Metabolic syndrome (MetS) is a highly prevalent condition that affects up to approximately 30% of adults aged >65 years worldwide [1]. It consists of multiple symptoms, namely abdominal obesity, glucose intolerance, hypertension, and high cholesterol as well as low high-density lipoprotein [2]. It is associated with a substantially increased risk of premature morbidity and mortality from diabetes and cardiovascular disease (CVD) [2]. Low levels of physical activity (PA) are strongly associated with MetS, including obesity and overweight [3], high blood pressure [4], and insulin intolerance [5]. Furthermore, low levels of activity are significantly associated with increased risk of complications of MetS, including diabetes and CVD [5,6]. In addition, research has found that losing weight by approximately 5% to 10% results in significantly reduced MetS-associated markers [1] in patients with existing disease, highlighting that MetS may be modifiable through lifestyle-related weight interventions. Dietary modifications, including reduced sodium, sugar, and fat intake, are also highly beneficial for reducing the risk of the syndrome and its complications [7].

In recent years, mobile health (mHealth) has increasingly been used to support behavior changes related to weight loss, including improving dietary intake and physical activity [8]. Research on the use of mHealth interventions has found support for a moderate effect size for assisting with weight loss [8]. This includes the use of SMS text messaging for behavior change and mHealth apps that target weight loss using a range of behavior change techniques (BCTs) [9], including self-monitoring, feedback, goal setting, education, tips, personal tailoring, reminders, encouragement, and social and professional support [8]. mHealth is a form of health care that enables timely accessibility, portability, and personalized medicine tailored to the needs of the user [10,11]. It includes smartphones, PDAs, MP3 players, iPads (Apple Inc), smart clothing, and smart watches [10,11].

Emerging research in the mHealth field has focused on developing conversational agents that can simulate human professional interactions for managing different health problems [12], including weight issues [13]. Furthermore, avatars have also been developed to display a conversational coach in addition to written conversational text, simulating real-life interactions with a professional, such as a live fitness coach [14,15]. Having a conversational coach complement or replace metabolic-related health advice from professionals may increase accessibility and enable more timely health monitoring and diagnosis of health conditions [15] such as MetS if physicians also gain access to patient data. Given that technology in the field is advancing, it is time to determine whether these conversational agents are effective for assisting with MetS-associated risk factors, including overweight, obesity, physical inactivity, and unhealthy dietary intake. There is also a need to better understand what types of weight-related and MetS-related studies have been undertaken using conversational agents and to identify challenges with the technology and future areas of research.

This review aimed to better understand the evidence surrounding the use of conversational coaches for metabolic-related risk factors and biomarkers. Furthermore, this review aimed to determine whether conversational coaches are effective for improving weight-related behaviors and metabolic indicators and whether conversational agents are acceptable for consumers as agents of behavior change.

This review included studies on chatbots or avatar conservational agents that acted as coaches for improving metabolic health behaviors, including dietary intake (sodium and sugar intake), PA, and weight (including abdominal obesity). Studies that evaluated one or more physiological indicators of metabolic health or risk factors for MetS, such as diabetes, glucose intolerance, hypertension, cholesterol, and serum triglycerides, were also included. Studies must have been published in the English language to be included. Chatbots that were used for survey reasons but not primarily for targeting weight-related or metabolic risk factors were excluded. Studies whose primary focus was not on conversational coaches were excluded (including those that had an avatar element but did not primarily focus on evaluating it). Studies on wearables that did not include avatars or chatbots were excluded. Studies in pregnant women were excluded.

The keywords included word variations for “chatbot,” “virtual assistant,” “virtual coach,” or “avatar”; weight-related behaviors, including “diet,” “exercise,” or “weight”; and metabolic risk factors, including “hypertension,” “cholesterol,” or “diabetes.” The search strategy is shown in Textbox 1.

Titles were screened for relevance to the RQs, followed by abstract and full-text retrieval of eligible studies that met the inclusion criteria. A second reviewer (LL) screened the abstracts and full texts against the inclusion and exclusion criteria to ensure agreement. Quantitative and qualitative data were extracted and summarized in a tabular format, including study characteristics, measures, outcomes, and intervention details.

LL and ME screened the final selected papers individually. A total of 52 full texts were selected [13,14,16-65]; however, after double peer screening, 1 protocol and 1 dated technology were removed. The final number included 50 papers [13,14,16-59,61-63,66]. Details of the search process and reasons for exclusion are illustrated in Figure 1 [67].

Most of the studies were feasibility and usability studies. A few studies were qualitative and explored consumer perspectives on conversational agents for weight-related behaviors [14,19]. The countries where the studies were conducted included Australia, the United States, Italy, Spain, and Taiwan [13,14,16-29]. Most of the studies explored virtual agents for diet and exercise, with only 2 (4%) exploring chatbots for hypertension management [17,19]. The majority were conducted among adults, but 3 (6%) were conducted among teenagers and preteens [14,26,29]. The study characteristics and results are summarized in Table 1.

A few studies evaluated the effects of conversational assistants for weight loss [13,22-24,44]. The study by Maher et al [13] in Australia found that the conversational assistant (chatbot) Paola assisted with a weight loss of 1.3 kg at 12 weeks follow-up (95% CI –0.1 to –2.5). In addition, there was a mean waist circumference reduction of 2.5 cm at follow-up compared with baseline (95% CI –3.5 to –0.7). The chatbot used a range of BCTs, including goal setting, self-monitoring, education, social support, and feedback to users on PA and the Mediterranean diet [13]. A study in the United States found that the Lark Weight Loss Coach, an artificial intelligence–powered bot, assisted participants with a weight loss of 2.38% (95% CI –3.75 to 1.0) with a mean use of 15 weeks [25]. The conversational agent was informed by cognitive behavioral therapy and used a range of BCTs, including education, encouragement, and reminders surrounding dietary and PA targets [25]. The determinants of weight loss included the duration of using the artificial intelligence program and engaging with it, logging meals, and the number of counseling sessions completed [25]. A large study in the United States examining the use of an avatar coach that targeted self-efficacy and modelled vicarious experiences for diet and PA (4 weeks) found that women lost an average of 1.6 (SD 1.7) kg at follow-up [22]. A study in India found that an avatar coaching app with calls from health professionals assisted with a weight loss of 1.39 kg (95% CI –0.63 to –2.01; P<.01) at 16 weeks [44]. A randomized controlled trial (RCT) with a qualitative component found that avatars increase motivation and PA self-efficacy linked with weight loss [33]. However, some studies did not report any significant weight loss [34,49].

A few studies evaluated the effects of conversational coaches (chatbots and avatars) on dietary intake and found that overall, the coaches assisted with ameliorating dietary habits and goals [13,16,25,28,34,45,49]. A study in the United States found that healthy dietary intake improved in 30% of participants who were using a conversational weight loss coach [25]. Another study found that eating behaviors improved in users of a conversational eating coach, which included increases in the mean scores for the perceptions of skills to eat healthily and self-control over their eating habits (0.7 increase in points) as well as confidence to control food consumption in social situations (1.0 increase in points; P<.01) [28]. The Paola chatbot study found a mean increase in the Mediterranean diet score [68] of 5.8 points at 12 weeks follow-up [13]. Similarly, a study of Karen, an animated counselor, found significant increases (F_3,103=4.5; P<.01) in fruit and vegetable intake in the diet intervention arm relative to the control group [34]. A further study found that eating behaviors were shaped by the appearance of the avatar, with healthier eating behavioral patterns in participants who had thinner avatars including reduced portions of ice cream and opting for healthier sugar-free drink alternatives [45].

A few conversational assistant PA coaches, including chatbots and avatars, were evaluated, and overall, they assisted with increasing PA [13,16,21,23,24,27,55,63]. Most of them involved exergames with the avatar. However, one of the studies did not find any improvements in PA among the 2 avatars, attributing improvements only to the web-based part of the intervention [23], and another study did not find a difference between the web-based intervention and the chatbot (only when considering a standard control) [63]. A preliminary usability study in Australia found that step count goals increased 59% of the time in users of the chatbot that targeted PA and that participants had a preference for personalization and greater knowledge-based content [16]. Another pilot study of Paola, the chatbot in Australia, found that it assisted with increasing mean step count by 109 minutes per week at 12 weeks follow-up (95% CI 1.9-217.7) [13]. A study involving an exergame that used a PA avatar coach in teens found that 75% of the time, participants engaged in 15.88 (SD 5.8) minutes of vigorous PA throughout the game [29]. Participants also wanted the avatar to have a supportive and nonpatronizing or nondisparaging tone in interactions regarding PA and found that it could motivate older adults when adequately personalized [55]. Similarly, a study in children also found that they desired the option to personalize the avatar, including controlling and customizing its physical appearance during game play when exercising [66].

The Proteus effect is a phenomenon wherein individuals embody and emulate the behaviors of their virtual characters such as avatars [69,70]. A few studies demonstrated support for the Proteus effect when it came to PA behaviors, although the type of avatar varied. A study in Taiwan found that younger looking avatars were associated with higher levels of PA than older looking avatars but only in women. Male participants had higher levels of PA than female participants who used an older looking avatar, highlighting differences between sexes [27]. A further study found a higher cardiac output resulting from increased intensity of PA in adult users of an avatar that resembled them and wore gym clothes when compared with avatars that appeared unfamiliar like strangers in regular clothing, which reduced heart rate [24]. Similarly, a study in Taiwan found increases in physical activity assessed in movements (986.7 points higher) in users of a “normal avatar”, more closely resembling them than the most muscular avatar [21]. They also found that self-efficacy was higher (0.66 points) for core muscle exercises in female participants assigned to normal avatars relative to their muscular counterparts and male participants assigned to the same standard avatar (0.9 points higher), with P<.05 [21]. Similarly, dietary behavior was also shaped by thinner embodied avatars in another study [45].

Most diabetes studies were feasibility studies. The results of diabetes conversational coaches were mixed. A few studies did not have positive findings concerning the applications with avatars for diabetes [42,53]. However, one study reported a usability score of 73, which is relatively high. Notably, the study integrated a range of BCTs, including goal setting, feedback, self-monitoring, social support, and counseling [32]. Low usability scores were reported in a few studies, including one that reported an overall score of 44.58 (SD 21.18) [42]. Similarly, an RCT of a diabetes coaching avatar did not find that knowledge increased relative to controls, but intervention participants in the computer-based programmed dynamic avatar had higher satisfaction levels (F₄=3.11; P=.01) [53]. Another study in the United States in participants with prediabetes found that 60% of patients had successfully completed the modules and met weight targets during 6 months of use (60% success rate) [57]. Engagement was also moderately high (50%) in a study in Singapore involving a chatbot, although usability was high along with retention (93%) [36]. In a study in teenagers, attrition was also low, and 80% of the participants were satisfied with the conversational diabetes coach [59]. A study that evaluated a coaching application for diabetes found an improvement of –11 mg/dL in fasting blood glucose levels [44]. However, the intervention also involved phone calls from health professionals [44]. Similarly, an avatar application with a ketogenic diet program assisted with a reduction in hemoglobin A_1c levels of 0.69% (SE 0.168%) [58]. Qualitative research found that avatars created an environment of social presence that facilitated social support and coherence for patients with diabetes [48]. In another study of avatars combined with robots, children preferred robots over avatars, but their friendship increased if the two had a greater similarity, which impacted usability [56].

A few studies evaluated the use of conversational coaches for CVD. One of them was a pilot study of the Tensiobot chatbot [17], a coaching application that teaches users how to properly check their blood pressure using recommended practice guidelines and provides users with graphic feedback and reminders. The study found that the chatbot group did not differ from the control group in terms of adhering to blood pressure measurement recommendations. However, there were significantly higher levels of knowledge (+6.53 points) with regard to checking blood pressure in the chatbot group than in the control group (P<.05) [17]. Blood pressure (diastolic) was significantly reduced, that is, by 1.43 mm Hg (SE 0.65; 95% CI –2.72 to –0.14; P<.01), in users of an avatar application that also involved a ketogenic diet [58]. In addition, a mixed methods study with a qualitative component found that users in general were interested in trying a hypertension chatbot for medication management as well as for health communication and self-care [19]. In addition to these studies, a general diet and PA chatbot study evaluated changes in blood pressure, but these changes were nonsignificant [13]. A study in Poland found high desirability for a CVD voice technology coach, in addition to accessing phone-based telemedical services by health professionals [43]. A further study in Brazil evaluated avatars for cardiovascular rehabilitation and found that an avatar with an exergame influenced heart rate, systolic blood pressure, and respiratory rate during the intervention and up to 5 minutes after its completion [31]. Furthermore, a study found that the avatar intervention increased the intent to improve lifestyle behavioral risk factors in patients relative to controls (P=.01), although confidence did not change [54]. Finally, a study evaluated a cardiovascular educational avatar application and found that it increased symptom recognition by 24% and knowledge of CVD by 15%, with a high satisfaction rate of 87% among patients [62].

Several studies found that users were interested in using conversational coaches for lifestyle behaviors [14,19,22]. Overall, participants enjoyed using the chatbots and avatars or found them helpful for diet, exercise, and hypertension management [17,22,25,26,29]. User-friendliness was reported by 83% of the participants in a study that evaluated a PA social media chatbot [20]. Similarly, 87.5% of women in a weight loss avatar intervention found it helpful [22]. With the exception of studies on diabetes conversational coaches, adherence or completion of tasks was high across studies on lifestyle (diet and PA) conversational coaches, ranging from 85% to 90% [13,16,17,29]. The qualitative study themes were related to the desirability for a conversational coach for hypertension and weight-related behaviors, especially for one that simulates human interaction closely, provides advice and goals for meals when cooking, and provides educational support [14] including for hypertension management [14,19].

A few tech challenges were brought up across the studies. Although users found that the conversational coach answered basic questions correctly, failure to understand and respond to more complex or spontaneous questions was reported in the studies. The percentage of failure for spontaneous or complex questions was 79% in one study [16], and participants in another study gave a high ranking for the chatbot’s failure to recognize their input [20]. Paola chatbot correctly answered spontaneous questions on diet in 4 out of 20 attempts, with a success of 20%, while the percentage of correctly answered simple and predetermined questions and responses was 96% and 97% [16].

This review aimed to better understand the effectiveness of virtual coaches for managing metabolic health and weight-related risk factors. It appears that virtual coaches hold potential for assisting patients with improving their dietary intake and PA behaviors, leading to subsequent weight loss. However, more studies that are larger and sufficiently powered RCTs are needed to establish a stronger evidence base. RCTs are the gold standard of evidence but are often costly and time-consuming [71,72]. Most of the studies were limited, as they were pilot studies. Ideally, it would be of interest to research long-term weight changes and cardiometabolic risk factor modifications over longer periods.

It appears that PA interventions may benefit from using avatars that embody the participant. The Proteus effect is based on the hypothesis that users adjust their behavior by modeling the virtual character with which they interact [73]. Thus, it seems that incorporating an avatar may enhance mHealth chatbot interventions, as it adds the element of user interaction and promotes the modeling of behavior through embodiment [73]. However, 1 (2%) study did not find that the avatars enhanced the effects of the web-based intervention [23].

We also found that consumers seemed to be interested in and enthusiastic about trying virtual coaches for managing their weight-related behaviors and blood pressure. Adherence to the intervention was also high throughout the studies, which indicates that this technology is acceptable and usable for patients. However, there is a need to undertake qualitative research on developing a MetS coach to further understand consumer perspectives. The main barrier to consider when developing future virtual agents is that the virtual agents did not always answer correctly to spontaneous responses. As consumers want personalized and tailored mHealth for weight-related behaviors [74], future applications should ensure that the virtual agents are sufficiently advanced to be able to interact with users in a natural and personalized manner.

It appears that diabetes virtual coaches should be improved to maximize engagement and adherence, as not all studies found that they were helpful. Although outside the scope of this review, we note that some studies used BCTs, which could suggest that future applications may benefit from integrating BCTs [13,20,22,25,32,42,51]. In addition, we identified some studies on blood pressure and CVD management, which demonstrated preliminary improvement in patients with hypertension as well as knowledge of CVD. However, we did not identify any virtual coaches for managing MetS. Therefore, there is a need to develop virtual coaches specifically tailored to this syndrome and its associated risk factors. Such virtual coaches could be integrated into a combined synchronized application that involves diabetes and CVD education and monitoring.

MetS is linked with high blood pressure, which is one of the main hallmarks of the disease. The theoretical mechanisms underpinning the development of hypertension in patients with MetS have included a combination of endothelial dysfunction, systemic inflammation, adiposity, and oxidative stress [75]. Dysfunction in the renin-angiotensin system has also been theorized to be a determinant [75]. Obesity itself has also been identified as a risk factor for high blood pressure in MetS [76]. Blood pressure is modifiable to some extent through lifestyle changes previously described, including dietary sodium restriction, PA, stress reduction [77], and medication [78]. Future virtual coaches may target hypertension as part of a MetS intervention, and this review found that patients are willing to try chatbots for managing their blood pressure.

MetS is also associated with high glucose levels of at least 100 mg/dL when patients are fasting [78], which indicates that they are in the prediabetes stage, as diabetes begins at fasting glucose levels of 126 mg/dL [79]. In a recent longitudinal study, patients who reduced their fasting blood glucose levels decreased their overall risk of diabetes by 54% when compared with their counterparts who did not improve their blood sugar levels (95% CIs exclude 1) [80]. A recent study found that individuals who consumed high amounts of sugar were 32% more likely to have MetS than their counterparts [81]. Thus, a future MetS virtual coach could target blood glucose monitoring and offer personalized advice on optimum sugar intake.

In addition to targeting dietary intake, PA is integral to managing this syndrome. A meta-analysis found that the risk of cardiovascular events was reduced by 30% in physically active individuals compared with those who were inactive [6]. A longitudinal study in middle-aged women found that increasing step counts significantly reduced, by 30%, the risk of MetS in this population and that they had clinically improved levels of the protective cholesterol high-density lipoprotein, whereas their serum triglycerides had significantly decreased [82]. A review found that walking on a daily basis reduced the risk of type 2 diabetes by nearly half [5]. Furthermore, recent research suggests that sedentary behavior, including sitting time, is an independent and significant risk factor for MetS syndrome [83]. Thus, PA chatbots and avatars, which were found to increase PA time, steps, and self-efficacy in this review, could be integrated into a comprehensive future MetS interventions.

Given that chatbots and avatars hold potential for increasing PA and reducing sedentary behavior, as well as improving dietary intake, studies are needed to evaluate their effectiveness for managing the symptoms and risk factors associated with MetS specifically.

In addition, stress is often an underlying determinant of maladaptive weight-related behaviors, including binge eating, emotional eating, and an unhealthy dietary intake as well as weight gain [84-88]. Future avatar and chatbot interventions for cardiometabolic factors could also consider integrating psychological supportive interventions such as mindfulness-based stress reduction, which assists with weight and stress [89-93], as an element.

In summary, we found that virtual coaches hold promise for regulating diet, PA, weight, and possibly hypertension. However, studies on virtual coaches are few in number; therefore, more research, including RCTs, is needed to confirm the effectiveness of virtual coaches. Overall, most participants in the reviewed studies were interested in using virtual coaches, including chatbots and avatars, for regulating their weight-related behaviors, and study adherence was good. Future interventions could be ameliorated to reduce technical challenges associated with these conversational agents and ensure that they respond correctly to complex and spontaneous questions. Furthermore, future research could involve developing a comprehensive conversational agent for MetS, such as a health coach that simultaneously targets diet (sodium, sugar, and fat intake), exercise, weight (including abdominal obesity), blood pressure, and diabetes, and evaluating it. This would include a health coach that simultaneously targets diet (sodium, sugar, and fat intake), exercise, weight (including abdominal obesity), blood pressure, and diabetes.