer to attempting suicide. Though this belief is widely accepted, it is a belief that is based on little-to-no scientific evidence. In addition to this issue, a team of researchers recently looked at all the studies published in the last half-century that examined risk factors for suicide to see how accurate the field is at predicting who will die by suicide or not. That team found that, overall, traditional risk factors and approaches to predicting suicide risk are little better than chance at determining who will die by suicide. Dr. Jessica Ribeiro, Assistant Professor in Psychology at Florida State University, aims to address these two problems.

Last year, Ribeiro used machine learning—a technique where a computer improves itself iteratively with almost no human input—to develop an algorithm that can accurately predict who will make a non-lethal suicide attempt using information from just days and weeks before their attempt. This algorithm performed better than traditional techniques used by most researchers. To develop the algorithm, Ribeiro used data collected from 1,000 people with recent and severe histories of nonsuicidal self-injury and/or nonfatal suicide attempts at four separate time points (i.e., baseline, 3-day, 2-week, and one-month follow-ups).

Now, Ribeiro plans on following up with those participants two years after collecting baseline data from them. With this 2-year follow-up data, Ribeiro plans to develop another algorithm using machine learning and examine how well this new algorithm predicts non-fatal suicide attempts compared to the short-term machine learning algorithm she developed before.

Ribeiro will also be able to compare which risk factors within these two algorithms are most important, which will show whether different risk factors for suicide become more important as the person’s suicide attempt draws closer. For example, access to means for suicide, such as guns or pills, may become more important the closer to a suicide attempt the individual gets.

This work, says Ribeiro, has the potential to substantially advance our ability to detect suicide risk. The methods to detect suicide risk currently used in military setting are inaccurate, which, Ribeiro says, results in considerable misallocation of resources. Improving the accuracy of suicide risk detection will improve resource efficiency and effectiveness by better identifying those at risk for suicide. In addition to improved accuracy, Ribeiro’s study will inform our understanding of the nature of suicide risk over time and create tools that detect those at risk for suicide in the distant future and those at imminent risk for suicide. This will help optimize the timing and frequency of risk detection efforts undertaken in the United States military.

No news on file at this time.