This study was designed as an explanatory qualitative study nested within the intervention arm of a cluster randomized controlled trial. It was intended to clarify perceived mechanisms of behavior change, contextual influences on use, and implementation-related experiences that were not captured by trial outcomes alone. The parent trial, registered at the Chinese Clinical Trial Registry (ChiCTR2300068952; registered on February 3, 2023), aimed to evaluate the effectiveness of AI-HEALS on glycemic control and self-management behaviors. In contrast, this qualitative substudy was designed to address questions that quantitative outcomes alone could not answer: how participants perceived the intervention to influence daily self-management, which contextual factors facilitated or constrained its use, and which implementation-related features shaped engagement in community-based primary care. Thus, the qualitative component was not intended to retest efficacy but to explain perceived mechanisms, user experiences, and real-world delivery processes within the intervention arm. The trial protocol has been published elsewhere [24].

The parent trial was conducted from August to December 2023 across 45 community health centers in the Daxing and Shunyi districts of Beijing. A total of 406 participants were enrolled using cluster sampling, with each community health center recruiting 8 to 10 patients.

Clusters were randomized 1:1 using a computer-generated sequence, assigning 208 participants to the intervention group (standard diabetes primary care plus AI-HEALS) and 198 participants to the control group (standard diabetes primary care alone). Complete blinding of participants and researchers was not feasible due to the nature of the intervention.

For this qualitative substudy, a purposive maximum variation sampling strategy was used to select information-rich cases from the intervention arm. Sampling criteria were designed to capture a broad range of experiences and perspectives, including sex (male/female), age (<60 y/≥60 y), diabetes duration (<5 y/5–10 y/>10 y), HbA_1c levels (<7%/≥7%), self-reported digital literacy (high/moderate/low), and level of platform use (frequent/moderate/infrequent based on login frequency and feature usage).

From the 208 participants in the intervention arm, we purposively approached 25 individuals who represented diverse combinations of the sampling criteria. Recruitment was conducted via telephone by a trained research assistant who explained the purpose of the qualitative interview and invited participation. Of the 25 individuals approached, 17 agreed to participate, yielding a response rate of 68%. Reasons for nonparticipation among the 8 decliners included lack of interest (n=3, 38%), time constraints due to work or family obligations (n=3, 38%), and concerns about privacy (n=2, 25%).

The sample size was determined by data saturation, defined as the point at which no new themes emerged from subsequent interviews. Saturation was achieved after 17 interviews, with no new themes emerging from the final 2 interviews. The final sample consisted of 17 AI-HEALS users from the intervention arm.

The intervention was delivered through a 2-layer architecture: the backend AI-HEALS and the frontend “Beiyi Sugar Manager” WeChat Mini Program (Tencent Holdings Ltd.). AI-HEALS used a hybrid logic architecture rather than a fully generative model. Specifically, the conversational function relied on vendor-provided NLP and natural language understanding modules to identify user intent and extract key terms, after which the system retrieved preconfigured responses from a curated diabetes education knowledge base. In parallel, reminder delivery and content recommendation were governed primarily by deterministic rule-based logic using baseline patient profile variables (eg, age, duration of diabetes, HbA_1c level, and self-reported digital literacy), recent article reading history, manual health logs, and regimen-related schedules. This architecture allowed the system to combine conversational access with stable, auditable delivery of education and reminders.

The “Beiyi Sugar Manager” Mini Program was the patient-facing application embedded within the WeChat ecosystem, through which participants accessed all intervention components without installing a standalone app. All communications—including educational content, AI chatbot interactions, automated reminders, and message notifications—were delivered via the WeChat Mini Program interface.

Educational messages were prioritized according to broad self-management domains (diet, exercise, medication, monitoring, complications) and adjusted using recent engagement patterns, such as categories previously read, reading duration, and frequency of health data entry. Reminder schedules were anchored to user-entered medication and monitoring routines and could be intensified when recent logging was irregular. The educational knowledge base and decision rules were developed by the research team with clinical and public health input before trial deployment. Because the dialog platform was commercially provided, the underlying training corpus, model weights, and proprietary internal validation metrics were not available to the research team for publication.

The mini program provides 4 main functional modules powered by the AI-HEALS backend:

1. AI-powered conversational agent (chatbot).

2. Personalized educational content library.

3. Physiological monitoring and reminder system.

4. Theory-driven automated messaging.

Semistructured interviews were conducted with all 17 participants 3 months after the intervention to explore their experiences and perceived influence of AI-HEALS on self-management behaviors, including diet, physical activity, and continuous glucose monitoring. Interviews, lasting 30 to 60 minutes, were held either in person or by telephone. Participants were encouraged to recount context-specific scenarios and critical incidents, describe their interactions with the system and challenges encountered, and propose suggestions for improvement related to behavior change. All interviews were audio recorded with permission and transcribed verbatim.

Thematic analysis of the transcripts was performed using NVivo software (version 12; QSR International). To enhance transparency and reproducibility, an audit trail of coding decisions was maintained. Two researchers (JW and YJ) independently coded the transcripts. After initial coding of the first 3 transcripts, the research team discussed the preliminary codebook and resolved any discrepancies through consensus; subsequent consensus meetings were held to refine codes and develop themes. The coding process was primarily inductive, allowing themes to emerge from the data. Following the completion of initial thematic analysis, member checking was conducted with 5 participants, who were asked to review a summary of the findings and provide feedback on whether the interpretations resonated with their experiences. This feedback confirmed the credibility of the themes and did not necessitate substantial revisions.

This study was approved by the Biomedical Ethics Committee of Peking University (IRB00001052-22058; approved on June 6, 2022) and registered with the Chinese Clinical Trial Registry (ChiCTR2300068952; registered on February 3, 2023). Written informed consent was obtained from all interview participants before data collection. Participants did not receive financial compensation for participation. Several measures were implemented to protect participant anonymity and data security within the WeChat environment. All data transmitted through the “Beiyi Sugar Manager” mini program were encrypted using industry-standard protocols. Participant identifiers were replaced with unique study codes, and data were stored on secure servers accessible only to authorized research personnel. Participants were informed that while WeChat’s platform has its own privacy policy, the research team implemented additional safeguards, including not collecting WeChat IDs or contact lists and advising participants not to share personal health information through unsecured channels. The mini program was designed to comply with China’s Personal Information Protection Law and relevant data security regulations.

Demographic characteristics of the patients are shown in Table 1.

The study included 17 participants with T2DM, with a mean BMI of 26.33 (SD 3.60)  kg/m² and a mean waist circumference of 97.88 (SD 13.45) cm. The majority were male (n=10, 58.82%) and of Han ethnicity (n=16, 94.12%), with approximately half reporting a monthly income of 3001 to 9000 CNY (based on an exchange rate of US $1=CN ¥6.906). Participants varied in occupation and education, with 41.18% (n=7) employed, 29.41% (n=5) retired, and educational attainment ranging from junior high school to a bachelor’s degree. Most participants were covered by Beijing medical insurance (n=15, 88.24%) and lived with family members, while 47.06% (n=8) reported a family history of diabetes. The duration since diagnosis ranged from 1 to 20 years, with 41.18% (n=7) of participants reporting it as unknown.

A comprehensive overview of each participant’s demographic and health-related information is available in Multimedia Appendix 1.

Participants accessed the AI-HEALS intervention via the “Beiyi Sugar Manager” WeChat Mini Program (hereafter referred to as “the mini program”). Four main themes emerged from the analysis (Table 2), each with several subthemes, illustrating the perceived impact of the AI-HEALS intervention on participants’ self-management journeys. Below, each theme and subtheme is described with illustrative quotations to demonstrate how the interpretation was grounded in participants’ accounts.

This theme reflects participants’ experiences in translating increased awareness into behavioral changes, with the mini program serving as a supportive tool.

This theme explores how participants interacted with the technology itself, revealing patterns of use, preferences for certain features, and barriers to full engagement.

Beyond practical behavior change, participants described a positive psychological impact from using the system, which contributed to a greater sense of control and a reduced emotional burden.

Participants’ reports that AI-HEALS deepened their understanding of diabetes complications align with the literature showing that knowledge is an important but often insufficient condition for behavior change in diabetes self-management [27,28]. In this study, awareness often extended beyond general education and became tied to concrete concerns about blindness, cardiovascular disease, and diabetic foot. This suggests that the intervention helped translate abstract risk into a more personally meaningful understanding.

The findings also indicate that the value of AI-HEALS lay not only in transmitting information but also in sustaining attention to diabetes in daily life. Repeated prompts and accessible educational content appeared to keep the condition salient over time, thereby helping to address the knowledge-behavior gap described in previous diabetes education research [29]. In this respect, AI-HEALS differs from more conventional education delivered in discrete sessions [5,6] by providing ongoing reinforcement. This pattern is also consistent with the MTM, particularly its emphasis on participatory dialog as a motivational process in behavior change [26].

At the same time, the nature of this awareness appeared to vary across participants. For individuals with a longer duration of diabetes, such as Patients D and N, the intervention seemed to reinforce and sharpen existing illness understanding, although the content of that awareness differed: one account reflected a broader physiological understanding of vascular risk, whereas another focused more strongly on feared complications such as blindness. By contrast, for a newly diagnosed participant such as Patient O, the intervention appeared to play a more foundational role in shaping initial self-management understanding. This pattern suggests that AI-enabled education may function differently across the diabetes trajectory, serving as reinforcement for some users and as initial structuring support for others. Family context may also have influenced how risk information was received, as prior household exposure to diabetes could heighten vigilance toward complications and self-management [30].

The sense of empowerment described by participants resonates with Anderson and Funnell [31] patient empowerment model, which emphasizes active patient ownership of diabetes management. In this study, the perceived value of AI-HEALS lay not simply in offering recommendations but also in helping participants translate broad self-management principles into concrete, day-to-day decisions.

Dietary change was the most frequently discussed area of improvement, which is consistent with systematic reviews identifying nutrition as one of the most responsive domains in diabetes self-management interventions [10,11]. Participants’ accounts suggested a movement away from rigid restriction toward more flexible and informed judgment, a pattern broadly in line with work on personalized nutrition and tailored self-management support [19]. Importantly, however, these changes did not appear uniform across users. For newly diagnosed participants, dietary guidance seemed to support the establishment of initial management routines. For others with more experience, it appeared to refine existing practices rather than create them from scratch. These differences suggest that the same digital intervention may support both habit formation and habit adjustment, depending on patients’ prior self-management history.

The findings on physical activity similarly indicate that participants valued guidance not just when it encouraged “more exercise” but also when it made activity more appropriate, feasible, and compatible with daily life. The accounts of individualized walking targets and indoor alternatives suggest that users valued optimized guidance over generic behavioral goals. This is in line with emerging evidence that, for adults managing chronic disease, the appropriateness of activity may matter as much as quantity [17]. The case of Patient P also suggests that such support may be particularly relevant for users who manage much of their care independently in everyday life.

At the same time, the findings point to an important tension between personalization and accessibility. Tailored guidance was clearly valued, but some participants also indicated difficulty understanding more complex content. This was especially evident in the account of a participant with lower educational attainment who remained engaged with simpler materials while avoiding more complicated ones. This suggests that digital support may be less effective when information is clinically relevant but insufficiently accessible to users with lower literacy or lower digital confidence. For intervention design, these findings support the importance of tailoring that is not only behaviorally targeted but also sensitive to comprehension and usability.

A notable finding was that many participants mainly engaged with lower-burden features, such as pushed articles and reminders, whereas more active functions, including the chatbot, were used less often. This pattern aligns with international research on digital diabetes interventions, which frequently report challenges in sustaining engagement and note that usage intensity does not necessarily correspond directly to perceived benefit [28,32].

In the present study, passive engagement still appeared meaningful. For some participants, reading articles and receiving reminders was sufficient to support awareness, routine, and behavioral adjustment. This pattern may reflect the effort-reward balance described in prior health technology research, whereby users gravitate toward features that provide clear benefits with minimal cognitive burden [33]. It may also suggest that interactive components, such as chatbots, need to demonstrate clearer added value if they are to become central components of engagement.

Participants’ accounts further indicate that engagement was shaped by the interaction of digital confidence, education, prior routines, and perceived need. The contrast between a participant with lower educational attainment who found “complicated” features difficult and a highly educated participant who used the system selectively because his daily life was already structured suggests that low use does not necessarily mean low usefulness. Instead, selective engagement may reflect different routes to perceived benefit. These findings complicate a simple digital divide narrative: age alone did not appear sufficient to explain engagement, and educational background, routine stability, and prior exposure to structured work or technology may have been equally important in shaping how participants used the intervention.

In addition, the continued engagement reported by several lower-income participants supports the practical value of delivering AI-HEALS through WeChat. By relying on an already familiar platform rather than a standalone app, the intervention may have reduced access barriers associated with downloading, registration, or cost. In this respect, the platform choice itself may have contributed to usability and equity in engagement.

The psychological benefits described by participants—greater confidence, reassurance, and reduced uncertainty—highlight an important dimension of digital self-management support that is often less visible in studies focused primarily on clinical outcomes. These accounts are consistent with the literature emphasizing the emotional burden of diabetes and the value of supportive interventions for coping and self-management [34,35].

Importantly, the perceived psychological benefit of AI-HEALS appeared to arise indirectly through better understanding, clearer routines, and more manageable daily tasks rather than through explicit psychological counseling. Participants described feeling less overwhelmed and more capable when management tasks became more understandable and structured. This suggests that psychological adaptation may be closely intertwined with practical self-management support rather than standing apart from it. In this sense, the intervention may have reduced resistance to the burden of ongoing self-management by transforming diabetes from an overwhelming threat into a more manageable set of daily tasks [36].

This pattern can also be related to the Health Action Process Approach, which distinguishes motivational processes from the volitional support needed to sustain action [37-39]. In this study, educational content appeared to strengthen motivation by reinforcing awareness of risk and outcomes, while reminders and monitoring functions appeared to support the volitional work of action planning, routine maintenance, and coping. The coexistence of behavioral change and reduced uncertainty in participants’ accounts is consistent with this dual-support process.

The findings also suggest that psychological adaptation varied according to illness experience and everyday context. For participants with more established self-management histories, confidence appeared to build on persistence that was reinforced by the intervention. For less experienced users, confidence seemed more closely linked to newly acquired understanding and clearer day-to-day guidance. Family and household context may also have shaped these experiences. For some participants, existing family support may have reinforced the intervention’s effects, while for others—particularly those with less immediate day-to-day support—reminders and accessible guidance may have played a more compensatory role. These observations underline that the psychological value of digital support is shaped not only by technological features but also by the life contexts in which those features are used.

The findings have several implications for implementation in routine primary care. First, AI-HEALS appears most suitable as an adjunct to community diabetes management rather than a replacement for clinicians. Community physicians or nurses could introduce the mini program during follow-up visits, confirm medication and monitoring schedules, and use subsequent visits to review questions or troubleshoot difficulties. Second, the predominance of lower-burden engagement suggests that pushed content and reminders may serve as the most feasible default components for routine implementation, while more interactive functions, such as the chatbot, may work better as optional layers for patients who wish to engage more actively. Third, the variation in participants’ accounts highlights the need for layered and literacy-sensitive design. Brief onboarding, plain-language instructions, and family-assisted use may be especially helpful for some users, while provider training, workflow integration, privacy procedures, and maintenance planning will be important for broader scale-up.

This study contributes qualitative evidence on an AI-enabled diabetes education intervention delivered through a familiar social platform in Chinese community settings. Its main contribution lies in clarifying how participants made use of digital support in practice: through the reinforcement of awareness, the translation of advice into everyday action, selective but meaningful engagement, and increased psychological reassurance. More broadly, the findings suggest that the perceived value of AI-enabled self-management support depends not only on algorithmic tailoring but also on whether the intervention is accessible, low burden, and well aligned with users’ daily routines.

Several limitations should be considered when interpreting these findings. First, the 3-month follow-up captures early perceptions and short-term behavioral experiences only; it does not allow claims about long-term sustainability, and statements in this paper about maintenance should therefore be interpreted as perceived early continuation rather than established long-term adherence. Second, although 17 interviews were sufficient for thematic saturation, the sample remained small and context-specific, which limits transferability beyond Beijing community settings. Third, selection bias is possible because interviewees were drawn from the intervention arm and agreed to discuss their experiences; participants who were more willing to engage with AI-HEALS may have been more likely to participate, potentially overrepresenting favorable views. Fourth, the commercial nature of the dialog platform limited our access to detailed training data, internal validation metrics, and model specifications, constraining technical reproducibility. Fifth, social desirability bias may have influenced participants’ accounts, particularly because interviewers were linked to the broader study team. Finally, missing information on diabetes duration for some participants reduced our ability to examine duration-related patterns in depth. Taken together, these limitations indicate that the findings should be interpreted as explanatory qualitative evidence of perceived mechanisms and implementation considerations rather than definitive evidence of long-term effectiveness or generalizable user responses.

This explanatory qualitative study provided insight into how patients with type 2 diabetes experienced AI-HEALS, a WeChat-based AI-enabled health education system. Participants generally perceived the intervention as helpful for strengthening awareness, supporting everyday self-management, and improving confidence, although experiences differed according to digital literacy, prior routines, and stage of illness. The findings suggest that meaningful benefits may arise even when engagement is concentrated in lower-burden features such as reminders and pushed content. At the same time, the study highlights the importance of cautious interpretation, literacy-adaptive design, and integration with routine primary care. Future studies should test the long-term sustainability of these perceived benefits and further evaluate how algorithmic tailoring, human support, and implementation context interact in larger and more diverse populations.