Abstract
<sec> <title>BACKGROUND</title> <p>Behavioral sleep problems are common in preschool children and respond well to parent-mediated intervention, but sleep promotion in China is typically generic sleep-hygiene education that is neither individualized nor scalable. Large language models (LLMs) could generate family-specific advice at scale, but unconstrained generation risks factually incorrect content. Retrieval-augmented generation (RAG) grounds output in a curated knowledge base; its feasibility for parent-facing sleep promotion is untested.</p> </sec> <sec> <title>OBJECTIVE</title> <p>This study aimed to describe the digital adaptation of an Intervention Mapping–derived behavioral sleep promotion program into the Sleep Island (安睡岛) WeChat mini-program and to evaluate its feasibility, acceptability, usability, and perceived usefulness</p> </sec> <sec> <title>METHODS</title> <p>Digital adaptation was followed by a prospective single-arm feasibility evaluation. Behavioral targets of the source protocol were mapped onto an electronic sleep diary, sleep trend visualization, a sleep behavior report, and an individualized recommendation set (the Sleep Silkbag). Recommendations were generated by an LLM (Qwen-Max) grounded through RAG in a knowledge base built from the protocol and pediatric sleep-health guidance, each reviewed by the research team before release. Parents of children aged 3-6 years were recruited over 2 weeks; they registered, completed a baseline assessment, recorded 7 consecutive nights, and reviewed the feedback. After 1 week they completed a 45-item questionnaire covering 6 domains rated from 1 to 5, plus items on engagement, burden, and technical problems</p> </sec> <sec> <title>RESULTS</title> <p>Of 30 parents recruited, 21 completed the assessment (70.0%, 95% CI 52.1%-83.3%; 95.5% of the 22 who consented), attrition occurring almost entirely at study entry. All 6 domains scored above 4.4: usability highest (mean 4.75, SD 0.32), sleep-diary feasibility lowest (mean 4.44, SD 0.37). Onboarding took under 15 minutes for 18 of 21 (86%) participants, all completed each diary entry within 5 minutes, and 17 of 21 (81%) recorded all 7 nights. The lowest-rated item was remembering the diary daily (mean 3.95); 11 of 21 (52%) cited forgetting and 9 (43%) competing demands, whereas none cited unclear questions or lack of value. All 21 viewed the trend page, 20 of 21 (95%) read the recommendations, and all rated the mini-program acceptable; 2 of 21 (10%) reported a technical problem. Items on parental knowledge and confidence (4.62-4.81) scored higher than items on observed child change (4.19-4.38). Across 19 children and 123 nights, screen-based video occurred on 54% of nights, nearly matching shared reading (57%).</p> </sec> <sec> <title>CONCLUSIONS</title> <p>The intervention was feasible and highly acceptable, and parents judged the retrieval-grounded, human-reviewed recommendations comprehensible, actionable, and consistent with their parenting values. The principal constraint was remembering the daily diary, a cueing rather than a burden problem, making automated reminders the clearest refinement priority. Perceived benefit concentrated in parental knowledge and confidence rather than child sleep change, as expected after one week.</p> </sec> <sec> <title>CLINICALTRIAL</title> <p>Open Science Framework q8ypn; https://osf.io/q8ypn</p> </sec>