Back to All Events

Leveraging Machine Learning Methods to Predict Active Suicidal Thinking in Young Adults with Schizophrenia

Leveraging Machine Learning Methods to Predict Active Suicidal Thinking in Young Adults with Schizophrenia

Damla Duendar,(1) Khang Le,(1) Amelia Blanton,(2) Anne Thompson,(2,3) Brittany Gouse, (2,3) Hannah Brown, (2,3) Archana Venkataraman (1)
1: Department of Electrical and Computer Engineering, Boston University, USA
2: WRAP Research Program, Boston Medical Center, Boston, MA
3: Boston University Avedisian and Chobanian School of Medicine, Boston, MA

Background: Suicide is a leading cause of preventable death following the first episode of psychosis. This study aims to develop a predictive machine learning model to identify college students with psychosis who are at heightened risk of depressive symptoms and suicidal thinking.

Methods: We leveraged sociodemographic data and psychological assessments of N = 227 college students with lifetime history of psychosis using the 2019-2020 Healthy Minds Study Data Set. Our model classifies individuals into three categories: no or mild depression without suicidal ideation (SI), moderate or severe depression without SI, and patient with active SI. Depression severity was assessed using the Patient Health Questionnaire-9. The study evaluated the performance of XGBoost and Random Forest models, using Recursive Feature Elimination (RFE) to select the top 30, 40, and 50 features. Additionally, a hierarchical classifier employing Random Forest as the local classifier was developed using the local classifier per parent node approach. In this approach, individuals were first classified by depression severity, followed by a local classifier to assess the risk of suicidal ideation.

Results: The performance of the Random Forest and XGBoost models was compared using the macro recall score across the outer folds of the nested cross-validation. The best macro recall score of 0.5720 ± 0.0511 was achieved using XGBoost with the top 50 features. The hierarchical classifier demonstrated superior performance, achieving an average recall score of 0.9286 ± 0.0426 on the training and validation set and a recall score of 0.8824 on the held-out test set.

Conclusions: The results of this study could help identify college students at a higher risk of suicidal thinking, allowing for timely support. Future work will apply this framework to the study of first-episode psychosis in order to identify patients in urgent need for intervention.