The importance of predicting depressive risk has become increasingly prominent in recent years. Research has shown dynamic associations between depressive symptoms and multiple components of the daily stress process model (e.g., stressor exposure, stress appraisal, and stress reactivity). However, an integrated analysis of the predictive effect of stress processes on depressive states is still warranted. More importantly, although studies have been conducted to improve the prediction accuracy of depression using machine learning algorithms, these prediction models have primarily focused on inter-individual differences in depressive risk factors while overlooking the intra-individual dynamics of stress processes and depressive states. Given that fluctuations in individuals’ depressive states can effectively guide clinical practice in answering the key questions of “when to intervene” and “for whom to intervene”, this study aimed to use ecological momentary assessment (EMA) data and adopt a multilevel decision tree approach to construct a dynamic prediction model of depressive states using stress processes.
A sample of 356 Chinese college students completed five momentary assessments per day for seven days. In each assessment, they completed measures of depressive states, stressful life events (stressor exposure), perceived stress (stress appraisal), positive and negative affect (affective reactivity), rumination and stressor anticipation (cognitive reactivity), present and anticipated stress coping (behavioral reactivity), and physical symptoms and discomfort (physical reactivity). A multilevel decision tree approach (i.e., generalized linear mixed model (GLMM) tree) was employed to account for the multilevel structure of the data and the differences in individuals’ general levels of depression (i.e., random intercept). In addition to the momentary score of each stress process factor, we also calculated the cumulative mean and deviation of each factor as indicators to further characterize the dynamics of daily stress processes. To effectively predict and warn individuals of potential depressive states in the near future, we constructed a dynamic prediction model of stress processes at the current moment on the depressive states at the subsequent moment (approximately three hours later).
Our analysis revealed several key findings. First, the model including negative affect (distress), stressors, and rumination accurately predicted whether individuals would experience depressive states three hours later, with distress levels (negative affective reactivity to stressors) emerging as the top risk factor. Second, even excluding affective factors, the model effectively predicted depressive states using present and anticipated stress coping, rumination, discomfort, and perceived stress. This has practical advantages when frequent assessment of affective states is not feasible and too intrusive, or when at-risk individuals may not disclosure their actual affective states if asked directly. Third, multiple components of the daily stress processes cumulatively acted on individuals, jointly predicting their subsequent risk of depression. For example, more stressors and higher levels of distress jointly predicted a higher tendency towards depressive states subsequently. Fourth, dynamic indicators such as cumulative means and deviations of stress processes played crucial roles in predicting depressive states. These findings highlight the complexity and multifaceted nature of stress processes in influencing depressive symptoms.
The study makes a substantial theoretical and practical contribution by examining depression prediction from a dynamic perspective. By integrating a variety of daily stress process factors and their dynamic characteristics, this study identified key stress process factors in predicting depressive risk and revealed the synergistic effects of their various combinations. These findings expand previous research on the relation between stress and depression and deepen our understanding of the complex predictive pathways of stress processes on depressive states. In addition, this study utilized multilevel decision trees and ecological momentary assessment to construct a near-term warning model of depression with both interpretability and predictive accuracy. This provides an effective decision tool for real-time monitoring and identification of potential depressive risk in daily life, guiding the implementation of just-in-time adaptive intervention for depression.
