JMIR Publications

ABSTRACT

Background:

Depression is a leading cause of disability worldwide, significantly impacting quality of life and imposing substantial financial burdens. Technology-delivered self-help interventions, including internet-based cognitive behavioral therapy (iCBT), have emerged as scalable and cost-effective solutions, particularly for underserved populations. However, fully self-administered interventions often struggle with limited effectiveness and low engagement. Recently, Natural Language Processing (NLP) has been applied to enhance treatment effectiveness and adherence by improving user experience through personalized interactions. Despite promising applications, previous studies have primarily compared NLP-based interventions to passive controls or heterogeneous alternatives, making it difficult to isolate NLP’s specific impact.

Objective:

This study aimed to evaluate the effectiveness of NLP-supported iCBT in a fully self-administered setting by directly comparing it with a standard iCBT program (without NLP) and a waitlist control, within a structured exploratory randomized controlled trial design.

Methods:

We implemented a randomized, double-blind, parallel-group exploratory trial to compare fully self-administered AI-supported iCBT (iCBT with NLP advisory or empathetic feedback as therapeutic contribution) and standard iCBT (without NLP feedback), both using an otherwise identical iCBT program, along with a waitlist control group. Participants (N = 1,187) were recruited online, and those meeting inclusion criteria (PHQ-9 ≥5) were randomly assigned to AI-iCBT, standard iCBT, or waitlist groups. The primary outcome was the proportion of participants with PHQ-9 scores ≥10 at Week 7 and Month 3. Secondary outcomes included changes in depressive and anxiety symptoms, functional impairment, adherence, and user satisfaction. The analysis followed the modified intention-to-treat (mITT) principle. Mixed-effects models for repeated measures (MMRM) were used for statistical analysis, adjusting for baseline imbalances and stratification factors (age, gender, and baseline PHQ-9 score).

Results:

At Week 7, the iCBT group showed the most significant reduction in the proportion of participants with PHQ-9 scores ≥10 (21.4%), followed by AI-iCBT (41.4%) and waitlist (40.0%). However, by Month 3, AI-iCBT achieved a significant reduction to 19.7% (p = 0.046 vs. control), while the iCBT group rebounded to 29.3% (p = 0.413 vs. control). This suggests that AI-iCBT may provide more sustained antidepressant effects. Adherence trends showed AI-iCBT participants had lower early participation rates but better retention after Week 4 (p = 0.026 for time-by-group interaction).

Conclusions:

To our knowledge, this is the first randomized trial directly isolating NLP’s impact on depression in fully self-administered iCBT. Results suggest that while standard iCBT may provide faster short-term symptom relief, AI-enhanced iCBT led to more sustained improvements and a greater reduction in the number of participants with moderate to severe depression over time. These findings highlight the potential of NLP to optimize digital mental health interventions, although further research is needed to refine its implementation and maximize its long-term benefits. Clinical Trial: UMIN00001922