Assessing the Feasibility, Usability, Acceptability, and Efficacy of an AI Chatbot for Sleep Promotion:...

Poor sleep is a concerning public health problem in the United States. Previous sleep interventions often face barriers such as high costs, limited accessibility, and low user engagement. Recent advancements in artificial intelligence (AI) technologies offer a novel approach to overcoming these limitations. In response, our team developed a prototype AI sleep chatbot powered by a large language model to deliver personalized, accessible sleep support.

This study aimed to examine the feasibility, usability, acceptability, and preliminary efficacy of the AI chatbot for sleep promotion.

We conducted a quasi-experimental, single-group study with adults in the United States aged 18 to 75 years who self-reported poor sleep. The chatbot was integrated into a commercially available messaging app. Participants were asked to engage with a virtual sleep therapist via texting over 2 weeks. The chatbot provided ongoing, individualized sleep guidance and adapted recommendations based on participants' prior conversations. Feasibility, usability, and acceptability were descriptively summarized. Sleep was assessed using questionnaires before and after the intervention.

Of the 107 adults who enrolled in the study, 88 (82.2%) completed chatbot registration. Among these 88 participants, 65 (73.9%) initiated interactions, and 44 (50%) completed the 2-week intervention. The final analysis included 42 adults (mean age 36, SD 11 years; n=12, 28.6% male). On average, participants engaged with the chatbot for 58 (SD 42) minutes, with each chat session lasting approximately 9 (SD 6) minutes. Most reported favorable experiences with the chatbot. The average usability score was 85.2 (SD 10.7) out of 100, which was well above the benchmark of 68. The chatbot was rated as highly acceptable, with a satisfaction score of 27.3 (SD 4.1) out of 32. All participants perceived the chatbot as effective, with ratings ranging from "slightly effective" to "extremely effective." The preliminary evidence showed improved sleep outcomes after chatbot use: total sleep time increased by 1.4 hours (P<.001); sleep onset latency decreased by 30.9 minutes (P<.001); sleep efficiency increased by 7.8% (P=.007); and scores improved for perceived sleep quality (mean difference [MD] -5.4; P<.001), insomnia severity (MD -7.9; P<.001), daytime sleepiness (MD -4.7; P<.001), and sleep hygiene skills (MD -13.2; P<.001). No significant change was observed in sleep environment (MD -1.1; P=.16).

Our AI chatbot demonstrated satisfactory feasibility, usability, and acceptability. Improvements were observed following chatbot use, although causality cannot be established. These findings highlight the potential of integrating state-of-the-art large language models into behavioral interventions for sleep promotion. Future research should include objective sleep measurements and conduct randomized controlled trials to validate the study findings. If confirmed, this AI chatbot could be scaled to support sleep health on a broader level.