Anxiety onset in adolescents: a machine-learning prediction

Abstract

Recent longitudinal studies in youth have reported MRI correlates of prospective anxiety symptoms during adolescence, a vulnerable period for the onset of anxiety disorders. However, their predictive value has not been established. Individual prediction through machine-learning algorithms might help bridge the gap to clinical relevance. A voting classifier with Random Forest, Support Vector Machine and Logistic Regression algorithms was used to evaluate the predictive pertinence of gray matter volumes of interest and psychometric scores in the detection of prospective clinical anxiety. Participants with clinical anxiety at age 18-23 (N = 156) were investigated at age 14 along with healthy controls (N = 424). Shapley values were extracted for in-depth interpretation of feature importance. Prospective prediction of pooled anxiety disorders relied mostly on psychometric features and achieved moderate performance (area under the receiver operating curve = 0.68), while generalized anxiety disorder (GAD) prediction achieved similar performance. MRI regional volumes did not improve the prediction performance of prospective pooled anxiety disorders with respect to psychometric features alone, but they improved the prediction performance of GAD, with the caudate and pallidum volumes being among the most contributing features. To conclude, in non-anxious 14 year old adolescents, future clinical anxiety onset 4-8 years later could be individually predicted. Psychometric features such as neuroticism, hopelessness and emotional symptoms were the main contributors to pooled anxiety disorders prediction. Neuroanatomical data, such as caudate and pallidum volume, proved valuable for GAD and should be included in prospective clinical anxiety prediction in adolescents.

Fig. 1. Inclusion flowchart.

AUDIT Alcohol Use Disorder Identification Test.

Fig. 2. SHAP values and importance of features at age 14 in the prediction of any future anxiety (N = 156) vs. healthy control (N = 424).

Each dot represents an individual in a given cross-validation iteration. Positive Shapley values indicate contribution of a feature value in favor of the positive class (future anxiety) prediction, negative Shapley values are in favor of the negative class (healthy control) prediction. Larger absolute Shapley values indicate larger impact on the model output. The 20 most contributing features are shown. BNST bed nucleus of the stria terminalis.

Fig. 3. SHAP values and importance of features at age 14 in the prediction of future generalized anxiety disorder (N = 42) vs. healthy control (N = 424).

Each dot represents an individual in a given cross-validation iteration. Positive Shapley values indicate contribution of a feature value in favor of the positive class (future generalized anxiety) prediction, negative Shapley values are in favor of the negative class (healthy control) prediction. Larger absolute Shapley values indicate larger impact on the model output. The 20 most contributing features are shown. BNST bed nucleus of the stria terminalis.

Fig. 4. SHAP values and importance of features at age 14 in the prediction of future multiple anxiety diagnoses (N = 42) vs. healthy control (N = 424).

Each dot represents an individual in a given cross-validation iteration. Positive Shapley values indicate contribution of a feature value in favor of the positive class (future multiple anxiety diagnoses) prediction, negative Shapley values are in favor of the negative class (healthy control) prediction. Larger absolute Shapley values indicate larger impact on the model output. BNST bed nucleus of the stria terminalis, DmOFC dorsomedial orbitofrontal cortex. The 20 most contributing features are shown.