Predicting generalized anxiety disorder among Chinese depressed adolescents: an explainable machine learning approach.

BACKGROUND

The co-occurrence of depression and anxiety in adolescents is associated with a greater risk than the presence of depressive symptoms alone. Predicting anxiety disorders among depressed adolescents is critical for interventions and therapeutic tools.

METHODS

We recruited 2316 depressed adolescents through the Chinese Adolescent Depression Cohort (CADC) and collected 34 predictive factors for model construction. The Light Gradient Boosting Machine (LightGBM) prediction model and Shapley Additive Explanations (SHAP) algorithm were implemented for in-depth interpretation of the predictive importance of different factors. Furthermore, chi-square automatic interaction detection (CHAID) and ordinal logistic regression were used to explore the factor interactions and validate the importance of the SHAP value-based factors, respectively.

RESULTS

Nine key risk factors were identified. In addition to depressive severity, rumination, perceived stress, sleep quality, alexithymia, peer victimization, academic stress level, emotion-focused coping, and parental overprotect were recognized as key risk factors for the onset of anxiety. Resilience was recognized as a protective factor. Interaction analysis captured critical interactions between depression and six other risk factors in relation to different levels of anxiety risks. Interactions between the protective effect of resilience and four risk factors were also analyzed. High-risk subgroups and low-risk groups for different levels of anxiety were identified through the CHAID decision tree. The high-risk subgroups for severe anxiety include (1) adolescents with severe depression symptoms, (2) with moderate depression symptoms and high rumination, and (3) with severe depression symptoms and high alexithymia. The low-risk subgroups are (1) adolescents with low depression and rumination, (2) with low depression, low alexithymia, and more parental care, (3) with low depression, moderate rumination, and moderate academic stress.

CONCLUSIONS

Utilizing an explainable machine learning approach enables us to identify the risk and protective factors for anxiety disorders among depressed adolescents. The SHAP analysis results suggest that depression severity was the most important predictor for co-morbid anxiety. CHAID decision tree further identified risk subgroups. These findings suggest that clinical workers take into consideration the above risk and protective factors as well as their interactions to develop appropriate therapies for the prevention of comorbid anxiety with depression.