Comprehensive Summary
Klee et al. performed a secondary analysis on the null DISCOVER trial, comparing different types of feedback (tailored, nontailored, and none [control]) delivered following an online depression screening to determine heterogeneity of treatment effects (i.e., which individuals' depression trajectory benefitted from receiving feedback). To do this, the researchers used causal forests to estimate conditional average treatment effects (CATEs) and evaluated model performance using the Area Under the Targeting Operator Curve (AUTOC). The analysis consisted of 946 participants who screened positive for at least moderate depression but had not been recently diagnosed or treated, and who were then separated into 3 groups for each experimental condition. Baseline measures included depression (PHQ-9), anxiety (GAD-7), somatic symptoms (SSS-8), illness beliefs, and other sociodemographic factors. The analysis investigated average treatment effects (ATE) and heterogeneous treatment effects (HTE), however, the researchers found no significant improvements or adverse effects across subgroups, whether comparing feedback to no feedback or between feedback types. Moreover, allocating feedback based on predicted CATEs did not enhance outcomes, suggesting that the benefit of automated feedback is uniform (and minimal) across individuals. Interestingly, Klee et al. found that participants with stronger beliefs in treatment control showed slightly less favorable outcomes when given feedback, though this effect was small. The authors conclude that while causal forests effectively capture complex interactions, there was no detectable harm or benefit from feedback provision after online depression screening, and that future research should examine whether screening alone can activate behavioral change.
Outcomes and Implications
Klee et al.’s analysis provides important insight into the potential utility of automated feedback following online depression screening. While the study found no measurable benefit or harm from providing tailored or nontailored feedback, it highlights the need to consider individual differences and psychological factors, such as treatment control beliefs, when designing digital mental health interventions. Given that feedback allocation based on causal forests-predicted CATEs did not improve outcomes, the results suggest that automated feedback alone may have limited capacity to influence depressive symptom trajectories. Nevertheless, this work underscores the potential for using machine learning approaches, such as causal forests, to explore complex interactions between baseline characteristics and treatment response in future trials. Though unaddressed here, the lack of detectable effect may reflect not only the aforementioned minimal impact of automated feedback but also the inherently impersonal nature of machine-generated results, which could feel dehumanizing or insufficiently engaging to participants. Nevertheless, the findings have implications for digital mental health strategies. Future research could investigate whether the act of screening itself prompts behavioral activation, or whether integrating additional support, follow-up, or human guidance might enhance effectiveness. Moreover, the study provides a framework for examining potential harms in vulnerable subgroups, ensuring that digital interventions are safe and equitable. Ultimately, Klee et al.’s work contributes to understanding the boundaries of automated feedback interventions and informs the design of more personalized, evidence-based digital mental health interventions.