JMIR Publications

ABSTRACT

Background:

AI-powered digital mental health interventions (DMHIs) are a promising approach to address barriers to traditional mental health care. However, there is limited evidence of the real-world, immediate, and sustained benefits that are attributed to these platforms.

Objective:

The purpose of this real-world, retrospective study is to explore patterns of perceived mood and stress change associated with the use of Yuna, an AI-powered DMHI. We aimed to: 1) describe user demographics and session characteristics, 2) quantify the magnitude of mood and stress change within sessions and over time with continued Yuna use, and 3) identify session-level factors associated with changes in mood and stress.

Methods:

Adult Yuna users (aged ≥18 years) who initiated at least one session with the Yuna app were included in the present study. Users self-reported mood and stress on visual analog scales (range: 0-1) before and after sessions. Linear mixed effects models were used to explore the immediate, within-session improvements in mood and stress and the sustained, between-session changes in symptoms of mood and stress. Linear mixed effects models were used to examine session-level predictors of within-session improvements, including baseline symptom severity, session duration, total number of unique therapeutic approaches used, safety guardrail activation, and gender. Analytic sample size varied across models based on data availability. Sensitivity analyses evaluated the robustness of findings by characterizing missingness and potential selection effects. Users and sessions with complete pre-post mood and stress data were compared to those without pre-post mood and stress data, and a logistic regression model evaluated predictors of post-session rating completion.

Results:

A total of 5,549 real-world users were included (52.3% female; mean sessions = 3.44). Users demonstrated significant, immediate within-session improvements in both mood (d = 0.55) and stress (d = 0.56), with sensitivity analyses yielding consistent results. Between-session analyses revealed gradual improvements in baseline mood (d = -0.009) and stress (d = -0.011). Baseline symptom severity was the strongest predictor of immediate, within-session change (mood: β = .118, SE = 0.003, p < .001; stress: β = .131, SE = 0.004, p < .001), followed by session duration (mood: β = .027, SE = 0.003, p = .003; stress: β = .029, SE = 0.003, p < .001). A greater number of unique therapeutic approaches used was associated with smaller improvements in both outcomes (mood: β = -0.007, SE = 0.003, p = .043; stress: β = -0.011, SE = 0.003, p = .002).

Conclusions:

Use of Yuna, an AI-powered DMHI, was associated with perceived within-session improvements in mood and stress, with preliminary evidence of gradual improvements in mood and stress across repeated sessions. However, the absence of a control group and potential selection bias preclude causal conclusions. These findings offer promising, real-world evidence for AI-powered DMHIs as accessible, on-demand support tools. Prospective, controlled designs are needed to establish causal effects and evaluate the sustainability of observed improvements.