Predicting suicide among Chinese adolescents with Major Depressive Disorder using the distal and proximal factor framework and machine learning (Preprint)

Abstract

BACKGROUND Depressed adolescents are at significantly higher risk of suicidal thoughts and behaviors, and socio-environmental, individual psychological traits and biogenetic factors are significant causes. However, our understanding of the specific determinants of suicidal thoughts and behaviors among depressed adolescents remains limited. OBJECTIVE The objective of this study is to develop a prediction model for suicide among Chinese adolescents with Major Depressive Disorder (MDD). Then, based on detailed and comprehensive feature contributions, this study also aims to elucidate and examine the main influencing factors and perform an in-depth interpretation of the output of the prediction model. METHODS We conducted a multi-center, cross-sectional survey from December 2020 to December 2021. A total of 2343 adolescents with MDD aged 12 to 18 years from 14 psychiatric general hospitals across 9 provinces of China were enrolled. Then, we constructed a dual path framework integrating distal and proximal factors and apply random forest (RF) classification model for risk prediction of suicide ideation and/or suicide attempt (SISA) among Chinese adolescents with MDD. Hyperparameter tuning, model training, and unbiased estimation were implemented through a stratified nested cross-validation procedure. Finally, to enhance the interpretability of the RF algorithm, we adopted the Shapley Additive exPlanations (SHAP) method for in-depth interpretation and analysis from a global and a local perspective. RESULTS Random Forest model achieved excellent performance with an accuracy of 95%, an F1-score of 95% and an AUC of 0.99. In the global analysis of the SHAP method, we found that depression severity and despair severity (proximal factors) contributed the most to SISA among the sample. Additionally, it was confirmed that borderline personality played a critical role as a distal factor. What’s more, prediction results indicated that support from significant others was a risk factor rather than a protective factor. According to dependence analysis, complex interactions between distal and proximal factors were identified, for example, the co-occurrence of high levels of borderline personality and depression may together lead to an increased risk of suicide. Furthermore, an indirect impact of cognitive reappraisal on suicide occurrence was also captured. Finally, in the local analysis, adolescent girls were at higher risk of SISA than boys in the depressive subtype. It can also be inferred that depression severity and despair severity were not necessary conditions of SISA even though they were main influencing factors. CONCLUSIONS The dual path framework integrating distal and proximal factors may provide a comprehensive perspective for suicide risk assessment. RF can be effectively used to predict suicide among Chinese adolescents with MDD and contribute to early detection, preventive interventions and intensive monitoring. As an analysis technique to interpret the prediction model with black box characteristics, SHAP can provide in-depth interpretation and reference for clinical practices.