The study aims to evaluate (1) the feasibility and acceptability of ChatACP and (2) the preliminary efficacy of ChatACP on family members’ readiness for ACP, engagement in ACP, and completion of ACP activities, in comparison to those who received a self-learning handout, at the postintervention time point and 3-month follow-up.

This will be an explanatory sequential mixed methods study including a pilot cluster RCT followed by a posttrial qualitative study with a postintervention time point and 3-month follow-up, as shown in the trial flowchart (Figure 1), over 2 years. To prevent contamination, we will enroll 6 care homes (N=60 family members of residents) and randomly assign 3 homes to receive ChatACP and 3 to receive the ACP self-learning handout.

The inclusion criteria are as follows: (1) age of 18 years or above; (2) providing care for an older adult (aged 65 years or older) living in a care home as a primary family caregiver (including family members in the traditional sense and the guardians and persons close to or significant to the older adult); (3) ability to read, write, and communicate in Cantonese or Mandarin; (4) the resident (care recipient) having a prefrail or frail status as defined by a score of 2 or above on the FRAIL scale [21]; (5) having a mobile device to receive instant messages and access to a cellular network; and (6) having engaged in minimal or no prior EOL care discussions with the care home resident as defined by a self-reported status of being in the precontemplation or contemplation stage regarding ACP discussions with the resident using the ACP staging algorithm illustrated in Figure 2 [22].

The exclusion criteria are as follows: (1) moderate to severe cognitive impairment, defined by an Abbreviated Mental Test score [23] of 5 or less, and inability to give consent; and (2) communication problems (eg, deafness or aphasia).

The care home will be the unit of randomization, and 6 homes will be pair matched based on type of home (public or private). An independent statistician will randomly assign one home from each pair to the intervention arm using a computer-generated algorithm.

This study received ethics approval from the University of Hong Kong and Hospital Authority Hong Kong West Cluster ethics review board (UW25-567) in December 2025. Written informed consent will be obtained from all participants, ensuring their right to withdraw per the Declaration of Helsinki. Participant data will be deidentified, stored securely, and accessed only by authorized personnel to ensure privacy and confidentiality.

Recruitment will be conducted in accordance with the protocol approved by the relevant ethics committee. Family members will be recruited from 6 care homes. All the eligible family members from the participating homes will be invited to join the study. Our RA will first contact the superintendent on their interest in joining the study. After obtaining agreement from the administration, we will promote the study in the homes during family events and using promotional posters. In addition, our RAs will work with the home staff (nurses, social workers, or other care providers) on identifying eligible family members. Staff at nursing homes will first obtain a verbal expression of interest from the family members before passing their contact information to our RA for recruitment. Regarding the qualitative interviews, family members of different sexes, relationships to the older adults, levels of intervention engagement, and motivation stages for ACP, as well as residential care home staff (approximately n=20 depending on the informational redundancy), will be invited for an individual semistructured interview upon completion of the intervention and at the 3-month follow-up.

The preliminary efficacy outcomes will be collected among participants in both arms at baseline (T0; before the intervention), after the intervention (T1), and at 3 months (T2). The feasibility of the proposed study will be evaluated using the feasibility and acceptability outcomes at the end of data collection. Following the completion of the T2 survey, participants will receive HKD 200 (HKD 1=US $0.13 as of June 5, 2026) supermarket coupon as an incentive.

Data will be analyzed in SPSS software (IBM Corp). Data will be screened for completeness, and the RA will contact participants if there are any missing data. Descriptive statistics will be used to summarize participant characteristics and study outcomes.

The preliminary effect outcomes (family members’ ACP readiness, engagement, and completion of ACP activities) will be analyzed using a generalized linear mixed-effects model to understand the treatment differences at each time point. Because data are nested within each care home (cluster), observations within the same cluster may be correlated, violating the assumption of independence. To account for this clustering effect, we will integrate a random intercept for care homes in the generalized linear mixed-effects model. Given that our study is not powered to examine effectiveness, the outcomes will be reported with the standardized effect size (scores for ACP readiness and engagement) or ORs (completion of ACP activities) with a 95% CI. A Cohen d of 0.2, 0.5, and 0.8 and an OR of 1.5, 2.5, and 4 will indicate small, medium, and large effect sizes, respectively.

The feasibility and acceptability outcomes will be assessed to determine the progression to a full multisite RCT based on the predefined progression cutoffs outlined in the outcome measurements section, which include a 50% recruitment rate, 70% retention rate, 90% intervention fidelity, 50% intervention engagement rate, 78% content usability, 70% acceptability, and 100% safety based on prior review [30].

For the qualitative evaluation, the voice-recorded interviews will be transcribed verbatim and then analyzed according to the thematic analysis proposed by Braun and Clarke [31]. The NVivo software (Lumivero) will be used to manage the collected qualitative data and support the analysis. To ensure rigor, credibility, transferability, dependability, and confirmability will be assessed. Findings from both the qualitative and quantitative portions after completing the qualitative study will be synthesized. A sequential interpretive integration of both types of data will be undertaken at the interpretation stage.

This represents, to our knowledge, one of the first experimental studies to examine an LLM-enhanced ACP consultation, addressing a critical gap in tailoring the presentation of ACP content to a layperson-friendly literacy level. In addition, this study offers a potential solution to address the phenomenon of information overload [32] using LLM-based consultation. The potential application of findings from the ChatACP pilot study could have significant implications for patients, their families, and the health care system.

This protocol describes a pilot cluster RCT with an explanatory sequential mixed methods design to evaluate the feasibility, acceptability, and preliminary efficacy of ChatACP, a novel digital intervention designed to empower family caregivers of frail older adults in Hong Kong residential care homes to engage in ACP. ChatACP addresses key cultural barriers in Chinese contexts—where older adults often defer EOL decisions to adult children, yet families frequently resist open discussions due to preferences for aggressive treatment—by progressing caregivers along the TTM stages from precontemplation or contemplation toward preparation and action. The intervention combines a 10-day digital education package (culturally adapted infographics, short artificial intelligence–generated videos, and a 24/7 customized LLM-based chatbot) with optional nurse-led motivational interviewing consultations. The active control—a standard self-learning ACP handout from the Hong Kong Hospital Authority—represents usual care, enabling a pragmatic comparison of whether theory-driven, interactive digital support can yield superior outcomes in ACP readiness, engagement, and behavioral completion.

The cluster randomized design, with care homes as the unit of randomization and pair matching by public or private status, is a core strength that minimizes the contamination risks common in shared residential settings. With 6 clusters, the modest sample (N=60) is appropriate for a pilot focused on feasibility rather than definitive efficacy. The mixed methods approach further enhances rigor: predefined quantitative progression criteria (eg, recruitment of ≥50%, retention of ≥70%, fidelity of ≥90%, engagement of ≥50%, usability of ≥78%, acceptability of ≥70%, and safety of 100%) provide objective benchmarks for advancing to a full-scale multisite RCT, whereas postintervention qualitative interviews will capture nuanced experiences, perceived burden, logistical challenges, and suggestions for refinement. The intervention’s innovation lies in its integration of accessible, low-literacy digital tools tailored to TTM stages, including interim self-reported staging to trigger personalized phase 2 coaching. The use of validated Chinese-language measures (ACP readiness scale and ACP engagement survey for SDMs) and binary behavioral outcomes ensures culturally relevant, reliable assessment. The chatbot’s fine-tuning on local guidelines, stakeholder-informed frequently asked questions, and fifth-grade reading level responses offer scalable, on-demand support, aligning with calls for flexible, family-centered ACP approaches.

Despite these strengths, there are important limitations. As this is a pilot study, the small sample and limited number of clusters reduce statistical power and increase vulnerability to baseline imbalances or cluster effects. Generalizability remains restricted to Hong Kong care home contexts, where family dynamics, digital access, and institutional practices may differ from those in mainland China or other Asian settings. Self-reported outcomes, while validated, are prone to social desirability bias or recall inaccuracies, especially on sensitive EOL topics. Digital delivery requires smartphone ownership, network connectivity, and technological comfort, potentially excluding less technologically savvy caregivers and raising equity concerns in an aging population. Recruitment relies heavily on care home staff cooperation and family member availability, which can be unpredictable amid competing caregiving demands or resident health fluctuations.

These considerations highlight the pilot’s primary role: generating critical data on the operational feasibility, participant engagement with digital components, and cultural acceptability of an LLM-augmented intervention. If the progression criteria are achieved, the findings will guide protocol refinements—such as supplementary support for low-engagement users, expanded offline options, or additional content modules—for a larger trial powered to assess longer-term impacts on ACP documentation, provider communication, and concordance between care preferences and actual EOL care. By addressing family empowerment through innovative, theory-informed digital strategies, this work contributes to developing culturally congruent ACP interventions suited to resource-constrained care home environments in aging societies.

The ChatACP intervention has the potential to improve the quality of care for care home residents by aligning medical and nursing care with residents’ preferences and values. By empowering family members to engage in ACP discussions, residents’ individual wishes can be better integrated into their health care plans, resulting in more patient-centered care. This approach may reduce unnecessary hospitalizations and invasive, often futile medical procedures while alleviating the emotional burden on family members who may otherwise struggle with decision-making during critical moments.

By adopting this low-cost digital ACP empowerment intervention, nursing homes may reduce the burden on health care staff and streamline communication with family members. The findings may provide valuable insights into the use of chatbot technology in enhancing nursing consultations and promoting ACP engagement. If proven effective, the ChatACP intervention could be integrated into routine nursing practice in care homes and other health care settings, leading to increased ACP awareness and engagement among family members and health care professionals. This could foster a more proactive approach to ACP and EOL care. Additionally, the findings could inform policymakers about the potential benefits of incorporating messaging and chatbot systems into life-and-death education in nursing homes and guide the development of similar interventions for other aspects of health care management, such as chronic disease management or mental health support.