Using Supervised Machine Learning to Predict the Role of Cognitive Control in the Suicide Ideation-to-Action Model

Research Aims and Hypotheses: The main objective of this study is to formulate a youth prediction model stemming from leading theoretical models of suicide that account for multiple interacting individual difference risk factors, with a thorough consideration of social determinants of health. Cognitive control is defined as a set of neurocognitive processes that guide cognitive inhibition, mental flexibility, and working memory to direct effortful and goal-oriented behaviours. I hypothesize that cognitive control deficits interact with well-known clinical risk factors of suicide attempt (e.g., psychosis, social withdrawal, substance use) to predict a future suicide attempt among youth. Given the greater risks of suicide in gender diverse youth, I hypothesize that gender identity will have moderating effects on existing risk factors in predicting attempt among youth.

Sample, Measures, and Procedures: I am conducting the first investigation to our knowledge to predict suicide attempt with supervised machine learning (ML) algorithms using clinical and cognitive variables as features in youth. The proposed project leverages the CAMH Toronto Adolescent and Youth (TAY) Cohort Study, which is recruiting youth ages of 11 to 24 with pre-existing mental health concerns over five years. Upon study entry, participants complete psychodiagnostic assessments, numerous multi-informant measures, and cognitive testing on an annual basis. Longitudinal data (N=750) for one year are expected by Summer of 2023. The Columbia Suicide Severity Rating Scale is used to assess previous suicide attempts at baseline and future attempt will be classified as an attempt that occurred between baseline and one-year follow-up.

Data Analysis: Machine learning models are compared to a generalized linear model for the most parsimonious solution. Comparisons between all models are evaluated based on area under the curve, precision recall curve, sensitivity, specificity, positive predictive value, and discrimination (i.e., effect size). These variables are coded together and separately to distinguish the impact of demographic variables in suicide attempt prediction. The pilot data (n=265) shows 97 (37%) participants endorsing an attempt in their lifetime and 37 (14%) within the past 12 months. These findings highlight the advantage of this study in that this enriched sample will likely have an event rate where prediction models can be developed. Moreover, 28.2% identified as gender diverse in our pilot data. Our sample is expected to be clinically heterogeneous and includes a variety of gender diverse youth which will lead to the development, validation, and implementation of an inclusive suicide prediction model for youth.

Potential Impact and Next Steps: Consistent with the current call, a large-scale longitudinal study of service-seeking youth would be ideally positioned to (1) identify a variety of risk factors that work together to predict suicide attempt in youth, (2) recognize vulnerable youth before they die by suicide, and (3) reveal underlying clinical risk factors and neurocognitive pathways predictive of suicide attempt and potentially responsive to early intervention. Knowledge generated from this study will be used to inform suicide risk assessment in youth mental health programs at CAMH and beyond. I aim to advance the scientific knowledge needed to reduce suicide rates in youth.