Evaluating the Potential and Pitfalls of AI-Powered Conversational Agents as Humanlike Virtual Health Carers in the Remote Management of Noncommunicable Diseases: Scoping Review

Introduction

Burden of Noncommunicable Diseases

Noncommunicable diseases (NCDs), also known as chronic diseases, are medical conditions that are not primarily caused by infectious agents (eg, viruses, bacteria, fungi, or parasites) and cannot be transmitted from one individual to another through close contact [1]. NCDs, such as cancer, cardiovascular diseases (CVDs), chronic obstructive pulmonary diseases, chronic respiratory diseases, chronic kidney diseases, cognitive disorders, metabolic syndrome, diabetes, and hypertension, are rising worldwide, with significantly higher rates in low- and middle-income countries (LMICs) [2-4]. In 2019, NCDs contributed to the highest proportion of total mortality (74.4%), accounting for 7.1 million additional deaths in 2019 as compared to 2009 [5]. Nearly half of the mortalities in Asia are attributable to NCDs, causing >80% of CVD and diabetes deaths, 90% of chronic obstructive pulmonary disease deaths, and two-thirds of all cancer deaths occurring in LMICs, resulting in 47% of the global burden of disease [3,6]. Major risk factors of NCDs include unhealthy dietary habits; physical inactivity; stress; and consumption of drugs, tobacco, and alcohol, which are generally modifiable due to lifestyle choices [5,7]. These chronic diseases are also major causes of long-term disabilities and prolonged costly treatment that may pose serious threats to a country’s health care resources and expenditure, especially among lower-income countries where health systems are not sufficiently equipped to tackle the escalating challenges [3,4,8].

In addition to regular visits to health care centers, NCDs require challenging self-care management, including compliance with medications, lifestyle modifications, and constant symptom monitoring to prevent disease progression; nonetheless, adherence to these procedures is generally low, especially among adults with limited health literacy who struggle to comprehend and follow instructions from their health care professionals [9,10]. Moreover, the shortage of health care providers and their limited time availability are substantial causes of patients’ being deprived of receiving adequate health education and support to make informed decisions required for the effective self-management of their chronic illnesses [11,12]. In addition, access to proper health services can be limited in many underdeveloped areas and rural communities due to poor health care infrastructure and mobility facilities [13].

The optimal prevention management strategy should incorporate elements of individual lifestyle management, societal health awareness management, national health policy decisions, and global health strategy [7]. However, the escalating prevalence of NCDs worldwide reveals that traditional disease management techniques are not sufficiently effective, thereby indicating an urgent necessity to develop effective supplementary management strategies to mitigate the substantial financial burden imposed by NCDs on many households, particularly in LMICs [2,3,14].

The Role of Telehealth in NCD Management

Telehealth applications have the potential to improve patient self-care and disease-specific knowledge as well as minimize hospitalizations and mortality [15]. There is evidence from several past studies suggesting the significant effectiveness of mobile-based telehealth apps in improving nutritional intake and physical activity with technology intervention, resulting in body weight loss and adoption of recommended lifestyle changes due to their convenience and accessibility [16]. This is because such apps, with the help of existing and emerging technologies, have the ability to assist patients in managing their chronic diseases more effectively by providing constant self-monitoring tools and promoting improved self-management of health problems [17]. Furthermore, the ever-growing ownership of mobile devices worldwide has greatly contributed to the shift toward digital health care services, including assessment, monitoring, and treatment of physical and mental health, thereby indicating the promising ability of mobile health apps in self-monitoring, assessment, and treatment of NCDs at a reduced cost [18-20]. Indeed, the COVID-19 pandemic has expedited the use of telehealth [21,22], which has been proposed as a cost-effective method of delivering better health care services to people with chronic illnesses in a more flexible, personalized, transparent, dynamic, and accessible way [21,23].

Potential Enhancement of Existing Telehealth Apps With Artificial Intelligence

Although mobile-based telehealth apps offer an ideal platform for systems designed to help patients manage their chronic illnesses due to the smartphones’ computational power, connectivity, and consistent availability, many individuals struggle with complex user interfaces of existing digital health technologies [10]. While the ubiquity of these apps reduces some acceptance barriers, most apps still overlook many other barriers, such as lack of motivational, psychological, and emotional support [10,24]. Nonetheless, technological advancement involving artificial intelligence (AI) seems to have the potential to further upgrade existing mobile health apps with more user-friendly features that can support individual user needs [25]. For instance, many patients lack the ability to effectively navigate conventional telehealth apps due to limited health-related or computer literacy and disabilities such as visual impairment; hence, such intelligent dialogue systems, specifically termed conversational agents (CAs), may help overcome these limitations and improve usability by providing an oral presentation of the apps’ contents in plain language [15]. AI-powered CAs are computer systems that can communicate with humans through text, voice, and images on mobile, web-based, or audio-based platforms using AI techniques such as machine learning (ML; a statistical method of training models using data for making predictions based on a variety of features) and natural language processing (machines’ ability to detect and interpret humans’ verbal and written languages) [26,27]. A CA may serve as an empathetic listener to understand patients’ problems as well as aid in monitoring a patient’s health 24/7 and notify physicians about an anticipated medical emergency [13].

The popularity of CAs, especially those that use unrestrained natural language, has increased over the last decade as consumers can use their smartphones to interact with CAs for daily tasks [27]. When deployed on mobile devices, CAs have the potential to augment human intelligence and demonstrate multiple benefits such as delivering health education and behavior change for a range of chronic health conditions [10]. Moreover, these AI-based CAs provide additional communication channels that are particularly effective for developing trust and therapeutic alliance to encourage adherence among users [10]. The human-like conversational features of CAs due to advancements in natural language processing, voice recognition, and AI are increasingly substituting human employees in service encounters, including the health care industry to deliver personalized care and support to individuals with chronic health issues [28].

Types of CAs

In this paper, we broadly categorize intelligent CAs into 3 main types (Table 1): chatbots (text based or voice enabled—voicebots) without embodiment, computer-based embodied digital avatars, and physically embodied humanoid robots [29]. Although robots are beyond the scope of this paper, AI humanoid or social robots can also be classified as CAs that can be used as human-like health caregivers for managing NCDs. Despite the differences, all CAs, including humanoid robots, aim to enhance relational outcomes through human-like communication [29].

Ethical Concerns of AI Agents

While AI agents have the capability to provide constant health surveillance support, challenges and risks associated with using AI-based CAs in health care remain. These include ethical concerns regarding data collection and interpretability of results, patient safety risks, biases encoded in algorithms, and cybersecurity [55,56]. Furthermore, customers are often reluctant to engage with such AI-based CAs due to many factors such as trust, reliability, learning curve, usability, privacy, and data security that should be addressed in the design, deployment, and use of AI applications [12,29,57-59].

State-of-the-Art Summary

Our study is novel regarding CAs as human-like digital agents for managing NCDs remotely. As AI-based CAs are a relatively new area, limited research has been conducted on applying these emerging technologies in health care. While existing research on CAs as virtual caregivers for NCDs serves as a seminal foundation, such research has other critical limitations. To illustrate, most studies have primarily focused on applications of CAs for mental health conditions, overlooking broader NCDs such as CVDs, metabolic syndrome, or diabetes [60,61]. A recent review [61] that focused on a different set of research questions from ours concluded the following: “A future chatbot could be tailored to metabolic syndrome specifically, targeting all the areas covered in the literature, which would be novel.” Our research looks specifically into this gap.

Furthermore, despite some favorable anecdotal evidence, the effectiveness of CAs in NCD management is seldom explored in large-scale trials, particularly in older adults, who have the highest risk of developing NCDs [62-64]. Some recent previous studies have conducted systematic reviews [27,65,66] or scoping reviews [61,62,67] on CAs in chronic disease management; however, to the best of our knowledge, no review has been conducted yet on the application of AI-based CAs as human-like digital agents for managing NCDs remotely. Such limitations and gaps as highlighted previously present both challenges and opportunities for research to advance our understanding of how CAs can contribute to managing NCDs.

Aim

This scoping review aimed to provide an overview of the existing evidence and research on using assistive humanoid AI-based CAs in health and social care for managing NCDs. Our primary objective was to explore the impact of AI-based CAs, including embodied avatars, as human-like health carers for the self-management of chronic diseases. By examining the current literature on this topic, we hoped to identify key areas for future research and provide insights into how these technologies can be effectively used in health care and personalize NCD management strategies.

Research Objectives

Our research objectives were as follows: (1) to explore the current state of research on the use of AI-based CAs as human-like virtual health carers for managing NCDs, (2) to identify the potential benefits and challenges associated with the use of these technologies in the health care field, (3) to explore the efficacy of AI-based CAs in the remote management of NCDs, (4) to discover the specific target users primarily studied, and (5) to provide recommendations for future research in this area.

Research Questions

Our research questions were as follows: (1) what is the current state of research on using AI-based CAs as human-like health carers for managing NCDs? (2) What are the limitations or challenges associated with the use of these technologies in health care? (3) What are the potential benefits of using AI-based CAs in managing NCDs, and how can they be effectively used to improve health care delivery and reduce health care burden? (4) What is the efficacy of the CAs in the remote care of NCDs? (5) What are the frequently targeted user groups for such virtual agents (eg, specific age groups and individuals with special needs)?

Methods

Search Strategy

We followed the methodological frameworks proposed by Arksey and O’Malley [68] and Levac et al [69]. Initially, research objectives and questions were formulated, followed by a systematic literature search conducted on July 31, 2023. For primary searching, we used 6 electronic databases that are considered relevant to the research focus—Ovid MEDLINE, Embase, PsycINFO, PubMed, CINAHL, and Web of Science—applying the same set of keywords, such as “conversational agents,” “artificial intelligence,” and “noncommunicable diseases,” including their associated synonyms (as shown in Textbox 1). The Boolean operators “*” and “OR” were used to expand and ensure that different word combinations were included. The operator “AND” was used for combining the main search terms to identify articles focusing on AI-based CAs (as health carers) only applicable in the health care field, particularly for managing NCDs or chronic diseases.

Additional studies were identified by hand searching the reference lists of included studies and relevant review articles. Furthermore, a supplementary manual search was conducted to identify specific articles from diverse sources, including ProQuest Central, ResearchGate, ACM Digital Library, and Google Scholar. Some of these articles were identified through pilot hand searching and preselected as they were closely relevant but not retrieved from the selected databases used for primary searching, whereas others were discovered through manual searches targeting specific authors well known in this field. The search strategy underwent refinement following expert recommendations, including insights from a coauthor with digital health expertise, who advised excluding “robots” from the search terms, recognizing it as a distinct field. These methodological steps collectively ensured a comprehensive exploration of the relevant literature.

Eligibility Criteria

Our scoping review used comprehensive inclusion and exclusion criteria to ensure the selection of pertinent studies. We did not impose any limitations based on gender or age groups in the selection of articles. The search scope was confined to scholarly articles in English published between January 2010 and July 2023, aligning with the substantial rise in CAs after 2010 [70], notably with the introduction of Apple’s Siri in 2011 [71]. Evidently, most of the selected papers were on recent studies conducted within the last 5 years due to the accelerated technological advancements and the latest evolution of AI. Specifically, we focused on empirical studies exploring CAs applied exclusively to human interaction within the health care context, emphasizing their role in the remote management of NCDs, ideally within home environments.

The exclusion criteria comprised the exclusion of conference abstracts, posters, reviews, protocols, position papers or viewpoints, and certain types of studies such as those involving noninteractive robotic devices. We also excluded studies related to CAs designed for medical education and non–patient-centered applications in hospital settings and those addressing specific health care domains such as surgery, dentistry, pregnancy or maternity, addiction or substance use disorders, and communicable diseases. Furthermore, we excluded studies centered solely on medical history data storage, telephone monitoring, data set construction methods, or user evaluations of commercial CAs for health care without their practical applications in the remote management of NCDs. Our criteria aimed to streamline the focus on patient-centered CAs contributing to the self-management and remote monitoring of NCDs, eliminating studies with a primary emphasis on clinical interviews, disease prediction, or decision support without a social interaction element.

Screening and Selection

In total, 2 of the authors independently searched each database. Titles and abstracts were screened for inclusion according to the aforementioned criteria, followed by an exclusion of duplicates, unrelated studies, and articles that could not be retrieved. The abstract screening yielded 264 articles eligible for full-text screening, of which 70 (26.5%) were review papers comparing different types of CAs in the health care field and 156 (59.1%) were empirical studies that explored different types of CAs used in the prevention, treatment, or rehabilitation of chronic diseases involving consumers, caregivers, and health care professionals.

Subsequently, the authors screened full texts of the remaining articles independently, and 40 full-text articles that met the inclusion criteria were selected for review. In addition, 10 [10] specific articles were obtained from hand searches of other sources (Google Scholar, ProQuest, ACM Digital Library, and ResearchGate), of which 3 (30%) [3] relevant ones were selected upon full-text screening. It was an iterative process; any discrepancies were discussed among the authors. The search and selection process is illustrated in Figure 1.

‎

Data Collation and Reporting

An Excel (Microsoft Corp) spreadsheet was initially created to aid the screening and selection process. Following the screening process, a total of 43 articles were eventually selected for synthesis. Quantitative and qualitative data from the included studies were extracted and summarized in a tabular format, including information such as intervention, type of CA, target population, number of participants with their age, methods, study duration, location, measures and outcomes, and limitations. The relevant extracted information was collated and summarized using the narrative synthesis approach, which was deemed appropriate for capturing the breadth of evidence in scoping reviews, identifying themes aligned with the research questions as well as patterns observed across the included studies and relevant reviews. The results were reported according to the PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines. Given the divergent methodologies of the included studies, predominantly using mixed methods and qualitative approaches, no quantitative synthesis or meta-analysis was conducted.

Results

Characteristics of the Included Studies

Most of the included studies (22/43, 51%) were feasibility, acceptability, and usability studies (Table 2). These studies mostly found positive results in terms of feasibility and usability of the proposed CAs. Some studies (4/43, 9%) used the System Usability Scale questionnaire to measure the system’s usability score and predominantly found high usability, with System Usability Scale scores of >70 [72-75]. Alternatively, multiple studies (10/43, 23%) used the subjective ratings of most of the participants to evaluate the feasibility and acceptability of CAs by applying indicators such as perceived usefulness [76-79], ease of use [78,80], user satisfaction [80,81], engagement rate [80,82,83], perceived closeness [73], and Net Promoter Score [84]. However, 2% (1/43) of the studies reported negative usability and acceptability outcomes, indicating unsatisfactory reliability, usability, and goal structure, which hindered health care professionals’ acceptance and trust [85].

A few studies (6/43, 14%) focused on evaluating the effectiveness of the CAs used for the self-management of certain chronic illnesses, such as atrial fibrillation [10], type 2 diabetes mellitus [24,82], and mental health conditions such as depression and anxiety [90,96,97]. All these studies found improvement in various parameters in management, such as significant reduction in glycated hemoglobin level (P<.001), as well as BMI, and weight (P<.001) in patients with type 2 diabetes mellitus [82]; notable improvement in quality of life measured using the Atrial Fibrillation Effect on Quality of Life questionnaire score (P<.05) or health-related quality of life utility score (P=.007) in patients with atrial fibrillation [10,24]; and substantial reduction in depression (9-item Patient Health Questionnaire score) and anxiety (7-item Generalized Anxiety Disorder Scale score) symptoms [90,97,105].

Types of AI-Based CAs in the Included Studies

Of the 47 AI-based CAs representing digital health assistants identified from the included studies, 9 (19%) were existing commercial CAs available as marketed products, which included voice assistants—Google Assistant, Amazon Alexa, Microsoft Cortana, and Apple’s Siri, as well as social chatbots—ChatGPT and Replika (a voice-enabled, avatar-based chatbot), and health care chatbots—Vik, Wysa, and Tess. The remaining 38 CAs were proposed designs or prototypes, where 19 (50%) of them were chatbots, 2 (5%) were voice assistants or voicebots [58,102], 13 (34%) were human-like digital avatars, 1 (3%) was a digital avatar version of a humanoid social robot [85], 1 (3%) was a text-based voice-enabled bot [74], and 2 (5%) were avatar-based chatbots [91,110]. In addition, 5% (2/38) were CAs as humanoid social robots [79,85], which were included as they particularly emphasized managing NCDs such as diabetes (although the robots topic was beyond the scope of this study).

Intelligence Level of the AI-Based CAs

Of the 38 CA prototypes presented, 27 (71%) were found to have indicated their AI algorithm implementation methods to generate appropriate health-related responses. A total of 30% (8/27) of them were implemented using rule-based algorithms [72,81,83,84,86-88,103], 7% (2/27) of the CAs were developed using network-based scripted algorithms [10,80], and 4% (1/27) of the CAs were implemented using prescripted algorithms based on behavior change theory, whereas 33% (9/27) of the CAs were implemented using ML approaches [58,74,75,89,93-95,97,99] and 15% (4/27) of the CAs were developed using a combination of rule-based and ML techniques [78,85,101,107]. However, one of the CAs identified from a study was a Wizard-of-Oz human-like AI avatar with prescripted text and human-controlled responses to evaluate the potential effectiveness of visually embodied avatars in the assessment of anxiety and depression through users’ behavioral expressiveness [96].

Types of Health Interventions and Study Duration

Most included studies (24/43, 56%) focused on behavioral or lifestyle interventions and patient education for improved self-management of physical NCDs such as CVDs [10,80,99,100], diabetes [24,72,74,79,82,85,89,93] or prediabetes [76], cancer [81,92], gastrointestinal diseases [104,106], and obesity [76,86,108].

A reasonable number of studies (13/43, 30%) focused on cognitive behavioral interventions related to the self-management of chronic cognitive impairments (eg, dementia and poststroke) [73,107], including chronic mental health conditions such as anxiety, depression, psychological distress, and bipolar disorders [83,87,88,90,96,97,101,102].

The duration of these studies was reasonably divergent, ranging from 5 days to 1 year. However, a substantial number of studies (14/43, 33%) could be considered long term, such as 8 weeks to 2 months [85,88,90,99,101], 12 weeks to 90 days [86,98,100], 16 weeks to 6 months [76,82,87,93], 9 months [94], and 12 months or 1 year [24,92]. A total of 26% (11/43) of the studies conducted trials for approximately ≤1 month [10,73,81,83,84,88,97,107] (ie, within 5-10 days) [72,78,79]. A total of 42% (18/43) of the studies did not mention the study duration.

Study Population and Location

Almost half (20/43, 47%) of the included studies did not specify the target population other than patients with a particular chronic disease that the health intervention was meant for. The studies that specified the target population or their age group (24/43, 56%) were relatively heterogeneous, with older adults [10,58,72,73,91,95] and middle-aged adults [24,99] being the most frequently targeted user group (8/43, 19%). However, some studies (6/43, 14%) also focused on younger adults [97], adolescents [76,108,110], and children [79,85].

The locations of the studies conducted were largely diverse, but most (27/43, 63%) were conducted in higher-income countries. A total of 9 studies were completed in Europe, where 2 (22%) of them were conducted in the United Kingdom [84,91], 2 (22%) studies were conducted in Italy and the Netherlands [73,85], 1 (11%) was conducted in Portugal [72], 1 (11%) was conducted in France [77], 1 (11%) was conducted in Greece [93], 1 (11%) was conducted in Spain [94], and 1 (11%) was conducted in Ukraine [97]. A total of 16% (7/43) of the studies were conducted in the United States [10,37,76,86,87,108,109]. In total, 4 of the studies were carried out in the Australasia region—3 (75%) in Australia [24,99,101] and 1 (25%) in New Zealand [110]. A total of 7 studies were conducted in Asia, where 4 (57%) of them were conducted in Japan [81,88,96,98], 1 (14%) was conducted in Singapore [74], 1 (14%) was conducted in Taiwan [78], and 1 (14%) was conducted in India [82]. A total of 5% (2/43) of the studies were conducted in Saudi Arabia [79] and Senegal in West Africa [75]. The remaining 33% (14/43) of the studies did not specify the study location.

Data Collection Methods of the Included Studies

The included studies mainly used quantitative (17/43, 40%) or mixed methods (17/43, 40%) approaches. A substantial proportion of these studies (11/43, 26%) were pretest-posttest intervention studies to observe the differences before and after the intervention within the same group of participants (single armed) or between intervention and control groups (2 armed). The study designs of the quantitative and mixed methods studies included observational experiments using 2-armed randomized controlled trials (7/34, 21%) [10,24,86,87,98,99,102], a nonrandomized prospective study (1/34, 3%) [88], retrospective studies (2/34, 6%) [82,105], quasi-experimental studies (4/34, 12%) [10,90,101,109], a single-armed pilot trial [94], and comparative evaluation studies (3/34, 9%) [89,96,100].

The qualitative studies (9/43, 21%) were mostly usability and acceptability assessments of prototypes through qualitative posttest surveys [103], postintervention interviews [107], focus groups [110], a participatory design approach with focus groups and interviews [108], and workshops [91]. A total of 22% (2/9) of the qualitative studies compared the proposed prototypes against similar existing applications [58,89].

Discussion

Principal Findings

The discussion of this scoping review synthesizes the findings in light of the overarching themes emergent from each research question, elucidating the interplay between the current state, challenges, benefits, and the targeted users of AI-based CAs within the scope of the examined literature.

Current State of AI-Powered CAs

Regarding the current state of research on AI-powered CAs as human-like health carers for managing NCDs, empirical studies have revealed the limited ability of existing commercial CAs to provide appropriate lifestyle advice and health recommendations [37,111]. Perhaps this is due to the lack of specified data training for these commercial tools in the medical context, which is currently restricting their integration in clinical and remote care [106].

Functionality-wise, commercial chatbots (eg, Wysa and Replika) tend to slightly outperform voice agents (eg, Google Assistant, Siri, Alexa, and Cortana) with improved response quality due to the former’s better dialogue capabilities and empathy support. At the same time, the latter are more susceptible to speech recognition errors, including greater dependency on simple web searches to generate advice [37]. Nonetheless, voice-enabled CAs offer hands-free communication, making them preferable over text-based chatbots for older adult users, as suggested by a recent scoping review conducted by Even et al [70] focusing on the benefits and challenges of CAs in the older adult population (Table 1). In addition, another study revealed that older adults preferred voice-based communications over SMS text messages or free-text entry as voice is regarded as a powerful mode for promoting motivation [112]. In addition, voice agents may also be more accessible compared to digital avatars due to their lower bandwidth requirements, making them suitable for locations with lower bandwidth connectivity or devices with limited computing power [48]. Conversely, digital avatars are generally more likely to be favored by users over other types of CAs for their widely accepted face-to-face conversation format [10] as they can support both text and audio outputs [47,113] and, therefore, can serve as a valuable alternative complement for remote care assistance [114]. This suggests that digital avatars remain a popular choice for their versatility and interactive capabilities, driving the rapid progression toward the anthropomorphism era as they represent realistic human-like traits in behaviors and appearances for enhanced interaction with the users [115].

Moreover, anthropomorphic CAs are more preferred over nonanthropomorphic CAs, as shown by Park et al [109], indicating that the human-like representations of a chatbot yielded a higher likelihood of intervention compliance than the machinelike chatbot because the former increased closeness and trust. Similarly, another study demonstrated that an augmented human-like embodiment of a digital avatar resulted in greater social closeness among the participants [116]. Correspondingly, strong social presence and emotional closeness are regarded as key facilitators of favorable attitudinal and behavioral outcomes among users [117]. Hence, anthropomorphism positively correlates to users’ perceived trust and acceptance of a virtual agent [12,117]. Indeed, digital human faces as intelligent CAs are considerably promising for achieving a human-like presence to enhance human-machine interaction [42]. Consequently, the application of AI-based conversational avatars as digital human-like health carers has been gaining attention recently, leading patients to prefer digital human avatars as medical professionals over chatbots due to the former’s perceived human-like and interactive nature, which is essential for maintaining the professionalism and emotional aspects in physician-patient relationships [39].

Nonetheless, this evolution may apparently raise perceived uncanniness, which is considered a genuine ethical concern with realistic anthropomorphic avatars, yet it does not appear to be a universal problem for the avatar appearance or limit users’ engagement with the virtual environment [55]. Realistic humanoid avatars representing virtual carers can appear more trustworthy and increase compliance among users [55,109]. However, striking a balance between realistic and stylized avatar appearances can be worthwhile to avoid visual overloading and discomfort among users while representing the required elements accurately to maximize perceived trust, closeness, and acceptability [25].

Challenges and Limitations of CAs

The following challenges with CAs, including limitations in intelligence level, empathy, data privacy, and ensuring patient safety, highlight significant barriers to their adoption and effectiveness in health care.

Intelligence Level of CAs

Most of the proposed AI-based CAs were rule-based or scripted systems with a predetermined set of rules and conditions, which are easier to design and implement but have limited capability to handle complex conversations (unlike the automated ML-based systems) [33,56,118] and, therefore, cannot generate personalized responses in a more human-like manner, which the users would mostly prefer [84,91,108-110]. Furthermore, few studies proposed integrating personality-tailored behavior in CAs, precisely known as personality-adaptive CAs, which can automatically recognize and adapt to the users’ distinctive personality traits accordingly to be able to support patients in a personalized and authentic human-like manner [119,120] as well as effectively persuade users to alter their attitudes and behaviors for healthy lifestyle habits [121].

Empathy of CAs

Studies have emphasized the importance of including the “empathy” element in the design of CAs [122,123], yet only one of the included studies notably aimed to incorporate “empathy” in the proposed chatbot by applying text-based emotion analysis to recognize the user’s emotion and respond accordingly using emojis, stickers, and Graphics Interchange Format images [78]. Furthermore, a scoping review on AI-based CAs conducted by Kusal et al [56] has suggested that the major shortcoming of CAs is their lack of competence to make communication with humans seem natural with the aid of empathy and sentiments, mainly due to their failure to comprehend the human context and the users’ emotions as their responses are restricted to the queries they are trained for, which can lead to a frustrating user experience. Indeed, a study conducted by Casas et al [124] determined the capability of an empathetic chatbot to surpass a benchmark bot or even humans in predicting the emotional state of textual messages. Therefore, CAs should be able to model their context and select important information to remember as communication depends on past messages [56].

Data Privacy

Surprisingly, the included studies demonstrated limited emphasis on the data privacy of the AI-based CAs while handling patient data, which is a common challenge faced in the design and evaluation of CAs [111,125]. Hence, it is recommended to prioritize data privacy and confidentiality factors during the development of such technologies to gain users’ trust and acceptance [63].

Patient Safety

Another area that needs to be focused on is patient safety [126], which was also not highlighted in the included studies. For instance, the inability of digital health tools to recognize the urgency of the patient’s health condition may generate inappropriate health advice or lead to delayed treatment, increasing the risk of further complications. In addition, limited validation and medical training for existing CAs (as mentioned previously) poses risk of inaccuracy and raises concerns about patient safety. Therefore, assessing the validation status of these digital health tools is crucial for patient safety and optimal health care outcomes.

Uses and Benefits of AI-Based CAs in Managing NCDs

This review suggests that CAs are primarily used for supporting nonpharmacological interventions such as behavioral or lifestyle medication for managing NCDs such as CVDs, diabetes, obesity, cancer, and gastrointestinal diseases, as well as preventive measures for healthy living, indicating their potential ability to complementarily assist in the long-term remote management of NCDs [114,127,128]. The ubiquitous self-care support through constant symptom monitoring, along with the personalized behavioral modifications provided by CAs, can possibly rescue patients from minor health concerns and prevent the progression of chronic diseases [10,104]. Health care professionals are optimistic about CAs’ potential ability to save patients time and money that they usually spend on physical clinic visits [103], thereby reducing the health care cost and burden and increasing accessibility [91,104]. Augmented reality applications could enhance the training of health care professionals and even patients in using these new technologies effectively [115,129,130].

Efficacy

The efficacy evaluation of CAs for remote care is still not thoroughly explored and well understood. According to Monaco et al [63], inadequate evidence regarding the efficacy of digital health tools may have hindered the potential beneficial outcomes of using such resources to control NCDs. Nonetheless, the few included studies that aimed to identify the effectiveness of CAs in managing the target NCDs found favorable outcomes [10,24,82,90,96,97,105]. Moreover, many reviews have emphasized the efficacy of CAs in health care support. For instance, a systematic review conducted by Milne-Ives et al [26] investigated the effectiveness and usability of AI-based CAs designed for health care and found them generally effective, with positive or mixed evidence, whereas a narrative review by Dingler et al [60] used context-aware voice assistants (eg, Google Home and Amazon Echo), demonstrating the utility of speech-based CAs in providing personalized remote health care support. However, a systematic review conducted by Hossain et al [131] exhibited low positive outcomes of digital health interventions among people with NCDs in India, requiring further exploration of advanced technologies for the equitable and sustainable development of digital health tools.

Targeted Users

Most of the target user groups were based on chronic disease conditions rather than any specific age group. A fair proportion of studies (6/43, 14%) focused on the application of CAs in the older adult population, exhibiting that this vulnerable user group may be most in need of such technologies as older adults have greater risk of developing serious NCDs [28,58,61] and many older adult patients lead an isolated, lonely life and, therefore, require constant remote monitoring as well as a virtual companion for healthy living [132,133]. Hence, in other words, the selected user groups were mostly based on targeted chronic diseases that may be more prevalent in older adults. However, we did not encounter any significant study that modified the CAs to enhance accessibility for older adults, but social networking sites could potentially enhance engagement with such CAs among this user group [134].

Moreover, it is crucial to recognize that the vast majority (41/43, 95%) of the included studies were conducted in high-income countries, underscoring the necessity of further research among underserved populations, particularly in LMICs where individuals may be more vulnerable to developing NCDs due to limited health awareness and access to health care services.

Limitations

Although this is an emerging field of research and we conducted an extensive literature search strategy, the possibility of noninclusion of relevant studies may still exist (eg, missing certain keywords or search terms, studies in languages other than English, studies not listed in the searched databases, and unidentified studies from gray literature). Similarly, our stringent inclusion and exclusion criteria were implemented to refine our focus on patient-centered AI-based CAs tailored for the remote management of NCDs, aiming to ensure the relevancy and applicability of our review to this particular domain of interest. However, this approach may have inadvertently narrowed the scope, potentially overlooking other relevant studies that could provide wider insights into our findings. Despite this, we strived to conduct a fair selection of studies meeting our inclusion criteria to mitigate selection bias and ensure a balanced representation of the literature. Nonetheless, it is important to acknowledge that other potential biases, such as publication bias, may still exist, which are beyond our control. In addition, new articles may have appeared after our search deadline (July 31, 2023), which may cause additional differentiations.

Furthermore, some common limitations were also identified within the included studies that may impact the overall findings of this review: (1) limited sample sizes (a vast majority of the studies had relatively small sample sizes, which could limit the generalizability of the findings), (2) short-term interventions or lack of long-term outcome measures (the duration of interventions in most studies was relatively short [80,82,88,101], potentially overlooking long-term user interactions and health outcomes essential for a comprehensive evaluation of CAs’ ability to sustainably assist in NCD management [10,37,77,85,91-93,96-98,103,109]), (3) limited demographic data (some studies lacked detailed demographic information of the participants [37,93,109], which may potentially limit the analysis of how demographic factors can impact the feasibility or effectiveness of the CAs), (4) potential bias (many studies had potential biases resulting from self-reported user data [37,77,78,101,109], sample bias [75,82,105], coding bias [108], and exclusion of variables or unaccounted confounding factors in analyses [100], which could impact the validity of the results), and (5) unaddressed technical aspects (some studies did not discuss the technical aspects of the proposed CAs, including their potential technical limitations or challenges [85,92,93]). Addressing these issues could contribute to the enhancement of the design, development, and usability of CAs in the future [135].

Implications and Future Recommendations

As CAs strive to facilitate personalized and empathetic human-like conversations in the health care field [113], the necessity for context-oriented and domain-specific training emerges as a crucial consideration for improved accuracy and relevance [136]. Although advanced AI systems such as large language models trained on vast textual data (eg, the generative pretrained transformer series and LLaMA) have proven transformative in various sectors such as marketing, education, and customer service, their application in the health care sector remains underexplored, primarily due to a scarcity of relevant high-quality medical data sets [137]. In enhancing the AI-based CAs for health care, only a few recent studies have aimed to investigate real-world interactions between patients and health care professionals [136-139]. While recognizing the dynamic and nonstandardized nature of historical human interactions in health care, our recommendation to study real-world communications between patients and health caregivers stems from the objective of training CAs to be contextually relevant, particularly in addressing the unique challenges associated with specific chronic diseases. This approach may seek to empower CAs to navigate the complexities of diverse health conditions with greater intelligence and empathy, thereby improving their adaptability through the use of high-quality data sets acquired from real-world scenarios [136] given the substantial amount of data and extensive knowledge base required for domain-specific contextual training [56]. However, to ensure effective training of AI agents in health care, it is essential to define their roles (eg, nurse, health assistant, nutritionist, and pharmacist), which can determine the appropriate level of training required, allowing for the selection of specific data sets tailored to the medical training and evaluation of CAs.

Findings from the literature suggest an overall higher user preference for realistic human-like representations in CAs, specifically in the form of anthropomorphic digital avatars, which can be accessed through mobile devices (unlike physical humanoid robots) [52,54], thus accommodating a broad user base including older adult users [58,140]. Therefore, future research should prioritize investigating the impact of anthropomorphic avatars or virtual humans in managing NCDs, including their potential ethical concerns (eg, uncanny valley) that should be addressed in the design considerations to increase acceptance of CAs for self-care [25,49,123].

Moreover, user preferences related to CA types and features may vary notably among different user groups. For instance, older adults and users with disabilities may require voice-activated CAs over text-based CAs due to their typing inabilities, and therefore, enhanced speech recognition functionality is required to incorporate into CAs to accommodate voice conversations in multiple languages [70]. In addition to language requirements, factors such as culture, ethnicity, and individual personality traits may also play a significant role in influencing the adoption of AI dialogue systems for remote care purposes. Indeed, the adoption of CAs can be impeded among populations traditionally experiencing health inequities due to the limited availability of such tools with localized and culturally tailored features. These populations, often at higher risk of NCDs, could greatly benefit from more culturally sensitive health care technologies, which could improve health care access and outcomes, thereby necessitating careful consideration in future work.

Conclusions

The review highlighted the promising acceptance of CAs by users for the self-management of chronic conditions, with feedback indicating helpfulness, satisfaction, and ease of use in most of the included studies. AI-based CAs present opportunities for communication and interaction in health care settings. However, understanding and optimizing the communication channels between humans and such CAs is crucial for enhancing their capabilities and potential benefits in health care. While our study confirmed the increasing role of CAs in augmenting self-care and potentially reducing health care costs, it also exposed critical limitations, particularly concerning conversational depth and emotional intelligence. This review also emphasized the lack of reliable and comparable evidence to determine the efficacy of AI-enabled CAs for chronic health conditions. Therefore, while user feedback and acceptance were positive, there is a need for more rigorous studies and standardized reporting to evaluate the effectiveness of AI-based CAs as human-like health carers for managing NCDs.