(PS1-82) Applying Machine Learning to Predict Depression Severity and Evaluate Its Impact on the Effectiveness of a Game-based Intervention

Abstract Text: Introduction and Problem Statement: Accurately identifying individuals at risk for severe depression is crucial for early intervention, yet traditional approaches often fail to capture the most influential factors driving symptom severity. Machine learning models provide a data-driven approach to enhance predictive accuracy and personalize mental health treatments. This study aims to (i) apply tree-based methods to identify key predictors of depression severity and (ii) examine whether baseline depression severity predicts the effectiveness of a game-based intervention for depression.

Methods: Participants included 101 adults aged 20-50 (M = 33.01) with mild-to-moderate MDD. Participants were randomized to the Game4Mood intervention or waitlist control group at a ratio of 3:1, with stratification to control for gender identity. Participants were mostly female (77.8%) and White (61.5%). Depression symptom severity (MADRS) was measured at baseline, week 4, and week 8. A regression tree and a random forest were used to examine predictors of depression severity, including baseline measures of positive and negative affect, rumination, anxiety, pleasure, pleasure, quality of life, and demographic variables). Linear regressions were then used to examine the association between baseline depression severity and post-treatment intervention effectiveness.

Results: The regression tree identified rumination, negative affect, and quality of life as the most influential factors. Individuals with rumination scores above 40.5, negative affect scores above 15.5, and quality of life scores below 11.8 were at the highest risk for severe depressive symptoms. A random forest model with 1000 iterations confirmed these key predictors. Linear regression analyses indicated that while baseline depression severity significantly predicted post-treatment scores in the control group (β = 1.22, p = 0.003), the relationship was non-significant in the treatment group (β = 0.32, p = 0.167). However, post hoc analyses revealed that the treatment group had significantly lower post-treatment MADRS scores than the control group only for individuals with Moderate-Severe depression (MADRS 22-35; β = -6.88, p = 0.002); this effect was not observed in individuals with Mild-Moderate depression (MADRS 14-21; β = -1.42, p = 0.818).
Conclusions and Implications: The regression tree analysis aligns with the psychosocial model, highlighting that environmental factors (quality of life), psychological factors (negative affect), and cognitive factors (rumination) all contribute to depression severity. While game-based intervention was effective regardless of baseline depression severity, its benefits were particularly pronounced for individuals with moderate-to-severe symptoms, suggesting that Game4Mood may be especially valuable for those experiencing more severe distress. This study underscores the importance of using data-driven approaches to personalize mental health treatments and improve early identification of individuals at risk for severe depression.