1\. Study Design and Randomization MechanismThis study utilizes a Pre-Post Parallel Intervention Design. To enhance internal validity, minimize selection bias, and account for unit-level characteristics, a Stratified Cluster Randomization approach is employed. Clinical units are stratified into three tiers based on assessed work intensity: High Intensity (e.g., ICUs and ED), Medium Intensity (e.g., medical and surgical wards), and Low Intensity (e.g., psychiatric and rehabilitation wards). Within each stratum, the nursing unit or ward serves as the cluster. Clusters are then randomly assigned (1:1 ratio) to either the Intervention Group or the Control Group, ensuring that all nurses within a single unit belong to the same study arm. The planned total sample size is N=120, distributed across approximately 18 clusters.2. Intervention Protocol (4-Week Period)The intervention period spans four weeks (28 days). The core tool is a customized GPT dialogue model integrated with the LINE Official Account (LINE OA) platform, widely used by the target population.Intervention Group (Interactive AI Support)Participants receive a two-part daily intervention: a daily push message and access to a two-way AI interaction. They are encouraged to actively engage in free-form psychological support dialogue with the LINE GPT assistant. The AI's response logic is rigorously defined via Prompt Engineering to focus on clinical work stress, emotional coping, and resilience-promoting strategies. Crucially, the AI is programmed with ethical guardrails to gently redirect users back to the clinical stress axis if they deviate to non-work related or sensitive topics, and strictly prohibit the provision of any medical, diagnostic, or therapeutic advice.Control Group (Static Message Control)Participants in the control group receive daily static text messages. The content themes are identical to the push messages received by the intervention group, but the channel lacks any interactive AI response capability.3. Study Procedures and Data Collection TimelineThe entire study spans a minimum of six weeks, structured into three distinct phases.The process begins in Weeks 1-2 with recruitment and baseline assessment. Following informed consent, participants are oriented to the study platform (LINE OA) and complete the Pre-Test Questionnaire. This baseline assessment, which takes approximately 25-30 minutes, captures all demographic/occupational variables and the initial scores for all primary and secondary outcome scales (ProQOL, GSE, PSS-10, BRS).Immediately following baseline assessment, the 4-week intervention period commences, spanning from approximately Week 3 through Week 6. Throughout this period, both the intervention and control groups receive their respective daily messages. To track subjective engagement and experience, all participants are required to complete a brief "AI Use Feedback Simple Form" weekly, which takes approximately 5 minutes.Upon completion of the four weeks, the study enters the Post-Test phase at the end of Week 6. All participants complete the same battery of outcome scales (re-measurement of ProQOL, GSE, PSS-10, BRS) and the AI Use Satisfaction Questionnaire. The Intervention Group is additionally requested to complete an Open-ended Feedback Questionnaire to provide rich qualitative data on their experience with the GPT assistant.4. Statistical Analysis PlanData will be anonymized using unique study IDs prior to analysis. Due to the stratified cluster randomization, the analysis will employ methods suitable for clustered data structures to account for Intracluster Correlation (ICC). The Primary Analysis will utilize Generalized Estimating Equations (GEE) to model the change in outcomes over time, specifically testing the "Time $\\times$ Group" interaction to assess the intervention effect. The nursing unit/ward will be specified as the cluster variable in the GEE model. Linear Mixed-Effects Models (LMM) may be used for supplementary or sensitivity analysis. Covariates such as nursing tenure and baseline stress levels will be included in the models for adjustment and exploratory analysis. Qualitative data from the open-ended feedback will be analyzed using Content Analysis to categorize emerging themes.
