Conversational Agents Supporting Self-Management in People With a Chronic Disease: Systematic Review

Introduction

Background

Chronic diseases are the leading cause of mortality and morbidity and impose a significant global burden on patients, their families, and the health care system [1,2]. Tackling these challenges requires strategies to empower patients in self-managing their disease effectively. Self-management is a promising strategy to treat chronic diseases, enabling patients to actively identify and solve illness-related problems to prevent complications or reduce disability [3]. Self-management interventions improve health behavior, health outcomes, and quality of life (QoL) and reduce health care use [4,5]. Conversational agents (CAs), web- or app-based computer programs that engage in 2-way interactions by simulating humanlike conversations, have emerged as a promising tool for self-management support [6].

Self-management refers to an individual’s ability to manage symptoms, treatments, physical and psychosocial consequences, and lifestyle changes [7]. Self-management involves role, medical, and emotional management tasks and requires problem-solving, decision-making, resource use, patient–health care provider partnership formation, action planning, and self-tailoring skills [5]. To deploy these skills, patients should be knowledgeable about their illness, symptoms, and available treatments, including self-care options, to make appropriate decisions about their disease management [5]. CAs facilitate self-management by providing knowledge (eg, answering questions about the disease [8]), promoting self-management skills (eg, decision-making based on self-monitoring data), and assisting with self-management tasks (eg, by offering emotional support [8]). CAs can serve as multicomponent technology-based behavior change interventions by incorporating behavior change techniques (BCTs). BCTs are observable, replicable, and irreducible components designed to alter or redirect causal processes that regulate behavior [9]. For example, a CA designed to support mental well-being in people with chronic headaches contains the BCT “goals and planning,” and the technology delivers this technique to the user by assisting them to plan a mindfulness-based activity during the next week [10]. Early evidence highlights that CAs hold significant promises to support self-management of chronic diseases, showing acceptability, usability, and potential efficacy for improving self-management [6,11,12].

CAs are present in diverse forms and functionalities, ranging from systems with pre-programmed responses to more sophisticated embodied artificial intelligence (AI) agents. The field of health care CAs is rapidly evolving, especially with the new developments in AI. A key development is the emergence of large language models (LLMs), deep learning models designed to process and generate natural language text. LLMs excel in generating meaningful humanlike creativity, reasoning, and contextually appropriate language [8,13], allowing users of CAs to converse with the CA as they would with other humans [8]. Well-known examples of CAs that use LLMs are ChatGPT and, more recently, DeepSeek. The rise of AI may further enhance the potential of CAs as self-management interventions.

CAs for chronic disease self-management operate in a multidisciplinary context, integrating intervention technology, (behavioral) psychology, and medical or biomedical sciences. Therefore, rigorous evaluation of their effectiveness in chronic disease management is complex, and a consensus on the best approach is lacking [6,11,12,14]. A comprehensive perspective that incorporates frameworks from all relevant disciplines is essential for effective evaluation [15]. Also, previous reviews typically focused on specific subsets of CAs—such as text-based [6], voice-based [16], AI-driven [11,12,14] or embodied [17] agents—resulting in an incomplete understanding of intervention content and design.

This Study

A previous scoping review [18] outlined various types of CAs used in chronic condition management, along with their characteristics and study designs. Expanding on this work, we adopt a multidisciplinary perspective by integrating frameworks from diverse domains, including the CONSORT-EHEALTH (Consolidated Standards of Reporting Trials of Electronic and Mobile Health Applications and Online Telehealth) checklist, the behavioral intervention technology (BIT) model, and BCTs. Through this synthesis, we provide a comprehensive overview of CAs developed for chronic disease self-management support and their characteristics and evaluation approaches. The findings offer practical insights for developers and researchers, ultimately contributing to the design of robust, evidence-based CA health interventions.

Methods

We adhered to the PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) guidelines for conducting systematic reviews (Multimedia Appendix 1) [19]. Our protocol was not registered.

Search Strategy

The search was performed in the PubMed and Embase databases, covering papers published between January 1, 2018, and April 15, 2024. We prioritized sensitivity over specificity to ensure that all potentially relevant studies were included. Therefore, we based our search string on synonyms for CA and derived no search terms from “self-management” and “chronic disease,” as these are broad concepts that are hard to capture within a set of terms. Our approach resulted in the following search string for PubMed: “Conversational robot*” [tiab] OR “conversational agent*” [tiab] OR “chat bot*” [tiab] OR “chat agent*” [tiab] OR “conversationalagent*” [tiab] OR “chatbot*” [tiab] OR “chatterbot*” [tiab] OR “virtual agent*” [tiab] OR “virtual robot*” [tiab] OR “relational agent*” [tiab] OR “conversational assistant*” [tiab] OR “speech assistant*” [tiab] OR “chat assistant*” [tiab] OR “virtual assistant*” [tiab] OR “chatassistant*” [tiab] OR “AI agent*” [tiab] OR “dialogue system*” [tiab] OR “voice assistant*” [tiab]” OR “ai agent*”[Title/Abstract] OR “dialog system*”[Title/Abstract] OR “voice assistant*”[Title/Abstract]. The search string was adjusted for use in Embase. The initial search was performed in PubMed on May 1, 2023, and an updated search in PubMed was performed on April 15, 2024. To improve the comprehensiveness of our search, we added "AI agent", "dialogue system", and "voice assistant" to our search string in PubMed on March 25, 2025, and repeated our full search in the Embase database on March 27, 2025.

Study Selection Criteria

Inclusion criteria were (1) full-text journal articles; (2) published in English; (3) from 2018 to April 15 2024, to account for rapid technological advancements; and (4) included a CA in; (5) the context of self-management for; (6) chronic diseases in; (7) adults; and (8) reporting primary research findings evaluating efficacy, including pre- and postuse measures.

Chronic disease was defined, according to the Centers for Disease Control and Prevention, as “conditions that last 1 year or more and require ongoing medical attention or limit activities of daily living or both.” [20] Our selection on self-management was guided by the definition of Barlow et al [7] as an individual’s ability to manage symptoms, treatments, physical and psychosocial consequences, and lifestyle changes. We added the criterion that CAs should at least offer information or education, as these components are essential for developing the knowledge and skills necessary to manage chronic diseases effectively [5]. Studies investigating one-time use of a CA, virtual reality agents, mental health, addiction, or neurodiversity were considered out of the scope of this review and were therefore excluded.

Study Selection

One researcher (TFP) screened the studies’ eligibility based on title and abstract, rating them as “potentially relevant” or “not relevant.” Doubtful cases were reviewed by 2 other researchers (SWvdB and NMdV). Subsequently, the full text of potentially relevant articles was assessed by one researcher (TFP). Uncertainties were discussed with the 2 other researchers (SWvdB and NMdV).

Data Extraction and Synthesis

Overview

One researcher (TFP) extracted the data in a standardized form using the abstract, main text, supplemental material, and previous publications describing the CA under study if referred to by the authors. To generate a comprehensive overview of the CA description and study methods, we used the BIT model [10] and CONSORT-EHEALTH checklist [21]. The BIT model defines the conceptual and technological architecture of BITs. CAs for self-management can be considered as BITs, as they function as mobile health or eHealth interventions that support users in changing behaviors and cognitions. The model includes reporting on the intervention aims (the “why”), behavior change strategies (the conceptual “how”), by what elements these strategies are implemented in technology (the “what”), when an intervention component is delivered (the “when”), and intervention characteristics to meet the user’s needs and preferences (the technical “how”). The CONSORT-EHEALTH checklist was developed to improve reporting of eHealth trials and serves as a basis for evaluating the validity and applicability of eHealth trials.

Characteristics of the CA

We described CAs using the “technical how” component of the BIT model and additional characteristics based on the CONSORT (Consolidated Standards of Reporting Trials) checklist, extracting data on mode of delivery, access, embodiment status, conversational techniques, in- and output modalities, and recommended dose.

Mode of delivery refers to how the CA is delivered to the users, such as via an app. Method of access reflects how users access the mode of delivery, such as by downloading the app through a link provided via WhatsApp. We specified conversational techniques, encompassing input processing and response generation methods, as rule-based or AI-driven. We considered the conversational techniques AI-driven if the input was processed using AI (eg, using natural language processing to understand user input, context, and recognize patterns) or AI was used to respond in a humanlike conversational manner (eg, using LLMs) or both. In contrast, rule-based systems do not adapt or learn from user input and provide preprogrammed responses. Input and output modalities refer to how users communicate with the CA (input modality, eg, text) and how the CA communicates with the user (output modality, eg, speech). Recommended dose refers to the planned or recommended amount (eg, duration or content completion, such as one lesson) or frequency (eg, once a day) of exposure to the intervention for optimal effectiveness.

Finally, we described any cointerventions, which are elements of the intervention that are not part of the CA itself.

Integration of BCTs

On the basis of the BIT model, we extracted the clinical aim (eg, reducing blood sugar), BCT clusters (eg, feedback and monitoring), and BCT delivery by technology (eg, providing self-care guidance based on monitored blood sugar levels). A BCT is an “observable, replicable, and irreducible component of an intervention designed to alter or redirect causal processes that regulate behaviour” and BCTs are categorized into clusters [9]. If not reported, 3 researchers (TFP, RJAvD, and SWvdB) identified BCT clusters based on intervention descriptions and BCT taxonomy [9] in consensus meetings.

Evaluation Methods

To describe the evaluation method, we extracted study characteristics (the study’s aim, design, follow-up period, location, participant eligibility criteria related to technical use and skills, intervention and control conditions, and demographics [age, gender, education, and technical skills and use]) and outcome measures. To ensure a comprehensive overview, we reported outcome measures in the following categories: clinical, feasibility, and acceptability outcomes. We adhered to the authors’ classification when provided and self-categorized otherwise. Clinical outcomes included disease-specific, psychological, behavioral, and knowledge and skills measures. Feasibility included recruitment, enrollment, attrition, accrual, exclusion, and retention rates, and outcomes based on use data. We categorized user experience measures under acceptability.

Primary or Coprimary Study Results

The focus of this review is not to draw conclusions on the effectiveness, feasibility, or acceptability of CAs for chronic disease self-management support. However, to get an overview of whether CAs seem to work for the self-management of chronic diseases, we chose to report the results of the primary or coprimary outcomes.

Study Risk of Bias Assessment

We only reported primary study outcomes in this review. Risk of bias for primary outcomes that evaluated efficacy or effectiveness was assessed using the Cochrane Risk of Bias in Non-randomized Studies-of Intervention (ROBINS-I) tool (version 2 [22]) for nonrandomized studies and the Risk of Bias 2 tool [23] for randomized controlled trials. One reviewer (TFP) individually assessed the quality of eligible studies; the assessments were discussed with SWvdB and NMdV.

Results

Overview

Our initial search in the PubMed database on May 1, 2023, yielded 1215 results. A subsequent search on April 15, 2024, resulted in 1114 additional hits. The expanded search of the additional terms yielded 243 hits, and the Embase database search yielded 3286 hits. In total, 5858 articles were screened. After screening, we excluded 5497 articles based on title and abstract. After full-text screening, 25 articles were included. Figure 1 illustrates the inclusion process details.

‎

Characteristics of the CAs

Table 1 summarizes the characteristics of 23 unique CAs included in this study. Two studies evaluated nurse Addressing Metastatic Individuals Everyday [24,25] and Wysa [26,27]. “Chatbot” was the predominant used term (13/23, 57%) [28-40]. Most CAs (18/23, 78%) operated through mobile apps [26,27,29-34,36-39,41-46] as stand-alone apps or integrated into existing apps such as Telegram [36] or Signal [44]. The majority were unembodied agents (14/23, 61%) that communicated through text [26-33,35,36,40,44,46-48]. Embodied agents used speech and animated behaviors [24,25,41-43,45]. CAs were often supplemented with multimedia, such as audio recordings or images. Embodiment status of 4 CAs [34,37-39] and output modalities for 4 CAs [34,37,39,40] were not reported.

Input modalities varied, including natural language text, voice, predefined choices, and combinations. Rule-based conversational techniques were most common (10/23, 43%) [24,25,29,30,33,36,41,42,44-46,48]. Some CAs (4/23, 17%) solely relied on AI-driven conversational techniques [26-28,34,39] while others (4/23) used a combination of AI-driven and rule-based techniques [24,25,32,40,43,47]. Some descriptions of conversational techniques or input modalities were incomplete [31,32,34,35,37-39]. The recommended dose varied from daily interactions to use frequency and duration based on the users’ preference [26,27,32,34,37-39,46] but sometimes lacked (5/23, 22%). Most CAs (15/23, 65%) were part of broader co-interventions, including additional app functionalities, physical and digital information sources, physical self-monitoring devices, and human involvement [24,25,31,32,34-41,43,45,46,48]. Human roles encompassed monitoring symptoms and activity, conducting weekly check-ins, and updating individualized goals [24,25]; monitoring of disabling conditions and addressing users’ questions [31]; responding to questions the chatbot could not answer and delivering contextually tailored messages [32]; reviewing patient data, providing personalized feedback, and answering queries [34]; and managing technical alerts [43]. Clinicians were also involved in responding to flagged participant responses [35,36].

Integration of BCTs

On the basis of intervention descriptions, we identified BCT clusters and delivery methods (Table 2). All authors reported the aim of their CA—except for one [37]. The most frequently identified BCT clusters were “shaping knowledge” (17/23, 74% [24-26,29-33,35,36,38,39,41,44-48]), “feedback and monitoring” (15/23, 65% [24-26,29-31,33-35,37,40,41,43,46,48]), and “natural consequences” (13/23, 57% [24-26,30,33,35,38, 39,41,42,44-47]). “Shaping knowledge” equips individuals with practical knowledge (eg, how to perform a behavior) and cognitive strategies to adopt and sustain desired behaviors. This was delivered mostly by behavioral instructions, for example, how to perform relaxation [46]. “Feedback and monitoring” refer to strategies to monitor or self-monitor behavior and its outcomes and deliver feedback. Monitoring often involved (prompted) assessments of health outcomes [24,25,29-31,35] or behaviors [35,36,48]. Feedback included self-care recommendations based on the data. “Natural consequences” involves strategies that help to understand the relationship between behavior and its consequences. This was delivered through information about the necessity of self-care behaviors and cognitions [26,30,38,46] or medication intake [41,42,44]. “Shaping knowledge” was often combined with “feedback and monitoring,” providing personalized self-care guidance based on the user’s symptoms [24,25,30,31,47]. In addition, “associations” were often incorporated (14/23, 61% [24,29-31,33,34,36,38,40-42,44-46]), often with prompts to use the CA. “Social support” was delivered by facilitating contact with health care providers [35]; videos of fellow patients [29,45]; advice on and arranging social support and practical help from friends and relatives [46]; or being the actor that provides the social support [24,25,30,31,33,45,46].

One study [46] integrated the BCTs into their intervention with explicit examples of delivery by technology. All others lacked explicit reporting on which and how BCTs were delivered. The details in which the CA was described varied. Ambiguous intervention descriptions hindered the classification of BCTs. As an example, “recommending evidence-based elements from cognitive behavioral therapy, behavioral reinforcement, and mindfulness, among others” [27] suggests the integration of many BCTs but does not allow us to identify which and how BCTs are delivered. The insufficient reporting on CA content made distinguishing “shaping knowledge” from “natural consequences” challenging [26,30,35,39,47]. For example, “providing information about medication” could refer to instructions on its use (“shaping knowledge”) or explaining the effect and importance of adherence (“natural consequences”).

Evaluation Methods

Study Characteristics

In total, 24 unique studies were reviewed (Tables 3 and 4), as the findings of 1 study were published in 2 separate articles [24,25]. Most studies (11/24, 46%) were randomized controlled trials [24,25,28,29,33,36-38,41-43,46]. Most (14/24, 58%) were early-phase studies, such as pilot studies, initial trials, or exploratory studies [24,26,29,30,32,33,35,37,41,42,44,45,47,48]. Follow-up periods ranged from 3 weeks to 1 year. More than half of the studies (13/24, 54%) reported primary or coprimary outcomes [28,29,31,33-35,38,39,41,43,45,46,48]. Primary or coprimary outcomes were predominantly clinical outcomes, and user satisfaction was assessed as a primary outcome once [39]. One study [29] lacked reporting of which outcome was primary. Many studies (17/24, 71%) selected participants on technical skills (eg, ability to install apps or communicate online) [37,40,41], device access or possession [28-30,32,33,35,45,46,48], or both [34,36,38,44]. CAs were accessible to the participants primarily in addition to usual care, although this was not always specified. In one study [28], the CA replaced parts of usual care. Some studies (7/24, 29%) involved active research team engagement by checking in with their participants to, among others, troubleshoot (technical) issues [24,25,30,37,38,40,43,45]. Preuse instructions were provided in some of the studies (9/24, 38%) [24,25,28,32,35,36,41,44,45,48]. Control conditions were typically usual care. Three studies chose an active control group in which participants received a nurse-led education program or discussion with nurses [28], a self-help book [47], or intensive specific health guidance [37]. One study [35] assigned their participants to the control or intervention condition based on CA use. Only 8% (2/24) of the studies [26,48] reported the technological experience of their participants.

Efficacy or Effectiveness Outcomes

Clinical manifestations of the disease were assessed most often (11/24, 46%). These include physical function examined with a chair test in cancer and a survey in chronic pain [25,26]; symptoms of cancer, pain, headaches, overactive bladder, and irritable bowel syndrome evaluated with surveys or a diary [25,26,30,33,38,46,48]; blood glucose levels determined with blood tests in diabetes [34,35,43,44]; BMI or weight assessed by a general practitioner in clinic or self-monitored data using a weighing scale in diabetes type 2, obesity, and hypertension [34,37,43]; and blood pressure examined using a blood pressure monitor in obesity and hypertension [37]. In addition, medical complications were assessed using unspecified methods, such as infection rate from peritoneal dialysis for chronic kidney disease [39] and department emergency visits and unscheduled hospitalizations of patients with cancer [31]. Side effects of chemotherapy for cancer were measured once with a survey [28].

More general outcomes were measured with a variety of (disease-specific) surveys. General outcomes encompassed the (health-related) QoL of patients with colorectal cancer, irritable bowel syndrome, atrial fibrillation, diabetes, and an overactive bladder [38,40,41,43,48] and mental health–related outcomes, with anxiety or depression being the most frequent, in patients with cancer, chronic pain, irritable bowel syndrome, diabetes, headaches, or an overactive bladder [25-27,29,38,43,44,46-48].

Self-management [30] or related concepts such as self-care confidence [35] and self-efficacy [45,46] were assessed using standardized questionnaires [30,45,46] or by adding one question to the test battery [35]. In addition, 13% (3/24) of the studies measured the effect of education with surveys assessing communicative literacy [32], health literacy [45], and blood pressure self-monitoring knowledge and skills [36]. Moreover, 8.3% (2/24) of the studies evaluated the participants’ intention to change their behavior [33,46]. Some (8/24, 33%) studies assessed changes in behavior with step counts using a pedometer [24,37], healthy habits using a survey [37], changes in food consumption and physical activity using a survey [40], application of BCTs using a survey [46], adherence to medication using pill counts [42,45], questions about (changes in) adherence [41,44], adherence to monitoring using number of correct measures [36], and monitoring quality by comparing self-assessed measurement to golden standards [41,42,44,45]. The effectiveness of self-care behaviors was assessed once [28]. The impact of the intervention on health care use was measured twice [31,44].

Feasibility Outcomes

One study [30] reported accrual and attrition rates as feasibility measures. Another study [25] defined feasibility as at least 50% of participants logging in for 30 out of the first 90 days. Most studies did not define feasibility but assessed metrics that we considered as feasibility. The majority (17/24, 71%) reported participant flow data from which enrollment, attrition, exclusion, or retention rates could be inferred [24-26,29,32-35,37,38,40-47]. Many studies (17/24, 71%) examined user-data metrics to assess intervention adherence [33,37,41], engagement [27,29,34,35,40,46], adoption [43], activation [35], intended use [46], and the extent of use [26,31,34,35,40,46]. “Intervention adherence” was assessed by conversation replies [33] and self-monitoring uploads [37]. Measures of “engagement” varied widely; for instance, one study [29] defined “engaged sessions” as the number of completed sessions, while other studies considered the quality of data reported by the participants [40] or the percentage of answered conversational turns [35]. One study [37] studied a similar measure, the ratio of conversations answered by the user and all conversations the CA initiated, rendering this as intervention adherence. Similarly, “extent of use” was measured inconsistently, varying from the number of consultations [31,35] to time spent in-app [46]. Some used “extent of use” to assess “intervention exposure” [26] and “received dosage” [38]. Others included use metrics without further specification of what they intended to measure [26,31,35,45] or reported what they measured without describing how [41]. In addition, 8.3% (2/24) of the studies did not evaluate feasibility [28,36].

Acceptability Outcomes

Some (10/24, 42%) studies reported acceptability. Acceptability was assessed with nonstandardized surveys [34,39,45], the extent of app use [45], usability using the System Usability Scale [32], or intervention satisfaction by a usability and acceptability survey combined with an interview to collect user experience and feedback [30]. One study [46] reported “engagement and acceptance,” including intention and commitment to change behavior, working alliance, participants’ sensitivity to triggers, and tendency to avoid triggers. Another study [44] reported on “patient use and acceptance of CAs,” encompassing use of CA (tools and reminders), acceptance (#users that did not uninstall signal), and usefulness (#times the patient was not understood by the CA). One study [25] defined an acceptability threshold as 50% of patients agreeing to participate. Some of these studies included additional metrics that we classified as acceptability, such as usability [25] and working alliance [33] using standardized questionnaires, helpfulness of the intervention [25], and perceived enjoyment using one single question [46].

Some authors did not report on acceptability but reported measures that we classified as acceptability, such as usability assessed with standardized questionnaires (ie, Chatbot Usability Questionnaire [28] or System Usability Scale [40,47,48]), CA feedback [29], helpfulness of the intervention [29], or user experience and satisfaction [35,36,40] using self-compromised surveys consisting of (single) questions, ratings, or the User Experience Questionnaire [40]. In addition, 13% (3/24) of studies interviewed their participants about the user experience [30,42,44]. One study [31] referred to patient satisfaction in their results, but the method description lacked, and another reported measures related to user experience but left them undefined [42]. The client-satisfaction questionnaire was used to address attitudes toward the intervention in one study [47] and usability in another [24,25].

Primary or Coprimary Study Results and RoB

Most study findings support the effectiveness of the CAs [28,31,34,38,41,46,48]. Their interventions were effective in elevating (health-related) QoL in patients with atrial fibrillation [41], diabetes [43], inflammatory bowel syndrome [38], and overactive bladder [48]. Other studies showed that their interventions improved mental well-being in patients with chronic headaches [46] and diminished symptom severity in patients with inflammatory bowel syndrome [38]. In cancer, the interventions led to fewer emergency department visits of patients with gynecologic cancer [31] and diminished chemotherapy-related side effects and improved effectiveness of self-care behaviors in breast cancer [28]. In diabetes, 8% (2/24) of studies found improved blood glucose levels [34,35] while another study found no effect on blood glucose levels [43]. Pain-related impairment in chronic pain did not differ between the intervention and control group [33]. Another CA intervention did not significantly improve medication adherence in young men who were HIV positive and have sex with men [45]. Furthermore, one study showed acceptability of their CA with a high overall user satisfaction (1/24, 4%) [39].

All included studies that reported primary outcomes were rated as having a high, serious, or critical RoB (Multimedia Appendix 2 [28,31,33-35,38,41,43,45,46,48]. Missing outcome data were a shared concern in randomized controlled trials and nonrandomized trials: 29% (7/24) of studies reported substantial missing data without providing reasons or using appropriate methods to account for the missing data [33-35,38,43,45,46]. In the randomized controlled trials, the most common source of high RoB was in outcome measurement [28,33,38,41,43,46]. Given the nature of the interventions, participants were often unblinded, and primary outcomes—such as QoL—were self-reported and susceptible to influence by participants’ expectations or beliefs about the intervention. In nonrandomized trials, bias also often emerged from the lack of controlling for key confounding factors, such as age, digital skills, comorbidities, and disease stage [34,35,45,48].

Discussion

Principal Findings

This review provides a comprehensive overview of the current state of intervention descriptions and evaluation methods for CAs supporting people to self-manage chronic diseases. Despite promising development, significant gaps remain in intervention descriptions and evaluations. We discuss these findings in further detail, highlighting existing strengths and directions for future research.

Characteristics of CAs and Integration of BCTs

Consistent with prior reviews [11,14], nearly all CAs were designed for specific diseases, such as cancer and diabetes. This allows CAs to be designed attuned to the disease-specific needs of people. Interventions aimed at improving adherence, self-care confidence, and behavior. However, intervention descriptions were inconsistent or incomplete, particularly conversational techniques, hindering understanding how responses are generated and insights into conversational flows, limiting replicability. To illustrate, one study reported about AI-powered decision support [34], but because the description of the conversational techniques is lacking, it is unclear how this functionality works. This aligns with earlier research highlighting inconsistent documentation of AI methodologies in CAs for chronic disease self-management [11]. Despite the rapid advancements in AI, the number of AI-driven CAs identified in our selection was lower than expected, and many AI-driven CAs relied on basic natural language processing rather than more advanced AI capabilities. This highlights the early-stage nature of AI integration in health care CA research. This is further supported by our observation that studies investigating more advanced AI-driven CAs evaluated for response accuracy rather than their effects on disease self-management.

BCT clusters “feedback and monitoring,” “shaping knowledge,” “natural consequences,” and “associations” were most prominently integrated. Findings from other studies examining eHealth interventions underscore the prevalence of feedback and monitoring, shaping knowledge, and associations [51,52]. “Associations” were often prompts to CA use, which can also be considered a persuasive feature of technology. Furthermore, CAs demonstrate the unique capacity to deliver “social support” by serving as a source of support themselves. Explicit reporting on BCT integration is often lacking. Only a few studies explicitly reported the theoretical frameworks guiding their interventions or the BCTs used, a concern echoed by a previous review on BCTs in self-management interventions for chronic obstructive pulmonary disease [53]. This raises the question of whether the intervention was designed using established theoretical frameworks or if such frameworks were absent. The latter could potentially result in less effective interventions. Conversely, if a theoretical basis exists, insufficient reporting generates ambiguity in two critical areas: (1) the chosen BCTs and (2) the role of technology in delivering these techniques. Clear reporting of the BCTs and their delivery—such as Ulrich et al [46]—is crucial to evaluate the theoretical underpinnings, providing insights into the mechanism behind behavior change and improving replicability and comparability.

Methodological Limitations in Intervention Evaluation

The primary findings of the studies supported the potential of CAs to improve self-management and health outcomes. A previous review similarly suggested that AI-powered chatbots may contribute to better health outcomes; however, the evidence was limited due to insufficient technical documentation [12]. These findings should be interpreted with caution, as all included studies exhibited a heightened RoB—a concern raised by earlier literature as well [12].

The field of CA interventions for chronic disease self-management remains in an early stage [11,14], with many studies being exploratory. Also, the heterogeneity in study designs, variability in taxonomy, and use of broad (not unified) outcomes present a major challenge [11,14,18]. Consequently, the possibility of conducting a meta-analysis is prevented [12,47]. Many studies relied on nonstandardized acceptability surveys. Clinical outcomes were often self-reported. Although self-reported clinical outcomes are valuable for capturing patient experiences, they are also prone to subjectivity, recall, and social desirability biases [6]. Combining self-reported outcomes with objective measures, such as blood glucose levels [34,44], offers a more comprehensive perspective on the patient’s health, reduces bias, and enhances reliability. Another issue is that few studies measured outcomes related to self-management (eg, knowledge), despite their critical role in improving health, highlighting a significant gap we should overcome to understand how CAs contribute to improving health outcomes.

Selection bias is a concern. Participants were often required to possess a compatible device. This eligibility criterion excludes less technology-oriented individuals or individuals with lower income, thereby limiting the generalizability of findings to more diverse populations. Dworkin et al [45] tried to overcome this bias by providing a study loaner phone. Moreover, minimal attention was given to patients’ technical proficiency, a factor that directly impacts engagement and outcomes, complicating the understanding of whether limited effects are due to the intervention or users’ technological comfort. Recommended would be at least an approach that describes the technological proficiency of the study group using standardized questionnaires [48].

While CAs are often proposed as solutions to health care resource scarcity, studies provided minimal details about the care environment where the interventions were implemented, and it was often unclear whether they opted to supplement or replace usual care. Also, the associated time investment of human involvement was not reported. Both hamper the understanding of the resource implication of implementing these interventions.

Limitations

This study was not without shortcomings. Only 2 databases were used in our search. Consequently, we might have missed some potentially relevant articles. However, our search in Embase and PubMed gives a good presentation of the articles published within the biomedical field—which was our key focus. Also, we included a considerable number of papers, and our main findings were consistently observed throughout the articles included. Therefore, even if some articles were missing, it is unlikely that additional articles would alter our results. The article inclusion and data extraction process were conducted by a single researcher, which increases the risk of subjective bias. To mitigate this, the inclusion and selection process was frequently discussed and reviewed with 2 additional researchers whenever uncertainties about an article’s inclusion arose, and the extracted data was cross-checked with the original articles to ensure accuracy. Moreover, the categorization of BCT clusters involved a degree of interpretation, which could affect the consistency of the findings. However, by categorizing BCTs in consensus meetings with 3 researchers, we minimized this risk. The limited detail in intervention reporting may have led to an underestimation of the presence and diversity of BCTs in this analysis. This concern is supported by the observation that the only study [46] that reported what and how BCTs were delivered included the most (diverse) BCT clusters. The same applies to the classification of outcome variables and conversational techniques, which require subjective judgment. However, to enable comparability across studies, this categorization was the preferred option.

Recommendations for Future CAs in Health Care Research

To improve the rapidly evolving field of research on CA for chronic disease self-management, our review highlights several areas for advancing the evaluation of CAs.

Standardize Reporting

Adopt frameworks like CONSORT-eHealth to ensure comprehensive documentation of the study and the BIT model for detailed reports of CA intervention components, separating between BCTs and their delivery by technology. We believe this enhances replicability and comparability across studies and allows us to evaluate what (combinations of) BCTs are most effective for chronic disease self-management and how they should be delivered to the user. This contributes to more robust and transparent evidence, including a better understanding of the working mechanisms of CAs as self-management interventions. If a study aims to further characterize the dialogue management systems of CAs, the framework by Laranjo et al [54] offers a valuable basis to refine CA characterization and technical evaluation.

Enhance and Broaden Methodological Rigor

Use validated, standardized evaluation methods for feasibility, acceptability, and clinical efficacy. Also, add objective disease-specific and technology (eg, log data for reach and engagement) measures alongside self-reported outcomes to strengthen the evidence of clinical effectiveness and adoption of the CA. Objective measures are essential to minimize bias, given the challenges of blinding participants in CA interventions. The use of shared taxonomy and clear definitions about the intended measure are key. This will improve reliability and comparability across studies and allow meta-analysis across multiple studies. As attrition is typically high in eHealth trials [21,55], including those in our assessment, robust strategies for handling missing data—such as sensitivity analysis [23], multiple imputation [56] using relevant covariates, and transparent reporting of missingness and its causes [23]—are essential to reduce bias. In addition, greater attention should be given to key factors, such as age, digital skills, health or eHealth literacy, comorbidities, and disease stage, which may influence CA engagement and effectiveness. Incorporate methodological frameworks of multiple domains, such as human-centered design research, to further CA research into a more agile field of science [57].

Conduct Real-World Evaluations

This includes evaluating the role of CAs in supplementing or replacing traditional care, reporting on associated resource implications, and examining their impact on health care use. We recommend evaluating the implementation of CAs for self-management using established, well-defined outcomes with respect to BIT use in routine practice settings [58] and using real-world evaluation frameworks that consider the complexity of the health care system as an evaluation context [59,60].

Leverage Advanced AI

Expanding the use of AI-driven personalization and decision-support functionalities may enhance user engagement and intervention efficacy. Improved personalization based on individual data and dynamic adapting and learning from the interaction can support self-management by fostering better decision-making (eg, predicting behavioral patterns and identifying potential challenges and proactively suggest solutions) and self-tailoring (refining suggestions and ensuring interventions remain relevant to the user). Furthermore, incorporating sentiment analysis into CAs enhances their capacity to deliver social support that is both empathetic and contextually appropriate [18].