A chatbot to improve adherence to internet-based cognitive-behavioural therapy among workers with subthreshold depression: a randomised controlled trial

Subthreshold depression is common and increases the risk of major depressive disorder, with substantial impacts on work performance and economic costs. Internet-based cognitive-behavioural therapy (iCBT) offers accessible, cost-effective, and scalable support for adults with subthreshold depression, including workers, and has demonstrated efficacy in reducing depressive symptoms and preventing major depression. However, iCBT faces lower adherence and higher dropout compared with face-to-face therapy, necessitating strategies to enhance engagement. Automated and human encouragement have been suggested to reduce dropouts. Chatbots, delivered via familiar messaging services, may provide human-like, immediate, and personalised encouragement to improve adherence without human guidance. This study tests whether adding a non-therapeutic, supportive chatbot to iCBT increases programme completion among workers with subthreshold depression.

Prior research shows iCBT is effective for subthreshold depression and prevention of major depression in adults and workers. Digital interventions often have lower adherence than face-to-face treatments, with reported dropout challenges. Approaches to improve adherence include periodic encouraging emails, adaptive user interfaces, and guidance (human or automated). The use of conversational agents in digital mental health has grown; RCT evidence (e.g., Woebot) suggests reduced dropout with automated agents, and supportive feedback features can improve engagement by about 20%. Yet, no study has isolated the effect of a chatbot specifically designed to encourage adherence (without delivering therapeutic content) on iCBT engagement and dropout. This gap motivated development of a personalised, friendly chatbot delivered via a widely used messaging app.

Design: Open-label, stratified block randomised controlled trial with two arms: iCBT plus chatbot (iCBT+chatbot) versus iCBT alone. The trial followed CONSORT guidelines. Participants: Full-time employees of Sony Group Corporation/Sony Corporation in Japan, aged 20–60, smartphone owners, consenting to use the iCBT app, Fitbit (Charge 4), Fitabase, and LINE messaging. Exclusions: inability to read/write Japanese; current mental health treatment; PHQ-9 ≥15 or 10–14 with item 9 score 2–3; plan to retire/change jobs during participation. Recruitment: April 2022; 334 screened; 149 eligible attended orientation, provided e-consent, and completed psychoeducation. Safety: Intervention discontinued for PHQ-9 ≥15 or 10–14 with item 9 score 2–3 sustained over 3 weeks; guidance to contact psychological/health services provided. Interventions: iCBT via the smartphone app ‘Resilience Training SE (Sony Edition),’ adapted from a student app to worker context. Components: psychoeducation (PE) followed by behavioural activation (BA), self-monitoring (SM), cognitive restructuring (CR), assertiveness training (AT), and problem-solving (PS), each targeted to one week. Each component included PE and a practice worksheet. Weekly self-check (PHQ-9) prompts; automated emails for missed self-checks; safety emails for high PHQ-9. Chatbot: ‘EPO,’ a cloud-like character, delivered personalised encouraging messages via LINE every morning and evening for 8 weeks, based on each participant’s iCBT progress retrieved from servers. Database of ~300 messages included progress-based encouragement, frequency-adjustment surveys, daily engagement prompts, and semi-open-ended questions to increase engagement; character stickers enhanced human-like interaction. Fitbit: All participants wore the device for 8 weeks; data collection identical across arms. Outcomes: Primary—completion rate of the five post-PE components within 8 weeks (56 days), defined as reading all lessons and completing the PS worksheet before the epilogue. Secondary—changes from baseline to week 8 in PHQ-9 (depression; weekly), GAD-7 (anxiety), CBT skills (SM, BA, CR, AT, PS), SWLS and WHO-5 (well-being), Presenteeism (HPQ), WSAS (functioning), and UWES (work engagement). Measurement schedule: GAD-7, CBT skills, SWLS, WHO-5, WSAS, UWES at baseline, week 4, week 8; Presenteeism at baseline and week 8. Sample size: Assuming a 20% improvement in completion with chatbot support, α=0.05 (two-sided), power=80%, required n=124; recruited n=150 anticipating nonattendance/declines. Randomisation: Permuted block randomisation stratified by pre-assessment PHQ-9 (≤4 vs ≥5). Allocation by automated system using consent timestamp; researchers (except CRC) concealed from assignment. Masking: None (participants and researchers unblinded). Analysis: Full analysis set (FAS) with intention-to-treat. Primary comparison of completion rates via χ² test (two-sided α=0.05). Secondary analyses: Mixed-effects model for repeated measures (MMRM); PHQ-9 analyses restricted to baseline PHQ-9 ≥5 to assess subthreshold depression impact. Fixed effects: intervention (chatbot vs none), time (nominal), age, baseline PHQ-9, and intervention×time interaction; unstructured covariance. Standardised mean differences (SMD) computed using baseline SD. The same MMRM approach applied to other secondary outcomes.

Participants: 149 randomised (74 iCBT+chatbot; 75 iCBT). FAS included 143 (excluded: 6 due to app login/psychoeducation issues). Primary outcome included 142 (one discontinued for high PHQ-9 per protocol). Baseline characteristics were balanced. Primary: Completion at 8 weeks—iCBT+chatbot 34.8% (24/69; 95% CI 23.5 to 46.0) vs iCBT 19.2% (14/73; 95% CI 10.2 to 28.2); risk ratio 1.81 (95% CI 1.02 to 3.21); risk difference 15.6% (95% CI 1.19 to 30.0); p=0.0358. Secondary: PHQ-9 change at week 8—iCBT+chatbot -2.21 (95% CI -3.21 to -1.22; ES=-0.75) vs iCBT -2.30 (95% CI -3.30 to -1.30; ES=-0.78); between-group mean difference 0.08 (95% CI -1.33 to 1.5; ES=0.03), not significant. Both groups improved in CBT skills (except PS), GAD-7, WHO-5, SWLS, and Presenteeism. UWES improved significantly more in the iCBT group than the iCBT+chatbot group (p<0.05). Adverse events: one hospitalization due to a traffic accident, judged unrelated; no serious adverse events attributable to the study.

A friendly, human-like chatbot delivering personalised, progress-based encouragement via a familiar messaging app significantly increased iCBT completion within 8 weeks. The enhanced adherence may stem from tailored pacing and regular prompts aligned to each participant’s progress, with more than half of chatbot users reporting that messages triggered their iCBT use. When the messaging ceased after 8 weeks, the between-group difference in completion attenuated by 10 weeks, suggesting the chatbot’s role in pacing within the intervention window. Despite higher adherence, depression and anxiety improvements at 8 weeks were similar across groups. Possible reasons include early benefits from psychoeducation and behavioural activation, and supportive effects of weekly self-checks and Fitbit life-log visualizations, which may independently improve mood and engagement. The overall completion rate was lower than expected, likely reflecting the intensity and breadth of content relative to participants’ busy schedules and the 2-month timeframe. These findings highlight the importance of optimizing iCBT dosing and duration for working populations and suggest that chatbots can effectively support adherence without human guidance.

This first RCT focused on a supportive, non-therapeutic chatbot add-on to iCBT demonstrates that personalised, automated messages can increase programme completion among workers with subthreshold depression. While adherence improved, short-term clinical outcomes (PHQ-9, GAD-7) improved similarly in both arms. Future research should refine iCBT structure and workload, evaluate long-term adherence and clinical impacts, and enhance chatbot personalisation to capture user characteristics and deliver more tailored support that may augment clinical outcomes.

The iCBT programme’s scope (psychoeducation plus five components over 2 months) was perceived by many as too demanding, contributing to lower-than-expected completion. The in-house sample of Sony employees may have higher digital literacy than the general population, limiting generalisability. The trial was not designed to test the chatbot’s efficacy for improving subthreshold depression symptoms; thus, clinical effects of the chatbot remain to be determined.