In this study, we analyzed data from the 24 included studies and created an evidence gap map (Figure 3) to illustrate the core AI technologies used in each study.

Among the included studies, NLP was the most widely used technology; nearly all studies involved text or dialogue processing for knowledge transfer, interactive learning, and behavioral guidance in hypertension health education. For example, rule-based dialogue agents and virtual nurse chatbots are typical applications of NLP technology, providing patients with personalized education and self-management support [34,36].

The introduction of generative AI marks a new development in this field, with its applications increasing significantly since 2023. In particular, the emergence of LLMs, such as the study by O’Hagan et al [41], which first evaluated the application of ChatGPT in hypertension health education, has led to a rapid increase in the number of LLM-related studies since 2024. This indicates that generative AI holds potential for improving educational interactions, personalizing information, and generating content.

Regarding the distribution of technology categories, expert system and knowledge base remain prevalent, indicating that traditional AI methods still hold value in the management of structured educational content and knowledge transfer. Furthermore, NLP and ML technologies are frequently applied together in many studies, suggesting that text processing and algorithm-driven intelligent assistance are jointly supporting the implementation of hypertension health education.

The evidence gap map reveals that, despite the widespread application of NLP and ML technologies, the use of generative AI is still in its infancy. The integration of specific AI technologies with certain areas of health education, such as symptom tracking, clinician-patient communication, and appointment management, remains relatively scarce, providing a clear direction for future research. Overall, AI applications in the field of hypertension health education are rapidly evolving from rule-based and traditional algorithms toward generative and interactive technologies.

As shown in Figure 4, we demonstrate how AI technology is applied to health education for hypertension.

Based on data extracted from 24 studies, we have categorized the applications of AI in hypertension health education into three types (Tables 2 and 3): (1) rule-based health education, (2) data-driven adaptive health education, and (3) generative AI–driven health education. Among these, rule-based health education and generative AI–driven health education are the most common. Rule-based health education is characterized by content determined by predefined rules and knowledge structures, thereby offering determinism and interpretability. Data-driven adaptive health education uses ML models trained on user behavior and physiological data to provide personalized intervention recommendations or real-time feedback to individuals while adaptively adjusting educational content. Generative AI–driven health education leverages LLMs, combined with techniques, such as retrieval-augmented generation (RAG) and intelligent agents, to support open-ended question-and-answering and personalized dialogue, thereby delivering health education. It is important to distinguish between studies that evaluated LLM outputs against reference standards (proof-of-concept accuracy assessments) and those that deployed LLM-based tools in interactive educational settings with real patients; the majority of generative AI studies in this review fall into the former category, reflecting the early developmental stage of this application scenario.

These 24 studies were categorized into 3 distinct application scenarios and exhibited clear patterns in terms of both time span and research methods. Rule-based systems (n=8) spanned the entire review period (2020‐2025) and accounted for 4 of the 6 RCTs. Data-driven adaptive interventions (n=3) began to emerge in 2024, primarily using nonrandomized designs. Generative AI–driven applications (n=13) have dominated the literature since 2023 but remain concentrated in the proof-of-concept phase. In total, 8 of the 11 studies used quantitative descriptive designs, and none were evaluated via RCTs.

We summarized the characteristics of health education based on data from the included studies, as shown in Table 4.

After analyzing the included studies, we found that their primary research objectives could be broadly categorized into 2 types: one aimed to evaluate the clinical efficacy of AI-based hypertension health education interventions, while the other assessed the performance of LLMs or AI-based hypertension health education systems. Studies in the first category reported outcomes related to clinically relevant measures. The second category of studies involved the measurement of common computer-related metrics, such as system usability, the accuracy of health education information, and readability. The outcome characteristics of the included studies are shown in Table 5.

We categorized all studies according to the definitions in the Digital Health Scorecard Framework [58]. The Digital Health Scorecard Framework encompasses 4 domains: technical, clinical, usability, and cost. In this framework, technical refers to evaluating whether a digital health solution can accurately and precisely deliver its claimed functionality, including considerations such as security, interoperability, and system architecture. The clinical dimension focuses on rigorous evaluation of evidence to validate whether the solution has demonstrated capacity to improve specific health outcomes, requiring comparison against relevant clinical gold standards; usability concerns the ease with which users can accomplish intended tasks with minimal effort, encompassing aspects like effectiveness, learnability, and user satisfaction; and cost includes user access fees, technology life-cycle investments, and integration expenses within clinical workflows.

In the included studies, accuracy was the most frequently evaluated metric. The performance of AI-based hypertension health education systems needs to be compared against reference standards, such as clinical guidelines, medical textbooks, and expert-developed question sets [40-43,46,48,49,52,55,56]. The completeness and comprehensiveness of health education content are also key metrics [46,48,55,56]. In addition, readability and credibility were evaluated [41], both of which are related to the accuracy and reliability of information delivery. One study further assessed readability using objective measures including the Flesch-Kincaid grade level and response length, providing additional evidence regarding the accessibility of AI-generated educational content [43].

The system’s performance in the face of external disturbances is reflected in the assessment of robustness when handling unexpected inputs [55]. According to the framework, privacy and security are explicitly listed as part of the technical evaluation, which aligns with the current requirement that AI applications in medicine must protect patient data [55]. Other relevant technical metrics include fidelity and medical fidelity [49], intent recognition and data extraction accuracy [52], applicability [48], and the appropriateness and professionalism of responses. Vinufrancis et al [44] additionally evaluated information quality and reliability using the Global Quality Scale and the modified DISCERN instrument, highlighting the importance of assessing the educational value and trustworthiness of AI-generated health information. Domain expert evaluations [49,52] and interrater reliability [40,44,55] serve as complementary methods for validating technical performance.

Blood pressure, as a key clinical end point, has been reported in numerous studies, with measurement methods including office blood pressure, home blood pressure, and 24-hour ambulatory blood pressure monitoring [34,36,37,39,45,47,50,51,53,57]. This underscores the need to compare any AI system used in clinical practice against established clinical gold standards. Other cardiovascular-related parameters, such as heart rate [36,47], subendocardial viability ratio, and brachial-ankle pulse wave velocity [47], as well as cardiopulmonary exercise test parameters and 6-minute walk distance [57], were also included in the assessment of clinical outcomes.

The process indicators emphasized by this framework are particularly evident in the assessment of medication adherence, specifically treatment adherence or adherence to clinical guidelines. The indicators used include medication possession rates, self-reported adherence, and the Morisky Medication Adherence Scale [34,37,47,50]. The number of antihypertensive medications and medication adjustments [37] are also used as process-related indicators. In one study [53], blood pressure control behaviors served as the primary outcome measure, and behavioral monitoring was reported—including blood pressure monitoring frequency, daily self-weighing, pedometer use, and dietary approaches to stop hypertension diet adherence [39,47]. Self-reported health behaviors [34,39] and salt intake [36] also meet the definition of process measures.

Other relevant clinical indicators include self-efficacy related to blood pressure [34] and physical activity [57], as well as knowledge about hypertension [37,47,50], academic performance [47], weight, BMI [34,36,39,47,57], waist, hip circumference [57], grip strength [57], mental health scales [57], and app-guided behavioral adherence scores [36].

Among the included studies, cost assessment was the least well-developed dimension, which is consistent with the framework’s observation that comprehensive cost estimation is a complex process. None of the studies included in this review reported formal cost-effectiveness analyses, technology lifecycle costs, or the long-term economic impacts of clinical efficacy improvements.

The included studies used various methods to assess usability. Griffin et al [38] used the System Usability Scale to capture users’ subjective evaluations of ease of use and likability. Task-level efficiency was measured by task completion rates, user error rates, number of clicks, number of voice commands, and interaction duration per task [38]. Chatbot response times [52] and qualitative feedback regarding their strengths and weaknesses were used to assess user experience. Two additional studies also assessed user experience using the “Chatbot Usability Questionnaire” and a custom-designed satisfaction questionnaire [50], combined with online surveys covering dimensions such as practicality, satisfaction, willingness to use, behavioral responses, and unmet needs [51].

Patient engagement is also a key indicator of usability, with metrics including user retention rates over a 24-month period [51], chat frequency [50], and app engagement rates [36,45]. For example, training time and the proportion of users able to operate the chatbot within 5 minutes [50] are also used to assess usability. Explorations of chatbot cognition, perceived use frequency, and barriers and facilitators [35] provide insights into user needs, reflecting the necessity of user-centered design for AI systems.

Usability for clinicians was assessed using the Physician Software Satisfaction Questionnaire and patient management time [54], indicating that AI-based hypertension health education systems should not increase the burden on clinical staff. The patient-oriented Mobile Health App Usability Questionnaire covered ease of use, satisfaction, system information layout, and practicality [54], while also incorporating patient-perceived comprehensibility, credibility, and emotional support [56], thereby reflecting usability at the subjective level. The evaluation of BP Coach in terms of practicality, simplicity, completeness, and clarity [54] also covered various aspects of effective and practical design.

This scoping review synthesizes the existing evidence on the application of AI in health education for hypertension. A total of 24 studies met the inclusion criteria, and this scoping review yielded 3 key findings.

First, NLP and ML form the technological foundation of AI in hypertension health education. The study identified 3 application scenarios: rule-based health education, data-driven adaptive health education, and generative AI–driven health education. The evolutionary trajectory from rule-based to generative AI methods reflects the overall trend in the application of AI to hypertension health education, with research based on LLM surging since 2023. However, the use of generative AI in hypertension health education is still limited to the proof-of-concept phase. While some studies have used quantitative descriptive designs, none of them have undergone clinical validation through RCT.

Second, health education is generally embedded within multicomponent AI platforms rather than implemented as a standalone intervention; in all 6 RCTs, the educational module was provided concurrently with monitoring, reminder, or clinical decision support functions, making it impossible to assess the specific effects of the educational intervention in isolation.

Third, the included studies used a multidimensional evaluation framework covering the 4 domains of the Digital Health Scorecard, but significant asymmetry was observed. Technical metrics (accuracy, completeness, and readability) were reported most frequently, clinical outcome metrics (blood pressure and medication adherence) were reported in more than half of the studies, usability metrics (satisfaction and engagement) were reported less than the clinical outcome metrics, and cost assessments were rarely reported. This evaluation gradient, ranging from robust technical validation to a complete lack of economic analysis, indicates that the current evidence base is insufficient to support real-world implementation decisions. This is due to the absence of large-scale RCTs, short follow-up periods (mostly ≤12 weeks), and a general lack of cost-effectiveness data. These 3 points collectively constitute the 3 most urgent gaps in the current evidence base that need to be addressed.

Rule-based health education systems, exemplified by early platforms such as the coaching app [34] and the digital therapeutic system [36], represent how AI technology was initially applied to deliver health education. Their strengths lie in content certainty and clinical interpretability: every piece of educational information can be traced back to predefined rules or validated knowledge structures, making these systems inherently traceable and suitable for integration into clinical workflows where transparency is paramount [37]. However, this certainty comes at the expense of flexibility; such systems cannot address novel queries outside their programmed knowledge domain and have limited capacity for personalization beyond predefined hierarchical criteria. This limitation echoes a longstanding critique in health education research that standardized, noncustomized materials often fail to meet the specific needs and circumstances of individual patients [59].

Data-driven adaptive approaches use ML to analyze user behavior and physiological data to enable dynamic educational functions. The personalized coaching platform developed by Leitner et al [45] is a prime example of this approach. By training personalized models, the platform identifies individual-specific associations between lifestyle factors and blood pressure. By providing dynamic feedback and adaptive health recommendations, these systems address the lack of personalization inherent in traditional rule-based frameworks. However, they also face their own challenges: achieving reliable personalization requires vast amounts of individual-level data, which raises concerns about data privacy. Furthermore, the effectiveness of these systems depends largely on users’ consistent use of monitoring devices, which may limit their applicability in settings where digital literacy or access to devices is limited [50].

The most transformative development has been the emergence of generative AI and LLM. Recent literature reviews on chronic disease management indicate that the use of LLMs in patient education has experienced explosive growth [22,23], a trend consistent with our findings: studies based on LLM have increased rapidly since 2023. By supporting multiturn natural language dialogue, these models overcome the one-way limitations of traditional health education, thereby enabling patients to actively seek information rather than passively receive it. However, LLMs also pose significant challenges to the reliability of educational content. As noted in evaluations of medical AI, LLMs remain prone to generating medical claims that appear plausible but lack evidence [60,61]. Within this review, LLM chatbot evaluations for hypertension education have demonstrated generally acceptable factual accuracy, yet revealed persistent concerns regarding readability and expert assessment consistency [42-44]. Patients, however, lack sufficient discernment, a factor that is particularly critical for the implementation of hypertension health education, which requires precise guidance on medication use and lifestyle adjustments [62,63]. To address this, recent studies advocate for the use of hybrid architectures to improve this situation by anchoring LLM to structured medical knowledge bases [64-66]. This aligns with the findings of this review, as multiple studies have used this approach to reduce hallucinations and enhance the accuracy of educational content [49,55,56]. Emerging evidence also suggests that performance gains can be achieved through prompt optimization alone. Li et al [67] demonstrated that structured prompt engineering strategies, particularly guidance-based and self-consistency approaches, substantially improved the accuracy and guideline adherence of LLM in hypertension treatment decision-making. This finding further highlights the potential of combining model architecture, knowledge integration, and prompt design to enhance the reliability and safety of AI-generated hypertension education.

This review found that classifying outcome metrics reveals heterogeneity in current evaluation practices. The technical dimension has consistently received significant attention, with its accuracy, completeness, readability, and reliability typically validated based on clinical guidelines or expert judgment [40-44,46]. This focus aligns with digital health validation pathways, which regard technical reliability as a prerequisite for clinical deployment [58]. However, a significant portion of the literature on LLM technologies for health education in this review remains at the proof-of-concept stage, assessing whether models can generate accurate information, rather than whether the information provided can change patient behavior or improve health outcomes. Currently, clinical validation of the use of LLM for hypertension health education is still in its early stages.

Blood pressure control is the most frequently studied clinical end point; however, most studies have relatively short follow-up periods, and only a few have used randomized designs to validate hard clinical end points based on established hypertension criteria [36,37,47]. Furthermore, the AI technologies used in these studies rarely involve LLM.

Even more striking is the complete absence of cost assessments across all 24 included studies. This omission aligns with evidence from implementation science, which indicates that even technically effective and user-friendly interventions struggle to achieve scale without data on implementation costs, technology lifecycle investments, and long-term health economic impacts [68,69]. Although usability evaluations covered user satisfaction, engagement rates, and system usability scales [38,50,54], these assessments primarily reflect short-term acceptance rather than long-term behavioral maintenance. Furthermore, research has found that content generated by LLM far exceeds the recommended reading level for patient education materials [41,43], raising fundamental equity concerns: if populations with lower health literacy, who are often older, less educated, and from lower socioeconomic backgrounds [70], cannot access AI-powered health education, these tools may exacerbate rather than alleviate existing inequalities in hypertension control.

The AI systems included in this study share a notable characteristic that warrants special attention: health education rarely exists as a standalone intervention. Instead, it is embedded within multifunctional AI-driven platforms as one of their components [36,47,54]. This embedded nature has significant implications for effect attribution. When a multidimensional AI intervention significantly reduces blood pressure, it is impossible to distinguish the specific contribution of the educational component from the effects of the monitoring, reminder, or decision-support components. This challenge reflects a long-standing methodological debate regarding the evaluation of complex interventions, namely, that synergistic effects between components may exceed the sum of their individual contributions [71].

This review found that the geographic distribution of the included studies exhibited distinct characteristics, with a significant concentration of research in China and the United States. This may reflect the policy momentum and technological investment in these 2 countries regarding the use of digital tools to manage hypertension. However, this concentration also raises an important question: whether the study findings are generalizable across different health care systems, reimbursement models, and sociocultural contexts. In total, 5 studies were conducted in low- and middle-income countries, 2 of which used relatively simple AI technologies rather than LLM. The study from Nigeria [50] demonstrated that a WhatsApp-based generative AI chatbot not only achieved high user satisfaction but also improved medication adherence in a resource-limited cardiology clinic. This study provides a key proof of concept but also highlights the significant gap in relevant evidence in low- and middle-income countries.

Research on multicultural and multilingual adaptability is almost entirely lacking. Only 2 studies have examined language-related performance differences, with one finding that blinded hypertension specialists rated English responses more favorably than Japanese responses [40], and this performance difference may align with the fact that current LLMs are primarily trained on English-based data. Beyond linguistic factors, although self-management behaviors for hypertension, including dietary choices, physical activity patterns, medication attitudes, and health care–seeking behaviors, are deeply influenced by cultural context [72,73], studies are needed to explore deeper dimensions of cultural adaptation.

This review found that only a few studies have systematically explored this topic. Montagna et al [52] proposed and compared various privacy-preserving architectures, revealing the inherent trade-offs between performance and data protection. Building on this work, Aguzzi et al [49] extended research on on-device deployment using enhanced RAG mini-language models, pointing the way toward the development of privacy-preserving AI tools that ensure data security without sacrificing performance. The design of digital health interventions must involve the active participation of the target population and be user-centered [74,75]. However, our findings indicate that this requirement has not yet been systematically implemented in AI-driven hypertension health education.