Effectiveness of an Artificial Intelligence Health Education Accurately Linking System (AI-HEALS) for the intervention of chronic kidney disease: a randomised controlled trial protocol

Abstract

Introduction

Chronic kidney disease (CKD) is a major global health challenge with high morbidity and socioeconomic burden. Despite medical advances, long-term prognosis remains poor, highlighting the need for sustained self-management and health education. Traditional face-to-face care is limited by inefficient follow-up and insufficient attention to psychosocial and behavioural determinants, making it inadequate for long-term support. Artificial intelligence and mobile health offer personalised sustainable health education interventions. Accordingly, we developed the Artificial Intelligence Health Education Accurately Linking System (AI-HEALS) to support CKD self-management.

Methods and analysis

This is a single-centre, randomised, assessor-blinded, parallel-group controlled trial. A total of 74 adults with stages 1–5 CKD will be randomly allocated in a 1:1 ratio to either the control group, receiving standard care or the intervention group, receiving standard care plus the AI-HEALS programme for 3 months. The primary outcome is health-related quality of life, assessed using the physical component summary (PCS) score of the 12-item Short Form Health Survey (SF-12). The primary endpoint is the between-group difference in change in SF-12 PCS score from baseline to the end of the 3-month intervention. Secondary outcomes include serum creatinine levels, urinary protein creatinine ratio, anxiety, depression, quality of sleep, self-efficacy, dietary behaviour and eHealth literacy. Outcomes will be assessed at baseline, at the end of the intervention and at 3, 6 and 9 months after completion of the intervention. Linear mixed-effects models will be used to analyse repeated measurements over time. The findings of this study will provide evidence regarding the effectiveness of AI-HEALS in improving health-related quality of life and supporting self-management among patients with CKD.

Ethics and dissemination

This study protocol has been approved by the Biomedical Research Ethics Committee of West China Hospital of Sichuan University (Ethics Approval Number: 2024–2106). The findings are expected to be disseminated through peer-reviewed journal publications, academic conference presentations and other forms of scholarly communication.

Trial registration number

ChiCTR2500099413.

STRENGTHS AND LIMITATIONS OF THIS STUDY.

• The intervention integrates a large language model with WeChat to provide personalised health education and self-management support through a commonly used digital platform.

• The study assesses a range of outcomes, including health-related quality of life, clinical measures, behavioural factors and psychosocial outcomes, with follow-up extending to 9 months after completion of the intervention.

• Blinding of participants and intervention providers is not feasible because of the nature of the intervention. However, outcome assessors and data analysts will remain blinded to group allocation.

Background

Chronic kidney disease (CKD) is a major public health concern affecting hundreds of millions of people worldwide.¹ According to the WHO, the global prevalence of CKD among adults is approximately 13.4%.¹ In China, recent epidemiological data suggest a prevalence of approximately 8.2%, indicating that CKD affects tens of millions of adults nationwide.² Beyond its high prevalence, CKD is associated with substantial morbidity and healthcare burden. As the disease progresses, patients may develop a range of complications, including cardiovascular disease, disorders of mineral and bone metabolism and neurological complications, all of which may adversely affect quality of life and increase the risk of mortality.^3–5 Patients who progress to kidney failure may require long-term kidney replacement therapy, such as dialysis, which can further impair quality of life and impose considerable financial costs.⁶ The prolonged treatment and care needs associated with CKD also place substantial economic and emotional demands on patients and their families and contribute to sustained pressure on healthcare systems.⁷ Given the high costs of long-term CKD management and the limited availability of curative treatment, CKD represents a considerable and ongoing burden on public health resources.^{5 8} Timely and continuous management is therefore important for slowing disease progression, reducing complications and delaying or preventing the need for dialysis or kidney transplantation, while helping to preserve patients’ quality of life.^9–11

Current approaches to CKD management include pharmacological treatment, lifestyle modification and patient education.^9–11 Pharmacological treatment focuses on controlling blood pressure and blood glucose levels and managing CKD-related complications, such as anaemia and mineral and bone disorders.^{12 13} Lifestyle modification, particularly dietary management and regular physical activity, also plays an important role in slowing disease progression.^14–16 Patient education is another key component of CKD management, providing patients with essential knowledge about kidney function, disease progression and warning signs that require medical attention.¹⁷ It also supports the development of self-management skills, including medication management, dietary adjustment, blood pressure monitoring and other lifestyle changes. By improving health literacy and self-efficacy, patient education may help patients become more actively involved in their care, make informed decisions and adhere to recommended treatment and self-management strategies.

Despite these approaches, several challenges remain. CKD requires sustained self-management over the long term,^{18 19} yet many patients continue to have unmet needs related to disease knowledge, self-efficacy and behavioural support.^{18 19} Previous studies suggest that effective self-management may help slow disease progression, improve symptom control, reduce treatment-related problems and lower the risk of complications and hospital readmission.^{18 19} It may also reduce the burden of ongoing care and contribute to better physical and psychological well-being and quality of life. However, conventional face-to-face interventions are often limited by constraints in healthcare resources, continuity and accessibility, particularly for patients living in remote areas who may have difficulty obtaining timely and sustained support.

Recent advances in digital health and artificial intelligence (AI) have created new opportunities for CKD self-management. Mobile and internet-based interventions have been used to support patient education, follow-up and lifestyle management, with some studies reporting improvements in self-management and disease-related outcomes.^{20 21} Although findings have not been uniformly significant, these studies suggest that digital interventions may help extend support beyond conventional face-to-face care. More recently, large language models (LLMs) have shown potential for providing accessible and personalised support in chronic disease management, including CKD.^22–24

In this context, Artificial Intelligence Health Education Accurately Linking System (AI-HEALS) was developed as a WeChat-based health education and self-management support system for people with CKD. By using a widely accessible mobile platform, the system is intended to provide patients with ongoing disease-management support outside clinical settings. The intervention is informed by the Health Action Process Approach (HAPA) and the Multi-Theory Model (MTM), both of which have been used to support health behaviour change in chronic disease management. Previous studies have applied HAPA-based interventions in conditions such as hypertension, diabetes and chronic obstructive pulmonary disease,^25–27 while MTM has also been used to guide interventions targeting physical activity among people with diabetes.²⁸ Together, these frameworks provide a basis for addressing health behaviours, self-management and psychosocial factors within the AI-HEALS intervention. The system includes knowledge-based question-and-answer services, monitoring of selected physiological and behavioural data, personalised reminders and tailored health information. Through these functions, AI-HEALS is designed to provide more continuous and individualised support for CKD self-management.

In summary, this randomised controlled trial aims to evaluate the effectiveness of AI-HEALS in improving health-related quality of life and self-management among adults with CKD, with a focus on the following aspects:

Primary objective

To determine whether AI-HEALS improves health-related quality of life compared with standard care among adults with CKD.

Secondary objective

To evaluate the effects of the intervention on secondary outcomes, including serum creatinine levels, urinary protein creatinine ratio, anxiety, depression, quality of sleep, self-efficacy, dietary behaviour and eHealth literacy

Methods

Study design

This study is designed as a single-centre, randomised, assessor-blinded, parallel-group controlled trial. The study population consists of patients diagnosed with CKD at a public comprehensive hospital in Chengdu, Sichuan Province, China, making it a single-centre study. The study is planned to be conducted from January 2025 to October 2026, and recruitment of participants is scheduled from May 2025 to July 2026. The overall research process and group allocation procedure are shown in figure 1.

Figure 1. Overall research process and group allocation procedures. AI-HEALS, Artificial Intelligence Health Education Accurately Linking System; CKD, chronic kidney disease.

Randomisation and blinding

In this trial, randomisation will be performed using a computer-generated random sequence, with participants allocated in a 1:1 ratio to either the intervention group or the standard care group. The randomisation sequence will be generated before participant admission. On enrolment, each participant will be assigned a unique study identification number corresponding to the randomisation sequence. Participants will then be admitted to separate wards according to group allocation: those in the intervention group will be admitted to Ward A and those in the control group to Ward B, in order to minimise potential contamination between the two groups.

To ensure allocation concealment, the randomisation sequence will be managed by an independent member of the research team who is not involved in participant recruitment or intervention delivery. Group allocation will not be disclosed to the recruitment or intervention teams before a participant has been enrolled and assigned. After allocation, the independent team member will inform the intervention team of the participant’s group assignment so that the appropriate intervention can be delivered.

Given the nature of the AI-HEALS intervention, blinding of participants and nursing staff involved in intervention delivery is not feasible. Several measures will therefore be taken to minimise contamination and potential bias. First, a private WeChat group will be established for participants in the intervention group, with access restricted to authorised study participants using designated accounts and passwords. Second, participants will be instructed not to share intervention materials with individuals outside the group. Third, general health information that is readily available from public sources will not be routinely posted in the group, thereby reducing unnecessary overlap with information accessible to the control group. Fourth, the research team will review group chat records weekly to identify any potential sharing of intervention materials outside the designated group. Fifth, outcome assessment, data extraction and statistical analysis will be conducted by researchers blinded to group allocation. Finally, sensitivity analyses will be undertaken to assess the potential influence of differential response and attrition. These will include: (1) comparison of baseline characteristics between early and late responders; and (2) where differential dropout between groups is observed, the use of appropriate methods, such as multiple imputation or inverse probability weighting, to assess the robustness of the findings.

Study participants

Patients with CKD will be screened for eligibility according to the following criteria.

Inclusion criteria

1. Aged ≥18 years, with no restriction on sex.

2. Clinically diagnosed by a physician with CKD stages 1–5.

3. Own a smartphone, are able to use WeChat and are willing to receive the intervention through this platform.

4. Able to understand the study procedures, provide written informed consent and willing to comply with the study protocol throughout the trial.

Exclusion criteria

1. Currently receiving kidney replacement therapy, including haemodialysis, peritoneal dialysis or kidney transplantation.

2. Presence of severe cardiovascular, cerebrovascular, renal or other major organ dysfunction.

3. Concurrent participation in another similar research study.

4. Unwilling or unable to cooperate with the study procedures, or otherwise considered unsuitable for participation by the investigators.

Withdrawal criteria

1. Participants may withdraw their consent and discontinue participation at any time without giving a reason.

2. Participants may be withdrawn from the study by the investigators if a medical condition develops during the study that could place them at risk if participation continues.

3. Participants may also be withdrawn if information obtained after enrolment indicates that they did not meet the eligibility criteria, if they subsequently meet an exclusion criterion or if substantial non-adherence to the study protocol may compromise participant safety or the integrity of key study data.

4. Any other circumstance in which the investigators consider withdrawal necessary in the interest of participant safety or study integrity.

Sample size

The sample size was calculated using G*Power software V.3.1.9.4. The primary outcome for sample size estimation was the between-group difference in change in health-related quality of life, assessed by the physical component summary (PCS) score of the 12-item Short Form Health Survey (SF-12) from baseline to the end of the 3-month intervention.

Based on previous findings reported by Li and colleagues,²⁹ an effect size of 0.68 was assumed for the expected difference between groups. A two-sided significance level of 0.05, statistical power of 0.80 and an allocation ratio of 1:1 were applied. The estimated minimum sample size was 58 participants, with 29 participants required in each group. Considering an anticipated dropout rate of approximately 20%, the final recruitment target was determined as 74 participants (37 participants per group).

Recruitment

Patients with CKD who meet the eligibility criteria will be recruited for this study. Potentially eligible patients will be identified and screened by the clinical team at the time of hospital admission and invited to participate. A member of the research team will then provide a face-to-face explanation of the study, including its purpose, procedures, potential benefits and possible risks. Eligibility will be further confirmed through assessment of relevant medical history, clinical information and CKD-related symptoms. Patients will be given sufficient opportunity to ask questions before deciding whether to participate. Written informed consent will be obtained from all participants before enrolment, and only those who provide consent will be formally enrolled in the study.

Informed consent

The informed consent process will be conducted in accordance with the approved study procedures and relevant ethical requirements. Participants who meet the eligibility criteria will receive a detailed participant information sheet describing the purpose of the study, study procedures, potential risks and benefits and measures for protecting personal data. The information provided will also explain how participants’ data will be collected, stored, used and protected. Study data will be stored on secure, encrypted servers, with access restricted to authorised members of the research team.

Written informed consent will be obtained from all participants using the approved consent form (see online supplemental informed consent form). The consent materials and procedures have been approved by the Biomedical Ethics Review Committee of West China Hospital, Sichuan University (Approval No. 2024–2106). Participants will be given sufficient time to review the study information, ask questions and consider whether they wish to participate before providing written consent. They will be informed that participation is voluntary and that they may withdraw from the study at any time without penalty or loss of entitled care or benefits. Participants will also be provided with contact details for the research team should they have any questions during the study and will be informed of any new information that may be relevant to their continued participation.

Participants who choose to discontinue the intervention will not be transferred to the control group. Their subsequent study status and the handling of data collected before withdrawal will follow the procedures specified in the informed consent form and the approved study protocol. Where appropriate, and with the participant’s agreement, a semi-structured interview may be conducted to explore the reasons for discontinuation. During the informed consent process, participants will be informed of the options available should they later decide to discontinue the intervention or withdraw from the study.

Interventions

Control group: standard care

Participants in the control group will receive standard CKD care provided by the hospital’s multidisciplinary clinical team. Standard care will include routine clinical assessment, pharmacological management, laboratory and diagnostic testing, nurse-led health education, discharge counselling and follow-up in accordance with usual clinical practice for patients with CKD.

During hospitalisation, participants will be assessed by the attending physicians for kidney function, CKD-related complications and treatment needs. Management of blood pressure, blood glucose, medications and other CKD-related conditions will be adjusted according to clinical indications. Participants will also undergo routine laboratory and diagnostic assessments, including kidney function tests, urine tests and any additional investigations required as part of their clinical care.

Nurse-led health education will be provided as part of routine inpatient care. The content will cover basic CKD knowledge, medication adherence, blood pressure monitoring, dietary management, lifestyle modification, follow-up requirements and warning signs that warrant medical attention. Face-to-face education will be provided during hospitalisation and before discharge. Each education session will last approximately 10–20 min, with the frequency of contact determined by routine clinical workflow and individual patient needs. Before discharge, participants will receive standard instructions regarding medications, diet, follow-up arrangements and when to seek further medical attention. Printed or verbal educational materials routinely used in the department may also be provided where appropriate. After discharge, participants will attend scheduled outpatient follow-up in accordance with standard nephrology practice, during which clinicians will review disease status, laboratory results, medication use and self-management issues and provide further guidance as needed.

Throughout the intervention period, participants in the control group will not have access to the AI-HEALS platform or any of its AI-supported functions, including AI-based question-and-answer services, personalised content delivery, AI-generated feedback, platform-based self-management reminders, monitoring of platform engagement or other AI-supported behavioural feedback. The comparison will therefore assess the additional contribution of AI-HEALS when added to standard CKD care, while the care received by the control group will remain consistent with routine clinical practice.

Intervention group: standard care + AI-HEALS

Participants in the intervention group will receive standard care plus the AI-HEALS intervention for 3 months through the WeChat-based service platform, ‘Chronic Kidney Disease Health Management Expert’. AI-HEALS is designed to provide mobile health education and self-management support for people with CKD. The system integrates a LLM, an expert-reviewed CKD knowledge base, patient self-management records, personalised reminders, automated delivery of health education content and provider-side monitoring functions.

The intervention will be supported by a multidisciplinary team comprising public health specialists, nephrologists, clinical nurses, psychologists, biostatisticians and technical support staff. Nephrologists and clinical nurses will review the health education content, oversee the clinical appropriateness of AI-generated responses and update relevant clinical content when necessary. Psychologists will provide guidance on psychological support and behaviour change strategies. Public health specialists and biostatisticians will contribute to the design of the intervention and the evaluation of process and outcome measures. Technical staff will be responsible for platform maintenance, back-end data management and troubleshooting.

The AI-HEALS intervention comprises four main functional components: (1) an intelligent CKD knowledge question-and-answer system, (2) recording and monitoring of selected physiological indicators and lifestyle behaviours, (3) personalised reminder services and (4) automated delivery of tailored CKD self-management information.

Intelligent question-and-answer system for kidney disease knowledge

The AI-HEALS platform will provide an intelligent question-and-answer function supported by the Doubao Pro-32k LLM and a CKD-specific knowledge base. Participants will be able to submit CKD-related questions by text or voice through the WeChat platform. The system will generate patient-oriented responses to support health education, disease understanding and daily self-management. To encourage further engagement, the system will also suggest three related questions based on each user query, allowing participants to explore relevant topics in greater depth.

The question-and-answer function is intended solely for general health education and self-management support. It will not provide individualised diagnoses, medication dose adjustments, emergency treatment recommendations or replace clinical judgement. A clear disclaimer will be displayed on the interaction interface stating that AI-generated content is for informational purposes only and should not be used as a substitute for professional medical assessment, treatment, or emergency care.

Recording and monitoring of physiological indicators and lifestyle changes

Participants will be encouraged to regularly record key lifestyle and self-management information through the AI-HEALS system, including dietary behaviours and medication adherence. These data will be stored electronically within the platform and will allow the research team to monitor participants’ self-management behaviours over time. The collected information will be used to assess changes in lifestyle and self-management during the intervention and to support the provision of more targeted health education and self-management guidance.

Personalised reminder services

To improve treatment adherence, the AI-HEALS system provides a customised reminder function. Participants can set their own reminders, and the system will alert them at the scheduled times. Reminders can be set according to specific time schedules. This function is particularly important for individuals with chronic kidney disease, as adherence to medical advice is critical.

Automated and personalised CKD self-management information delivery

Participants will receive one to three automated educational messages per week through the AI-HEALS platform. The content will focus on key aspects of CKD self-management, including diet, physical activity and medication use, with the aim of supporting health literacy and day-to-day self-management. In addition, the system may recommend relevant educational articles based on participants’ interactions with the chatbot and their expressed information needs. The research team will periodically review platform usage data, such as the number and types of articles accessed, to assess engagement and refine the delivery of educational content when necessary.

The LLM used in this study is Doubao Pro-32k, a transformer-based model developed in China. The model has been adapted for medical knowledge applications through the incorporation of biomedical literature and CKD-related clinical guidelines. It also underwent general language pre-training to support language understanding and knowledge-based response generation. Model performance was evaluated using the third-party benchmarking platform FlagEval. As shown in figure 2, Doubao Pro-32k achieved an overall score of 77.75, compared with 73.51 for GPT-o1-preview. It also achieved a higher score in the knowledge application domain (91.14 vs 86.71). Before implementation in AI-HEALS, domain experts conducted a preliminary review of model-generated responses to assess their consistency with current medical knowledge and their suitability for use in patient health education.

Figure 2. Performance comparison between Doubao Pro-32k and o1-preview across multiple evaluation dimensions.

As AI-HEALS is delivered through a WeChat-based platform and incorporates third-party technical services, a predefined data governance framework will be applied to protect participant privacy and manage AI-generated and interaction data appropriately. The framework covers informed consent, data minimisation, secure data transmission, third-party data processing, data localisation, access control, de-identification, data retention, and management of AI interaction logs.

AI interaction and data use consent

During the informed consent process, participants will be informed that the AI-HEALS intervention involves interaction with an AI-based question-and-answer system, receipt of platform-delivered health education, submission of self-management records and monitoring of platform use by the research team. The consent form will describe the types of data collected through AI-HEALS, including account-related platform identifiers, AI interaction logs, text and voice inputs, system-generated responses, records of educational content accessed and reminders used, self-management records and other study-related data generated through use of the platform.

Participants will also be informed that AI interaction logs may be used for purposes related to the present study, including intervention monitoring, safety review, quality control and refinement of the CKD knowledge base and AI response rules within the scope of the approved protocol. Participation in AI-HEALS is voluntary, and participants may withdraw from the study at any time without affecting their usual medical care.

Data minimisation and de-identification

AI-HEALS will adhere to the principle of data minimisation. Only data necessary for intervention delivery, safety monitoring, outcome assessment and system quality control will be collected. Each participant will be assigned a unique study identifier. Direct personal identifiers, including names, national identification numbers, telephone numbers, home addresses and medical record numbers, will be stored separately from study data where applicable.

Before data are used for analysis or expert review, direct identifiers will be removed or replaced with the corresponding study identifier. The linkage file connecting study identifiers with participant identities will be stored separately on a password-protected hospital server, with access restricted to authorised members of the research team.

Data transmission based on WeChat

AI-HEALS will be accessed through a WeChat service account and mini-programme. Data submitted by participants through the WeChat interface, including text and voice inputs, self-management records and study-related platform interaction data, will be transmitted to the AI-HEALS back-end system for intervention delivery and research monitoring. Participants will be informed that WeChat serves only as the user-facing access channel for the intervention.

The research team will not collect or access personal WeChat social data unrelated to the study. Only study-related interaction data generated through participants’ use of the AI-HEALS platform will be collected and used for the purposes specified in the study protocol.

Third party technology providers and access to commercial systems

Technical support for AI-HEALS will be provided by third-party technology providers, including ByteDance and Mitochondrial (Beijing) Technology. Their roles and access to study-related data will be limited to the technical services necessary for system operation, including system maintenance, model access and troubleshooting.

To minimise privacy risks, direct personal identifiers will not be transmitted to the LLM. When processing user queries through the AI question-and-answer module, only the minimum information necessary to generate relevant health education responses will be provided to the model. Identifiable clinical information, telephone numbers, home addresses and medical record numbers will not be included in model prompts. Where contextual information is required to improve response relevance, only de-identified or coded information related to CKD self-management will be used.

Third-party technology providers will not use identifiable participant data for purposes unrelated to the study. AI interaction data and other study-related back-end data will not be used for commercial marketing or advertising. Any use of de-identified interaction data for system debugging or quality improvement will remain within the scope of the ethics-approved protocol and the purposes described in the informed consent form.

Data localisation, storage and access control

Identifiable research data will be managed by designated members of the research team and stored separately from the de-identified analysis dataset. Role-based access controls will be implemented so that researchers can access only the data required for their assigned responsibilities. Access to identifiable data will be restricted to authorised personnel involved in participant recruitment, follow-up and safety management.

All research team members and technical personnel with access to study-related data will be required to sign confidentiality agreements. Where technically feasible, data access activities will be logged to support monitoring and audit. De-identified datasets will be used for statistical analysis, expert review and study reporting.

Management of AI interaction log

AI interaction logs, including user queries, input modality, AI-generated responses, response status and review results, will be stored in the AI-HEALS back-end system. These logs will be used to monitor intervention delivery, identify potentially inaccurate or unsafe responses, assess system performance and support refinement of the CKD knowledge base and AI response rules.

During the intervention period, trained researchers will review the back-end interaction logs daily. Clinical experts will independently review a sample of AI-generated responses each week, covering at least 30% of responses generated during that period. Responses identified as incorrect, incomplete, ambiguous, hallucinated or potentially unsafe will be documented and reviewed, and appropriate corrective actions will be taken, including updates to the knowledge base, response templates or prompt rules where necessary. High-risk interactions will be escalated in accordance with predefined AI safety governance procedures.

Data retention and destruction

Research data will be retained only for the period required by applicable institutional policies, national regulations governing biomedical research and the conditions of the current ethical approval (Approval No. 2024–2106; approved on 14 January 2025). During this mandatory retention period, identifiable data and de-identified analysis datasets may be retained only for purposes related to the conduct, monitoring, audit, quality control and validation of the AI-HEALS study.

Identifiable information will be stored separately from the de-identified research dataset. The linkage file connecting study codes with participant identities will be stored separately and will be accessible only to authorised members of the research team. At the end of the required retention period, direct identifiers and the linkage file will be securely destroyed in accordance with institutional requirements.

The current ethical approval does not cover the long-term retention or reuse of de-identified datasets for independent secondary research. Any future secondary analysis, reuse of study data or retention beyond the scope of the current approval will require separate ethics review and, where applicable, additional written informed consent from participants before such activities are undertaken.

Privacy event management

In the event of unauthorised access, data leakage, serious system failure or other privacy-related incidents, the research team will promptly assess the scope and potential impact of the incident. The principal investigator, ethics committee and relevant institutional departments will be notified in accordance with institutional procedures. Appropriate corrective measures will be taken as needed, which may include temporary suspension of relevant system functions, technical remediation, participant notification where appropriate and documentation of the incident and subsequent actions.

Operation specification and safety management of AI-HEALS

To support consistent, auditable and safe delivery of the AI-HEALS intervention, predefined procedures will be followed for knowledge base development, content review and validation, automated message delivery, clinical oversight, escalation of potentially unsafe situations and management of inaccurate or unsafe AI-generated responses.

Knowledge base development

The AI-HEALS knowledge base will be developed specifically to support CKD self-management. It will cover basic disease information, aetiology, clinical manifestations, disease progression, warning signs requiring medical attention, medication use, blood pressure and blood glucose monitoring, dietary management, physical activity, lifestyle modification, prevention of complications, follow-up and daily self-management.

Sources used to develop the knowledge base will include clinical guidelines, nephrology textbooks, expert consensus documents and patient education materials routinely used in nephrology practice. Relevant information will be adapted into patient-oriented language to improve readability while preserving clinical accuracy. The knowledge base will be organised into structured thematic modules, including disease knowledge, medication management, diet and nutrition, physical activity, physiological monitoring, follow-up and examination guidance, psychological support and frequently asked questions.

Content development and clinician validation

Health education content will be developed by the research team and reviewed by nephrologists and clinical nurses before being released on the platform. The review will consider clinical accuracy, consistency with current CKD management recommendations, readability, suitability for patient education and the risk of potentially misleading or unsafe information. Content will be released only after review and approval by at least two clinical team members. Any disagreements will be discussed with the principal investigator or a designated nephrologist until consensus is reached.

AI-generated responses will be limited to health education and self-management support and will not provide medical diagnoses or individualised treatment decisions. Where relevant information is available, the system will be configured to prioritise content from the expert-reviewed CKD knowledge base. Before deployment, the research team will test representative patient queries to assess the accuracy, clarity and clinical appropriateness of the generated responses.

During the intervention period, trained researchers will review the back-end interaction logs daily. The review will focus on unresolved queries, inaccurate or incomplete responses, irrelevant content, potentially unsafe recommendations and interactions requiring clinical review. In addition, clinical experts will independently review a sample of AI-generated responses each week, covering at least 30% of responses generated during that period. Responses identified as inaccurate, incomplete, ambiguous or potentially unsafe will be documented and reviewed, and appropriate corrective actions will be taken. These may include revision of the knowledge base, response templates or prompt rules, as appropriate.

Escalation of clinical or safety issues

AI-HEALS will be used solely as a health education and self-management support tool and will not replace professional medical diagnosis, treatment or emergency care. During the informed consent and training process, participants will be advised that AI-generated responses are for informational purposes only and that individual clinical decisions should be made in consultation with qualified healthcare professionals.

Prevent and manage incorrect or unsafe AI responses

Several safeguards will be implemented to reduce the risk of incorrect or unsafe AI-generated responses. First, the system interface will display a clear reminder that AI-generated content is for informational purposes only. Second, the system will be instructed not to provide individualised prescriptions, medication dose adjustments, diagnostic conclusions or emergency treatment recommendations. Third, when a query falls outside the scope of the knowledge base or requires individual clinical decision-making, the system will advise the participant to consult a healthcare professional rather than provide a definitive answer.

All interaction logs will be stored in the back-end system and reviewed daily by trained researchers. Potentially inaccurate, incomplete, irrelevant or unsafe responses will be flagged and documented. In addition, clinical experts will independently review at least 30% of AI-generated responses each week to assess their clinical appropriateness and safety. Identified problems will be reviewed and, where necessary, addressed through revisions to the knowledge base, response templates, or prompt rules.

If recurrent errors of the same type, serious technical faults, unstable model access or major security issues are identified, the affected AI function will be temporarily suspended. The technical team and clinical experts will jointly investigate the cause, implement appropriate corrections and retest the function before it is restored. Any adverse events potentially related to the intervention will be documented and reported in accordance with the study safety monitoring procedures.

Back end monitoring and intervention optimisation

The AI-HEALS back-end system will record study-related platform use, including question categories, chatbot use, text or voice input, access to educational articles, reminder use and completion of self-management records. These data will be reviewed to monitor engagement and identify potential adherence issues. Participants with low engagement or incomplete self-management records may receive additional reminders or support through WeChat, telephone contact, or routine clinical follow-up.

The knowledge base and intervention content may be refined during the study in response to participant queries, daily back-end reviews, weekly clinical expert reviews, identified AI response problems and emerging self-management needs. All changes to the knowledge base, response templates, or system rules will be documented to maintain transparency and enable the intervention procedures to be traced over time.

AI security governance

Since AI-HEALS includes an interactive LLM, a predefined security governance framework will be implemented to reduce the risks of hallucinations, inappropriate medical advice, unsafe responses and delayed management of high-risk clinical information. The framework includes a patient-oriented disclaimer, real-time red-flag response rules, restrictions on personalised clinical recommendations, near real-time manual supervision, a clinician escalation pathway, and continuous error correction.

Disclaimer for patients

AI-HEALS will be used only as a health education and self-management support tool. It will not replace professional medical diagnosis, treatment, emergency medical care or personalised clinical decision-making. During the informed consent process and platform training, participants will be informed that AI-generated content is for reference only. In addition, a visible reminder will be displayed on the system’s interactive page, stating that all AI-generated content is for reference only. Participants should consult qualified healthcare professionals for diagnosis, treatment decisions, drug adjustments, or emergency symptoms.

Daily review

During the intervention, trained researchers will review back-end interaction records daily to identify user input or AI-generated responses that may indicate an urgent clinical risk or require clinician involvement. Predefined red-flag categories will include emergency symptoms, rapidly deteriorating conditions, markedly abnormal physiological indicators, drug-related adverse events, requests for personalised dose adjustments, suicidal ideation or serious psychological distress and other potentially unsafe clinical situations. Examples of red-flag information include, but are not limited to, chest pain, severe shortness of breath, loss of consciousness, seizures, sudden weakness or speech difficulties, severe headache, severe oedema, significant reduction in urine output, persistent vomiting or diarrhoea, suspected severe infection, extreme hypertension, symptoms suggestive of electrolyte disorders, severe drug-related adverse reactions and expressions of self-harm or suicidal thoughts. If a potential red-flag interaction is found during the daily back-end review, the researcher will immediately report it to the designated clinical team members. The clinical team will assess the risk level and contact the participant via WeChat, telephone or routine clinical follow-up as appropriate. If necessary, the participant will be advised to seek outpatient consultation, specialist evaluation or emergency medical care. Participants are advised to seek immediate emergency care for life-threatening symptoms or risk of self-harm. All red-flag interactions, reporting procedures, clinical evaluations, actions taken and outcomes will be recorded in the study records. These records will also be used to improve the knowledge base, refine response rules, and optimise safety management procedures during the intervention.

Management of drug-related queries

Drug-related issues will be treated conservatively. AI-HEALS can provide general educational information about medication adherence, common precautions and the importance of following prescription treatment plans. However, the system will not provide personalised prescriptions, advice on initiating or discontinuing medications, dose adjustments, drug substitutions or management of complex adverse reactions without clinician involvement.

If participants ask whether to start, stop, change or adjust a medication dose, or report suspected drug-related adverse reactions, the system will recommend that they consult their treating clinician. Such interactions will be flagged during the back-end review. If the query indicates potential hazards, such as serious adverse reactions, drug overdose or unsafe medication discontinuation, the case will be reported to the designated clinical team members for further evaluation.

Limitations of individualised clinical recommendations

AI-HEALS will be instructed to avoid individualised clinical decisions. The system will not provide a definitive diagnosis, personalised treatment plan, complex interpretation of laboratory results, dose adjustments or emergency management decisions. When participants’ questions are beyond the scope of health education or require personalised clinical judgement, the system will provide a boundary-setting response, explain that the issue requires professional assessment and recommend that participants consult a clinician.

AI-HEALS will only provide general educational explanations for issues involving laboratory indicators, changes in renal function, complications or treatment options at an appropriate time, and will recommend clinical consultation for personalised advice. This approach is intended to reduce the risk of over-reliance on AI-generated responses.

Prevention and management of hallucinations and inappropriate reactions

A number of safeguards will be implemented to reduce hallucinations and inappropriate AI-generated medical advice. First, when relevant content is available, the AI system will be instructed to prioritise the expert-reviewed CKD knowledge base. Second, when there is insufficient information in the knowledge base to answer a question, the system will be instructed to acknowledge the limitation rather than generating unsupported content. Third, the system will avoid speculative statements, unsubstantiated claims and personalised treatment recommendations.

During the intervention, trained researchers will review the back-end interaction log daily. The daily review will focus on identifying hallucinations, inaccurate or incomplete responses, irrelevant responses, unsafe suggestions, unresolved issues and red-flag interactions that require clinician involvement. In addition, clinical experts will conduct independent third-party reviews of AI-generated responses every week, with at least 30% of responses being reviewed. Any hallucinatory, incorrect, incomplete, ambiguous or potentially unsafe responses will be recorded and corrected. The corresponding knowledge base content, response templates, or prompt rules will be updated accordingly.

Researcher supervision

A hierarchical manual supervision approach will be established. Daily problems and low-risk self-management issues will be handled through the AI-HEALS education and feedback process. Issues requiring clarification, involving potentially inaccurate AI responses, or indicating poor self-management adherence will be reviewed by trained researchers during daily back-end monitoring. Problems involving possible clinical risks, drug-related harm, emergency symptoms, severe psychological distress or personalised treatment decisions will be reported to clinical experts.

When high-risk content is identified, the designated clinical team members will contact participants via WeChat, telephone or routine clinical follow-up depending on the urgency of the situation. If necessary, participants will be advised to seek outpatient consultation, expert evaluation or emergency medical care. All escalated cases, actions taken, and outcomes will be recorded in the study records.

Temporary suspension and system correction

If repeated hallucinations, repeated unsafe responses, serious programming errors, unstable model access or major security problems are identified, the affected AI functions will be temporarily suspended. The technical team and clinical experts will jointly review the cause, modify the knowledge base, response rules, prompt settings or system interface, and retest before reactivation. Serious adverse events or safety issues related to the intervention will be reported according to the study safety monitoring and ethical reporting procedures.

Patient and public involvement

Patients and the public were not involved in the development of the research question, the selection of outcome measures, the design of this study or the recruitment and conduct of the trial. The decision not to involve patients was based on the nature of this study: a protocol for a randomised controlled trial evaluating a newly developed, standardised AI-HEALS. The intervention and study procedures are technology-driven and standardised, with no requirement for direct patient input into the study design or modification to the intervention delivery. The burden of the AI-HEALS intervention was not assessed by patients prior to this trial, as the intervention is delivered via an automated system that imposes minimal participant burden, involving only the receipt of health education messages through the AI-HEALS platform.

The results of this trial will be disseminated to study participants and the wider public through publication in a peer-reviewed journal, presentation at scientific conferences, and provision of a plain language summary to participants on request.

Strategies to enhance adherence to the intervention

To encourage participants to comply with regulations, the following strategies will be implemented: (1) During the recruitment process and when signing the informed consent form, a comprehensive explanation of the participants’ responsibilities during the study period will be provided. At the same time, efforts will be made to establish strong relationships with participants to promote mutual trust; (2) Throughout the research, researchers will ensure that participants are fully educated on disease management and prevention strategies. This will include an in-depth discussion about disease characteristics, with emphasis on the need to adhere to the recommended treatment plan, as well as strategies to prevent complications that may develop in the course of treatment; (3) The research team will also assess participants’ interaction with the system on a weekly basis to ensure adherence. This will help to identify problems and ensure that corrective measures are put in place to maintain participants’ continued participation.

Outcomes

Primary outcome

The primary outcome is quality of life, assessed by the SF-12 PCS score. The primary endpoint is defined as the between-group difference in change from baseline to the end of the 3-month intervention. Scores at 6 months post-intervention and at other time points will be analysed as secondary endpoints.

Secondary outcomes

Clinical outcomes related to renal function and CKD will be evaluated using standardised CKD indicators as secondary outcomes. These outcomes will include serum creatinine levels and urinary protein creatinine ratio. Serum creatinine levels will be measured by conventional laboratory methods and reported in μ mol/l. Urinary protein creatinine ratio, which will be expressed as mg/g or mg/mmol according to the reporting standards of the hospital laboratory.

In addition to the above physiological indicators, depressive and anxiety symptoms will be assessed using the Patient Health Questionnaire-4 (PHQ-4), quality of sleep using the Short Form of the Pittsburgh Sleep Quality Index (PSQI-SF), self-efficacy using the Short Form of the New General Self-Efficacy Scale (NGSES-SF), eHealth literacy using the Chinese version of the eHealth Literacy Scale (eHEALS) and dietary behaviours using the Chinese version of the Self-Regulation of Eating Behaviour Questionnaire (SREBQ).

In addition, the number of participants recruited, the number who withdrew and the reasons for withdrawing will be documented. During the analysis, data homogeneity will be assessed by comparing demographic characteristics and outcome measures between participants who completed the study and those who withdrew, in order to determine whether withdrawals may have biased the results.

Variable measurement

Sociodemographic variables will be collected at baseline. Physical examinations and questionnaire assessments are planned at baseline, at discharge, at the end of the intervention and at 3, 6 and 9 months after completion of the intervention (table 1).

Table 1. The schedule of enrolment, interventions and assessments.

Study period	Recruitment	Intervention		Follow-up
Time point	0	Discharge	3 months*	3 months	6 months	9 months
Eligibility screen	√
Informed consent	√
Allocation	√
Sociodemographic variables	√
SF-12	√	√	√	√	√	√
PHQ4	√	√	√	√	√	√
NGSES-SF	√	√	√	√	√	√
PSQI-SF	√		√	√	√	√
SREBQ	√		√	√	√	√
eHEALS	√

^*Main observation time nodes.

eHEALS, eHealth Literacy Scale; NGSES-SF, New General Self-Efficacy Scale-Short Form; PHQ-4, Patient Health Questionnaire-4; PSQI-SF, Pittsburgh Sleep Quality Index-Short Form; SF-12, Short Form Health Survey; SREBQ, Self-Regulation of Eating Behaviour Questionnaire.

Sociodemographic variables

A customised demographic questionnaire will be used to collect information on the following variables: sex, age, marital status, household registration type, average household size, educational attainment, occupation, income status, as well as clinical data related to CKD.

Anthropometric variables

The height, weight and waist circumference of each participant will be measured twice using certified instruments to ensure accuracy. Height and weight will be measured using fully calibrated and validated electronic devices. Body mass index will be calculated by dividing body weight (kg) by the square of height (m²). The waist circumference will be measured horizontally at the navel using a flexible tape measure.

To maintain continued participation after discharge, the research team will communicate with participants via WeChat, phone or text messages. Through these interactions, participants will receive timely updates on study progress, guidance for resolving any new issues and reminders about upcoming follow-up evaluations.

Standardised data entry for this study will be conducted using an electronic data capture system. Initially, double data entry will be performed independently by two operators to ensure accuracy, followed by logical validation checks of all raw data. Sensitive information will be encrypted using a Unicode-based encoding scheme. Data will be stored on password-protected servers with role-based access control, such that only authorised researchers can access de-identified data. In addition, real-time backups will be sent to off-site cloud storage to ensure data integrity and disaster recovery capability.

After collecting participant data, all personally identifiable information will be immediately deleted to protect confidentiality. This process involves replacing personal identifiers with unique codes. Only a limited number of authorised research team members can access the keys that link these codes to actual identities. The key will be stored in a separate, highly secure location. Under no circumstances will any personal information be shared with third parties without the participants’ explicit informed consent. All research team members must sign a confidentiality agreement to ensure compliance with the highest privacy protection standards. Throughout the study and thereafter, all personal information will be securely stored on dedicated servers and accessed only by authorised researchers.

Instruments

Short Form Health Survey

The quality of life will be evaluated using the SF-12 questionnaire, which is a simplified version of the SF-36 health survey.³⁰ SF-12 consists of 12 projects that evaluate the quality of life in two summary areas, physical health (PCS) and mental health (psychological composition summary, MCS), as well as eight dimensions: general health, physical function, role limitations caused by physical health, physical pain, role limitations caused by emotional problems, mental health, vitality and social function. The original PCS and MCS scores will be converted into standardised scores ranging from 0 to 100 using established scoring algorithms, with higher scores indicating better health status. SF-12 exhibits good internal consistency, with Cronbach’s alpha of 0.775.

Patient Health Questionnaire-4

Depression and anxiety will be evaluated using the PHQ-4, a brief self-report screening tool designed to assess symptoms of depression and anxiety over the past 2 weeks.³¹ The PHQ-4 comprises four items, with the first two assessing depressive symptoms and the latter two assessing anxiety symptoms. Each item is rated on a four-point Likert scale ranging from ‘not at all’ (0) to ‘nearly every day’ (3), yielding a total score between 0 and 12. The PHQ-4 has been shown to possess good validity and reliability, and studies conducted in China have reported a Cronbach’s alpha coefficient of 0.833.

New General Self-Efficacy Scale-Short Form

The NGSES-SF is used to measure individuals’ confidence and perceived ability to successfully cope with specific tasks or challenges.³² This scale also comprises three items representing three aspects of self-efficacy: perceived competence (Item 1), strength of confidence (Item 2) and generalisation of confidence (Item 3). Items are rated on a five-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’), with total scores ranging from 3 to 15. Higher scores indicate stronger self-confidence and coping ability. The NGSES-SF has demonstrated good internal consistency, with a reported Cronbach’s α of 0.876.

Pittsburgh Sleep Quality Index-Short Form

The PSQI-SF is a short form of the Pittsburgh Sleep Quality Index that measures the quality of sleep for an individual.³³ The scale comprises six questions that cover five major aspects: sleep efficiency (Items 1 and 2), sleep latency (Item 3), sleep duration (Item 4), sleep disturbances (Item 5) and sleep quality (Item 6). The bedtime and wake-up time data are also used to calculate the sleep efficiency. The scores range from 0 to 15. Higher scores on the PSQI-SF indicate poor quality of sleep. The PSQI-SF shows good internal consistency reliability with a Cronbach’s α of 0.79.

Chinese version of the Self-Regulation of Eating Behaviour Questionnaire

The SREBQ is used to evaluate an individual’s ability to regulate dietary behaviour.³⁴ This scale consists of five items, rated according to the Likert five point scale, ranging from 1 (‘never’) to 5 (‘always’). The total score ranges from 5 to 25 points, with higher scores indicating stronger ability to regulate dietary habits. The scale showed acceptable internal consistency, with Cronbach’s alpha of 0.75.

Chinese version of the eHealth Literacy Scale

The eHEALS is designed to comprehensively assess individuals’ knowledge, perceived ease of use and skills in searching for, evaluating and applying electronic health information to address real-world health problems.³⁵ This scale includes five items rated on a five-point Likert scale ranging from 1 (‘strongly disagree’) to 5 (‘strongly agree’). Total scores range from 5 to 25, with higher scores indicating higher levels of eHealth literacy. The scale demonstrates excellent internal consistency, with a reported Cronbach’s α of 0.913.

Statistical analysis

This study will follow the intention-to-treat principle and include all randomised participants. Per-protocol analysis will be used as a sensitivity analysis, defined as participants without significant protocol deviations and with sufficient intervention exposure.

Descriptive statistics will be presented according to data distribution characteristics. Continuous variables will be described as mean±SD deviation or median (IQR). Categorical variables will be described as frequencies (percentages).

The primary outcome is quality of life, assessed by the SF-12 PCS score. The primary endpoint is the between-group difference in change from baseline to the end of the 3-month intervention. A linear mixed-effects model (LMM) will be used to analyse repeated measurements. The fixed effects will include group, time (as a categorical variable) and the group-by-time interaction. A random intercept for each participant will be included to account for intraindividual correlation. The primary efficacy estimate is the adjusted between-group difference in SF-12 PCS change from baseline to 3 months, derived from the group-by-time interaction term. In the event that the prespecified primary LMM fails to yield stable estimates, alternative approaches will be considered, including generalised estimating equations, generalised linear mixed models, repeated-measures analysis of variance and analysis of covariance. The final analytical method for the longitudinal repeated-measures data on the primary outcome will be selected based on the data distribution, completeness and model fit characteristics.

The model will force adjustment for the following prespecified baseline covariates: age, sex, baseline SF-12 PCS score and CKD stage. CKD stage will be included as an ordered categorical variable. If the model does not fit well due to sparse cells, CKD stages will be prespecified to be combined into two levels: ‘stage 1–3’ and ‘stage 4–5’. To ensure model robustness and statistical validity, the final specification and categorization of covariates in this study will be optimised based on the empirical data distribution and model fit performance. All covariate adjustment strategies and data-driven optimisation approaches will be fully documented in the final statistical analysis plan and the resulting manuscript.

Continuous secondary outcomes, including serum creatinine levels, urinary protein creatinine ratio, anxiety, depression, quality of sleep, self-efficiency, dietary behaviour and eHealth literacy, will be analysed using appropriate longitudinal analytical methods consistent with the primary outcome analytical framework, chosen according to their data characteristics. Feasibility, intervention acceptability, system usage and safety-related outcomes will be summarised descriptively with appropriate summary statistics.

All secondary outcome analyses are predefined exploratory analyses, aiming to provide preliminary evidence for the potential effects of the intervention. The multiple-comparison adjustment strategy will be determined comprehensively based on the number of secondary endpoints, the observed data and mainstream statistical standards for clinical exploratory studies. All secondary outcomes will be reported with 95% CIs and corresponding p values.

Missing data will be examined by checking the missing pattern and reasons for missingness by group and time point. Under the missing at random assumption, the primary model will use all available repeated measurements. If the missing rate for the primary outcome exceeds 15%, multiple imputation by chained equations will be used as a sensitivity analysis. The imputation model will include treatment group, time point, baseline covariates, CKD stage, observed outcome values and variables related to missingness. Results from the combined imputed datasets will be summarised using Rubin’s rules.

The following sensitivity analyses will be performed to verify the robustness of the results:

1. Complete case analysis.

2. Per-protocol analysis.

3. The main analysis repeated after excluding participants with CKD stage 1.

4. The main analysis repeated after excluding participants with CKD stage 5.

All statistical tests will be two-sided, with a significance level of 0.05. The effect estimates will be reported with 95% CIs. Data analysis will be conducted using IBM SPSS Statistics (V.24.0; SPSS Inc., Chicago, Illinois, USA) and Stata (V.14.0); StataCorp, University City, Texas, USA.

Qualitative data collection and analysis

Data collection

To understand the fatigue, frustration and subjective burden experienced by participants while using the AI-HEALS application, the researchers will select a subset of participants from the intervention group to conduct semi-structured one-on-one interviews. The interviews will be conducted within 2 weeks after the end of the 3-month intervention period. Each interview is expected to last approximately 30 min and may be conducted by telephone or face-to-face. When selecting interviewees, sociodemographic characteristics such as gender, age, residence, education level and CKD severity will be comprehensively considered to ensure sample diversity. The interviews will be conducted by a trained qualitative researcher holding a Master’s degree, with prior experience in semi-structured interviews in healthcare settings. This interviewer was not involved in the development of the AI-HEALS application to minimise potential bias. Before data collection, all interviewers will receive standardised training on the interview guide and procedures. The research team will develop an interview guide based on the Technology Acceptance Model and previous digital health feasibility studies, covering the following topics: (1) frequency and degree of fatigue when using the application; (2) situations that cause frustration; (3) subjectively perceived time, energy and cognitive burden; (4) suggestions for improving negative user experiences. An interview will be conducted with two participants prior to formal data collection to refine question wording and interview flow, with these data excluded from the final analysis. Participant recruitment will continue until data saturation is achieved. All interviews will be audio-recorded, transcribed verbatim and anonymised. The full semi-structured interview guide is provided as online supplemental interview guide.

Data analysis

Qualitative data will be analysed using the Theoretical Domains Framework as the analytical framework, with NVivo V.12 (QSR International, Doncaster, Australia) used to support data management and coding. The qualitative findings will be integrated with the quantitative results to provide a broader understanding of participants’ experiences with the intervention and to help interpret the feasibility, implementation and preliminary effects of AI-HEALS. Two researchers with experience in qualitative data analysis will independently code a subset of the transcripts and compare their coding to develop and refine the coding framework. The agreed framework will then be applied to the remaining transcripts. Coding discrepancies will be discussed between the two researchers, and a third senior researcher will be consulted if consensus cannot be reached. Coding decisions and their rationale will be documented to maintain an audit trail.

To enhance trustworthiness, this study incorporates several strategies. Data saturation will be pursued during recruitment. Independent double coding will reduce coder bias. A complete audit trail, including original audio files, verbatim transcripts, coding notes and records of disagreement resolution, will be maintained to ensure transparency and traceability of analytic decisions. Given participant burden, member checking will not be performed; instead, priority will be given to transparent reporting of the analytic process.

Reflexivity

We acknowledge that the researchers’ personal backgrounds, disciplinary training and prior experiences in digital health research may influence data collection and interpretation. To mitigate this, the primary interviewer will maintain a reflexive journal throughout the study to document her/his positionality, assumptions and interactions with participants. The research team will also hold regular debriefing meetings to discuss emerging interpretations and identify potential biases.

Oversight and monitoring

Composition of the coordinating centre and trial steering structure

This study is a single-centre trial designed, implemented and coordinated by West China Hospital of Sichuan University. Day-to-day trial support will be provided by the following personnel:

• Principal investigator: responsible for overall trial oversight and clinical care of participants.

• Data manager: responsible for organising data collection and ensuring data quality.

• Research coordinator: responsible for trial registration, coordination of follow-up visits and submission of annual safety reports.

• Study physicians: responsible for identifying potential participants, obtaining informed consent and ensuring protocol-adherent follow-up.

The research team will convene meetings every 2 weeks. No formal trial steering committee or patient and public involvement group will be established for this study.

Composition, responsibilities and reporting structure of the data monitoring committee

The Data Monitoring Committee (DMC) will comprise at least two members from the Ethics Committee of West China Hospital of Sichuan University, all of whom have no conflicts of interest with the study. The primary responsibility of the DMC is to regularly monitor and evaluate study progress and report monitoring outcomes to the ethics committee, thereby ensuring independence from the research team and maintaining appropriate distance from the principal investigators. In the event of serious adverse events (SAEs) or adverse events (AEs), these must be reported to the committee within 24 hours. The DMC will assess the relationship between SAEs and the intervention (definitely related or possibly related) and, if a relationship is identified, will recommend further safety measures.

Adverse event reporting and harms

If any deviations from the approved study protocol or unauthorised modifications are identified during monitoring, the DMC has the authority to suspend or terminate the study. Throughout the trial, the research team must promptly report all SAEs and AEs to the ethics committee. Should any participant experience an SAE or AE, regardless of its direct relevance to the study, investigators are required to report the event to the research team within 24 hours and suspend the trial as necessary. In addition, if serious programming errors occur within the AI system or major issues are identified during user interactions, the trial must be immediately halted. The research team will then address technical issues and improve the AI system’s user interface. Following resolution and re-evaluation of the identified problems, the study may be modified and restarted to ensure participant safety and the scientific validity of the trial.

Trial auditing frequency and procedures

In accordance with the Ethics Committee of West China Hospital of Sichuan University, an independent research monitor will be appointed to conduct study-specific audits. The independent monitor will perform two on-site visits per year to verify the presence and completeness of study documentation. In addition, a random sample of 25% of participants will be selected for source data verification, including review of informed consent forms, assessment of compliance with inclusion and exclusion criteria, verification of source data and evaluation of the reporting and documentation of AEs, SAEs and suspected unexpected serious adverse reactions. Further details are provided in the monitoring plan.

Communication of important protocol amendments to relevant parties

The research team will first discuss all proposed amendments to the study protocol. After reaching an agreement, the proposed changes will be submitted to the hospital ethics committee and research management department for approval. Once approved, a formal document detailing the revisions will be provided to all relevant parties (such as trial participants and ethics committees). Electronic versions will also be provided if necessary. Any modifications to the protocol will be recorded and signed by the principal investigator. These records will be kept together with other study documents. After approval by the research management department and ethics committee, the principal investigator will update the clinical trial application accordingly.

Ethics and dissemination

This study protocol was reviewed and approved by the Biomedical Ethics Review Committee of West China Hospital, Sichuan University, approval number 2024–2106 (approval date: 14 January 2025). Please refer to the online supplemental ethics approval. Written informed consent will be obtained from all participants prior to enrolment in the study.

The trial has been registered in the Chinese Clinical Trial Registry (ChiCTR) under the identifier ChiCTR2500099413 (registration date: 24 March 2025).

The research team will prepare a summary of the RCT results and distribute it via email to participants who have indicated, in their informed consent, an interest in receiving study findings. Study results will be disseminated through publications in national and international peer-reviewed journals and presentations at national and international conferences. In addition, the findings will be shared with other hospitals and made available to healthcare professionals on request. Study results will also be provided on request to interested parties.

Discussion

This study protocol describes a randomised controlled trial evaluating the effectiveness of the AI-HEALS intervention among adults with CKD. The study aims to determine whether AI-HEALS, when added to standard care, improves health-related quality of life and supports self-management. Health-related quality of life assessed using the PCS score of the SF-12 will serve as the primary outcome, while secondary outcomes will include clinical, psychological, behavioural and eHealth literacy measures. This study will provide evidence regarding the potential role of AI-based health education in CKD management. The AI-HEALS intervention comprises four main components: a LLM-based question-and-answer system supported by an expert-reviewed CKD knowledge base, a module for recording and monitoring selected physiological indicators, personalised reminder services and tailored health education and self-management information. Together, these components are designed to provide ongoing support for CKD self-management and health behaviour change.

Systematic reviews published in recent years indicate that, although numerous RCTs have evaluated mobile health interventions for people with CKD, most have relied on conventional approaches such as text messaging, telephone support or standard mobile communication platforms.^36–38 In contrast, interventions integrating LLMs with mobile social media platforms remain relatively uncommon. By combining an LLM with a widely used mobile platform, this study explores a new approach to supporting CKD self-management and health education. In addition to health-related quality of life, the study will examine a range of physiological, psychological and behavioural outcomes to provide a broader understanding of participants’ responses to the intervention. The findings may help inform the further development and evaluation of AI-supported mobile health interventions for CKD and other chronic conditions.

This study has several limitations. First, because of the limited recruitment period and sample size, randomisation will not be stratified by CKD stage. Although CKD stage will be considered in prespecified covariate-adjusted analyses, subgroup analyses and sensitivity analyses, imbalances in disease stage may still occur because of the relatively small sample size and could affect the estimated intervention effects. Future studies with larger samples could use stratified randomisation by CKD stage to improve balance between groups and provide more robust estimates.

Second, although participants will be admitted to separate wards to reduce contamination during hospitalisation, some risk of contamination may remain after discharge because the intervention is delivered through WeChat. Participants in the control group may communicate with those in the intervention group or obtain similar health information from other online sources, which could reduce the observed differences between groups.

Finally, because of the nature of the intervention, participants and staff delivering the intervention cannot be blinded to group allocation. In addition, several outcomes are based on participant self-report, which may introduce performance and reporting biases. These limitations should be considered when interpreting the findings.