Category: Suicide and Self-Injury

Symposium

Symposium 50 - Machine Learning Techniques Predict Suicide Ideation, Attempts, Death: Implications for Theory and Practice

Friday, November 17
3:30 PM - 4:30 PM
Location: Sapphire Ballroom K & L, Level 4, Sapphire Level

Keywords: Suicide | Longitudinal | Methods
Presentation Type: Symposium

Recent meta-analyses show that the prediction of suicidal thoughts and behaviors (STBs) is on par with random guessing and have not improved across 50 years of research (e.g., Franklin et al., 2017). A major reason for this is that, based on theories that propose the suicide is caused by a small handful of factors (usually 1-5 factors), most prior studies have attempted to predict STBs with 1 to 5 factors. We reason that the causes of STBs are far more complex and that accurate prediction could be produced by machine learning techniques that find the optimal combination of tens or hundreds of factors. The talks in this symposium provide evidence for this view, showing that machine learning techniques powerfully outperform traditional techniques to predict STBs in very large samples. Talks will use area under the curve (AUC) as their primary metric, with 0.50 indicating random guessing and 1.0 perfect prediction.


In the first talk, Dr. Colin Walsh will draw on a large longitudinal dataset (3,250 suicide attempters; 22,958 total controls) to show that hundreds of constructs combined with machine learning techniques (AUCs = 0.83 - .91) produces significantly more accurate prediction of future suicide attempts compared to the same set of constructs combined with traditional multiple logistic regression (AUCs = 0.62 - 0.63). In the second talk, Dr. Jessica Ribeiro will present similar evidence from large longitudinal dataset that includes suicide decedents (502 suicide decedents; 38,290 controls), with machine learning techniques (AUCs = 0.85 – 0.88) predicting future suicide death with significantly greater accuracy than traditional multiple logistic regression (AUC = 0.58). In the third talk, Xieyining Huang will draw on a smaller dataset (but still large for the field; N = 917) to present similar machine learning evidence on distinguishing suicide ideators from non-ideators. Machine learning techniques produced superior classification accuracy (AUCs = 0.73 - 0.79) compared to traditional multiple logistic regression (AUC = 0.50). As discussant, Dr. Joseph Franklin will build from these findings to propose a novel theory of suicidality.

Learning Objectives:

Joseph Franklin

Assistant Professor
Florida State University

Presentation(s):

    Send Email for Joseph Franklin

    Joseph Franklin

    Assistant Professor
    Florida State University

    Presentation(s):

      Send Email for Joseph Franklin

      Colin Walsh

      Assistant Professor
      Vanderbilt University Medical Center

      Presentation(s):

      Send Email for Colin Walsh

      Jessica D. Ribeiro

      Assistant Professor
      Florida State University

      Presentation(s):

      Send Email for Jessica Ribeiro

      Xieyining Huang

      Graduate Student
      Florida State University

      Presentation(s):

      Send Email for Xieyining Huang


      Assets

      Symposium 50 - Machine Learning Techniques Predict Suicide Ideation, Attempts, Death: Implications for Theory and Practice



      Attendees who have favorited this

      Please enter your access key

      The asset you are trying to access is locked. Please enter your access key to unlock.

      Send Email for Machine Learning Techniques Predict Suicide Ideation, Attempts, Death: Implications for Theory and Practice