A Survey of Conversational Agents and Their Applications for Self-Management of Chronic Conditions

Abstract

Conversational agents have gained their ground in our daily life and various domains including healthcare. Chronic condition self-management is one of the promising healthcare areas in which conversational agents demonstrate significant potential to contribute to alleviating healthcare burdens from chronic conditions. This survey paper introduces and outlines types of conversational agents, their generic architecture and workflow, the implemented technologies, and their application to chronic condition self-management.

I. Introduction

Chronic conditions (e.g., diabetes, hypertension, and cancer), are long-term illnesses that last one year or longer and require ongoing medical attention [1]. 60% of Americans have at least one chronic condition and 40% have two or more [2]. To illustrate, over 30 million people, approximately 10% of Americans, have diabetes, and two-thirds of them also have hypertension [3]. Chronic conditions are expensive for both clinical and economic reasons. They are the leading causes of death and disability globally and a leading driver of 4.1 trillion in annual healthcare costs in the United States alone [2]. As life expectancy globally rises and the prevalence of associated risks (e.g., unhealthy diet and smoking) are increasing, the burden of chronic conditions gets heavier.

Self-management represents a promising strategy for successful healthcare and positive health outcomes for people with chronic conditions [4]. Self-management refers to the set of relevant skills of individuals to manage and monitor their symptoms, treatment, physical and psychosocial consequences and lifestyle inherent in living with chronic conditions [5]. Self-management prevents the development of secondary morbidity and mortality, helps people with chronic conditions act in concert with providers, and take actions (e.g., self-monitoring) [6], [7]. Self-monitoring, a process of self-management, in turn helps them aware of patients’ symptoms, bodily sensations, and cognitive processes, as well as create the potential to provide information for action [8]. However, people with chronic conditions are often confronted with a range of challenges in managing their conditions routinely such as difficulties in tracking accurate health indicators (e.g., weight, blood sugar, and blood pressure) [9], difficulty in negotiating self-care commitment, and vertical communication healthcare providers and patients [10].

Conversational agents, or dialogue systems or chatbots, have been developed and used in the field of health among many of their applications (e.g., commerce, helpdesk, and websites). Conversational agents are computer systems that communicate with users in natural language in multiple communication channels notably by text or speech [11], [12]. Self-management of chronic conditions is one of the popular applications of these systems via two-way communication and tailored responses that mirror therapeutic processes (e.g., cognitive behavioral therapy) by emulating human-to-human interactions. Through the interactions, the agents engage users in self-management by promoting goal-setting, self-monitoring, provision of feedback, alleviation of negative emotions, and so on. The agents also can provide rich contextual information about the users’ health status [12]. The underlying motivation is to promote preventive measures and continuous monitoring with the help of technologies so that costs for emergency care and hospital admissions can be avoided [13].

Due to the advancement in computer technologies, particularly natural language processing (NLP) and artificial intelligence (AI), conversational agents have the significant potential to play an increasingly important role in healthcare. In particular, in the context of chronic conditions, it has a range of potential to improve patients’ self-management. This opportunity also comes with some challenges regarding the design of interfaces and the application of personal touches for communicating with different users [14], [15]. Thus, this paper aims to broadly outline the concepts, building blocks, underlying techniques, and applications for chronic condition self-management.

II. Conversational Agents

A. Definition

Conversational agents refer to dialog systems that engage with users in the natural language [14], [16], [17]. Dialog systems indicate any human-computer interaction systems that use natural language to understand and generate content to communicate with the user [16]. In other words, a premise of conversational agents is engagement, thus the capability of handling natural language and the contents should not be constrained to a predetermined set of words or limited sentence structure [14], [16]. In this sense, conversational agents are not necessarily the same as chatbots. In literature, the term conversational agents are often synonymously used with chatbots. Chatbots are, however, software components that are a type of text-based and goal-oriented conservational agent [16]–[18]. Chatbots are often designed to simulate conversations with the user or perform a particular task and respond to the conversation in human language with predefined replies and are mainly based on pattern matching [19], [20]. On the other hand, conversational agents are more context-aware, based on theoretically motivated techniques that enable conversation, and use more advanced technologies such as deep learning and natural language understanding [16], [18]. In that regard, conversational agents are said a pragmatic usage of computational linguistics techniques [17].

B. Types of Conversational Agents

Conversational agents can be classified by different criteria such as their communication mode, capability, embodiment, methods of response generation, and domains. Accounting for that natural language can be either written or spoken, the mode of communication of conversational agents can be differentiated by text, voice, or combined. To facilitate engagement, conversational agents are required to process the input and provide relevant output (e.g., advice or feedback) in a form of text, speech, graphics, haptics, gesture, handwriting recognizer, etc [21]. Text-based agents work through text messages or web platforms [22]. The design interfaces of this type of agent aim to support text messages. For example, chatbots like ELIZA [23] developed by Joseph Weizenbaum in 1966, simulate conversations similar to those in social media and service platforms (e.g., banks, bookings, and e-commerce), using text communication mode [18]. Speech or voice-based agents communicate with users through voice. Assistants like Amazon Alexa, Apple Siri, Microsoft Cortana, and Google Assistant Now are more commonplace examples. This type of agent integrates speech technology to recognize human speech [14]. The inputs from users trigger specific information retrieval or function, and then the agents respond with synthesized speeches [14]. Speech technology in voice-based agents is to support a range of purposes, including conducting small talk or expressing emotions [24], supporting learning [25], and fostering communication with others [26]. Some studies address that voice-based agents can better assist children and the elderly in a more natural form of interaction [27], [28]. Conversational agents with multimodal communication models use text, voice, non-verbal communication (e.g., hand gestures and facial expressions), and body stance in combination [29]. This type of agent encompasses physically- (e.g., social robots) and graphically embodied (e.g., virtual agents with a human-like body on a computer screen or avatars) agents [29].

Conversational agents also can be classified into two categories based on their capabilities [14]. Conversational agents’ main capabilities are often constrained to communication only and do not execute actions. ELIZA, Mitsuku [30], and Cleverbot [31] are examples. ELIZA is a basic person-centered or Rogerian psychotherapist chatbot. Its users type questions and ELIZA answers in the input box. ELIZA simulates conversations based on the pattern matching directives, MAD-Slip, that were provided in separate scripts and used rules to respond with to user inputs [32]. Mitsuku is a social chatbot that tells stories and plays games with its users while simulating a young female from Leeds [33]. Other agents, on the other hand, are capable of taking specific actions either in the physical or virtual world such as turning on lights or booking a flight [17]. Alexa [34] and Bluebot [35] are examples of those with capabilities of conducting actions.

The conversational agents also vary by their forms of representation: none embodiment, virtual (static or interactive), and physical embodiment. Those agents with no embodiment lack any form of representation and have a natural language interface only [36]. Embodied conversational agents have representation that demonstrates the same or similar properties as humans in face-to-face conversation, including the ability to communicate through both verbal and nonverbal means such as gestures, facial expressions, and body posture [29]. A physical-based embodied agent has a physical body such as a social robot or mobile robot. Embodied conversational agents encompass both virtual, graphical, and physical agents [17]. These forms of embodiment were found to influence users’ perception of a conversational agent and offer a range of design options [36], [37]. For example, an overly realistic representation of the agent negatively influences the users’ willingness to disclose personal or sensitive information [38], [39]. Users perceive a conversational agent with matching ethnicity of virtual representation as more enjoyable, sociable, and useful [40]. These embodiments of conversational agents also found to create a personal connection with a user through verbal and non-verbal interactions [14].

How to generate a response is another criterion to categorize conversational agents into two groups: rule-based and generative-based [41]. Rule-based agents are provided with pre-determined rules and a database of responses, rather than generating their own answers. ELIZA is an example of a rule-based agent that uses pattern matching. Generative-based agents use machine learning algorithms to generate responses based on the data provided and keep on learning and improving based on learning models. These types of agents with AI technologies are also called AI-based or intelligent agents.

The agents can also be differenced on the basis of domain: general or domain-specific. In the general or open domain, conversational agents are to support conversation in various topics and disciplines including entertainment [40]. Examples include Botta, an Arabic dialect education chatbot [42], AliMe [43], SOGO, a semi-automatic negotiation dialog system [44], and ChatGPT [45]. Goal-oriented conversational agents focus on particular topics in a specific domain to assist target users with completing a particular task, such as healthcare, education, and business [41], [46]. For example, in the medical domain, conversational agents can be applied to socially support patients with medical conditions or learners in the learning process with answers to specific health-related queries on particular health conditions such as HIV/DIDS, children’s health, and mental health [41].

III. System Architecture and Enabling Technologies

The rise and maturity of technologies, such as speech recognition and NLP have enabled the omnipresence of conversational agents. Thus, this section outlines a high-level genetic architectural overview, a foundational set of building structures [47], which can be used for developing a range of applications and related design technologies.

A. Architecture

The specification of conversational agent architecture varies by type and domain, but the generic diagram suggested by Kahn [47] comprises layers of presentation, business, service, data, and utility. The presentation layer encapsulates the various user interface components that are implemented and displayed for the users. The users interact with the agent via this layer using communication channels such as email, SMS, text messaging, voice, or IoT devices [20]. The core responsibility of this layer is supporting multi-channel and multi-platform to make the agent scalable and extendable enough to work with different channels and platforms (e.g., iOS or Android) [20]. To support multiple channels and platforms, user interfaces may contain any interface components that enable easy communication with the current and new platforms. The presentation layer then interacts with the underlying business layer components to access the functional capabilities of the agent [47].

The business layer is responsible for data processing, data formatting, and dialog management [47]. The data processing involves multiple steps that have to be performed in the correct order. Besides, the data from the service layer also need to be converted into real-world business entities (e.g., orders and products). The data processing components often gather data from services and identify common entity usage patterns and implement them as separate components. Data formatting is to convert the data into the required format that is in line with UI requirements in the presentation layer and capabilities. All the data from the service layer components needs processing to be structured as per the requirements of an agent. This process is needed to make a conversational agent extendable and scalable enough for any existing and new platforms and channels [47]. Dialog management is responsible for managing dialog with the user. This component is one of the core and unique components of a conversational agent solution that determines the user experience. For a flawless experience for the user, dialog management should be able to manage the context of dialog, the user’s preferences, and profiles [47]. With the assistance of specific AI/NLP services, Natural Language Understanding (NLU) enables the agent to understand and respond to the user by converting an abstract statement into a natural language surface utterance [20].

The service layer is responsible for integrating with external party services (e.g., business functionality and middleware) for sharing the internal and external data for further usage [47]. The components of the service layers can be deployed on the same tier or a separate tier, depending on the performance and requirements of the implementation. The service layer can consist of NLP services, data access services, and external service interfaces. An NLP service has NLU as a fundamental service component that can determine the success or failure of an agent. The degree of efficiency of an NLP service relies on its machine learning capability which is based on the number of and advancement of AI algorithms. Data access services serve as adaptors to provide access to a set of services that convert the data from those services into a format that the other components of an agent understand. An agent may have this layer separately or on the same layer, depending on the requirements of the implementation. External service interfaces manage the integration of a different set of external services (e.g., social media, CRM, and support center). The specific services may vary by the functionalities provided by an agent [47].

The data layer concerns access to local data collection, which is provided to other layers via the service layer [47]. These storage systems serve as the basis of many services and components of an agent. For instance, the history of communication with users, analysis of collected data, and performance of machine learning techniques rely on this layer. In case of an intelligent agent based on AI may consist of Knowledge Base and user interaction history in this layer [20]. The knowledge Base represents the domain-specific source content for training an agent [20].

The utility layer provides common services such as security, logging, and configuration. This layer includes configuration and security. These are not considered to be a functional part of an agent but play a significant role in the overall operations of any system implementation. The components of this layer affect the success of an agent’s feature and functionality by determining the scalability and repeatability of characteristics of an agent, and security management. Fig. 1 portrays an architectural view of the generic conversational agent solution and its ecosystem.

Fig. 1.

A generic architectural overview of conversational agents (redrawn and adapted from Kahn [47])

B. Workflow and Related Techniques

A common workflow for both rule-and AI-based agents is pre-processing, processing, and generation [46] as shown in Fig 2. The figure depicts the main three steps and associated components of a conversational agent. These steps involve different design approaches that may employ various design methods and techniques that are introduced below.

Fig. 2.

Workflow and associated components of conversational agent (redrawn and adapted from Maroengsit et al.,[41])

1). Pre-processing step

Once a user launches his/her interaction with an conversational agent via its user interface, the pre-processing step revolves around NLP to collect external data and format data for other processes of the agent. This step involves design methods such as pattern matching, parsing, Artificial Intelligence Markup Language (AIML), and statistical techniques like Term Frequency-Inverse Document Frequency (TF-IDF) and word2vec:

• Pattern matching technique classifies the user input as “pattern” and produces a suitable response stored in “template” [46]. Although the complexity of algorithms of this technique varies, it basically uses “pattern”-“template” pairs which are hand-craft. This technique is the most commonly used approach which is used in both early and modern conversational agent design [46]. A disadvantage of the technique is limited capabilities because manual building restricts the scaling, making the agent’s responses predictable and repetitive [46].

• AIML is to create conversational flow in an agent. AIML is made up of data objects, which consist of two units: topics and category units [46]. The topic has a name attribute ad a set of categories related to the topic. A category is a rule which comprises two elements “pattern” and “template” [48]. A user’s input is matched using “pattern” and “template” is used to generate a response using [48]. AIML is a powerful design tool but requires NLP programming too.

• Parsing is a method to extract meaningful information from a text input by parsing that text input into a set of words that can be stored and manipulated. The extracted word or keyword is matched against the documents in the corpus to find an appropriate response [23]. Semantic parsing, which is a more advanced form of parsing, converts the user’s input to a machine-understanding representation of its meaning [23].

• Chatscript is an open-source authoring tool that combines a natural language engine and dialog management system for interactive conversation [46]. This rule-based engine uses rules dialog flow script creates. Machine learning tools can also be used to improve dialog flow [46].

• Domain ontologies are also used to replace a manually built knowledge base. Domain ontologies allow an agent to explore the concept nodes of an ontology to establish the relationship between concepts [49] or enable to yield standardization and sharing [50].

• TF-IDF is a statistical technique to quantify the importance or relevance of text representations (e.g., words, lemmas) in a document among a corpus [51]. It is often used as a weighting factor in information retrieval, text mining, and user modeling. The weights of text representations are measured by calculating word frequencies (TF) and by multiplying TF with inverse document frequency [51]. TF-IDF is a basic way to convert a document of words into numbers but it can be used to obtain a vector of words frequencies [52].

• Word2vec is a group of models for NLP that turns every word in text into a unique vector called “word embeddings” based on the context in which they appear in text [53]. This NLP technique can be used in conversational agents to improve the ability of the agent to understand the meanings of user inputs and to generate more appropriate responses. An example of using word2vec in a conversational agent is using pre-trained word2vec embeddings to initialize the word representation weights for individual utterance [54].

2). Processing step

The processing step mainly relies on NLU to collect and manage conversations based on pre-processed user input data [41]. Intent classification and name entity recognition (NER) can be integrated into NLU service of a conversational agent [20]. When an agent has machine learning components, it can consist of NLP, NLU, and decision engine.

• NER is an NLP technique to extract and classify mentions of designators from text such as biological species, temporal expressions, and names. In conversational agents, NER can be a component of NLU [55]. NLU maps an input text into a pre-defined semantic slot according to scenarios. In this process, NLU typically uses two types of presentations: the utterance level such as the user’s intent category; and the world-level information extraction such as named entity recognition and slot filling [20].

• Intent classification is the process of automatically categorizing users’ intent by analyzing the user input into intents such as purchase, unsubscribe, and demo requests [56].

• Decision engine decides what response to generate based on information in the knowledge base. The engine learns domain-specific knowledge from the source data to generate the response [20]. Many recent intelligent agents employ neural networks in NLP, such as language modeling [57], paraphrase detection [58], and word embedding extraction [59]. The use of machine learning has risen in the conversational agent development [46].

Especially artificial neural network (ANN) models and their variants, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), sequence-to-sequence neural networks (Seq2Seq), and long short-term memory networks (LSTMs), are one of the learning algorithms within machine learning [46]. The presence of a learning algorithm in the ANN approach distinguishes intelligent AI-based agents from rule-based ones [46]. ANN-based agents can use both retrieval- and generative-based approaches for response generations. The use of deep learning neural networks, which are capable of learning unsupervised from unstructured or unlabeled data, has dominated intelligent conversational modeling, using especially RNNs and LSTMs [46]. Recently NLU with transformer model also have proven to offer benefits as a building block of conversational agents [60].

• An RNN is a recursive ANN. RNNs save the output of a layer and feeding this saved output to the new input to predict the next output. Due to their ability to remember the previous computation, RNNs useful in understanding the semantical meaning and conversational context based on the previous words in the sentence. This ability to handle the sequential nature of natural language changed the landscape of conversational agent development as the understanding of conversational context is essential to understand the user inputs and generating more contextually appropriate responses [61].

• Seq2Seq models are designed based on RNN architecture and have two RNNs, which are an encoder and a decoder. The encoder encodes the <status> or input sentence. The decoder decodes the <status> and generates the desired responses [61]. These two neural networks enable the model to process variable length input sentences. In conversational agents, this model is widely used to generate responses by converting between input <status> and output <response> [62].

• LSTMs are a special type of RNN, which is designed to resolve the RNN model’s long-term dependency issue [63]. To hold the previous information for a long period of time, memory cells and gates are introduced in LSTMs. The memory cells store information in, write new information, and read information from them, and the gates are used to control the flow of information [64]. The capabilities of LSTM show higher performance in learning from experiences, classification, and processing and prediction of time series as compared to other available RNNs, hidden Markov models, and other sequence learning models. These abilities make LSTMs useful in designing conversational agents [64].

• Transformer proposed by Vaswani et al., [65] is a new architecture that is based on encoder and encoder that are built on self-attention mechanism, and parallelizable to understand speech sequences [66]. This model utilizes pretrained models that are trained for the patterns and features of a sizable dataset, can be adjusted for other tasks with a minimal amount of additional training data [15], [17], [21]. A transformer model has ability to processes many words at the same time which overcomes slow training time of LSTM and RNNs [65]. Examples of those models include BERT (Bidrectional Encoder Representations from Transformers), RoBERTa (Robustly Optimized BERT Pretraining Approach), and GPT (Generative Pre-trained Transformer). In order to learn the semantics of words and contexts, BERT, developed by researchers [67] at Google in 2018, is an open-source machine learning framework for NLP. BERT consists of Transformer encoder layers and is pre-trained on two tasks: language modeling and next sentence prediction that enable BERT to learn latent representations of words and sentences in context [67]. After pretraining, the model can be quickly and easily adjusted to suit various tasks with fewer resources on smaller datasets to optimize its performance on specific tasks such as sequence-to-sequence-based language generation (e.g., question-answering and conversational response generations) [67]. RoBERTa is built on BERT by Yinhan et al [68]. This language model optimized key hyperparameters of BERT and trained with a larger training data size [68]. RoBERTa has the same architecture as BERT but is trained with bigger batches over more data and longer sequences, removed the next sentence prediction objective, and uses dynamic masking [68]. GPT is another family of language models by OpenAI that uses several blocks of the transformer [69]. A GPT model is generally trained on an enormous corpus of publicly available datasets and generates human-like texts based on a given input [69]. The latest version is GPT-4 as of March 2023, which can accept images as well [69]. Conversational agents with transformer models hold the potential to understand context around a word over a longer distance in a more effective way, so that the agent delivers a better response.

3). Generation step

In this last step, an agent responds to the user whether the response is either to provide information to the user or to ask a question to gain more information from the user [41]. This step evolves around natural language generation (NLG), converting an abstract statement into a natural language surface utterance [20]. To construct a response and render the output, dialog manager first uses response models (e.g. generative and retrieval-base), and then returns a response based on the priority [70]. Retrieval-based agents search a user utterance in a database and return the best matched response [20]. Agents used generative approaches typically employ machine learning models aforementioned to synthesize new responses. For instance, an LSTM-based agent incorporates the encoder-decoder to incorporate the question information, semantic slot values and dialog act type to generate correct responses [20].

IV. Conversational Agents for Chronic Condition Self-Management

Healthcare is an area where conversational agents have gained their ground along with business. Besides, conversational agents show the potential to play an increasingly more significant role, such as assisting healthcare providers and patients with consultation and supporting health behavior changes in their own living environment [17]. Self-management supports, such as continuous feedback and reinforcement, from conversational agents present the potential to reduce the health system burden and to empower people with the skills, knowledge, and confidence needed to take an active role in managing their health conditions [71]. This section examines how conversational agents have been discussed for self-management.

A. Characteristics of Chronic Conditions

Conversational agents that are applied for chronic conditions are domain specific that focus on specific conditions and tasks. Psychological conditions, such as depression and anxiety [72]–[74] are the most commonly addressed target condition in literature [14], [75], [76]. Other conditions include obesity [77], [78], diabetes [79], [80], cancer [81], asthma [82], [83], neuropsychiatric disorders [74], [84], cardiovascular diseases [85]–[87], neurological disorders [88], chronic obstructive pulmonary disease (COPD) [83], arthritis [89], hypertension [90], and hearing impairment [91]. These agents are designed to assist with maintaining a healthy lifestyle [77], [92]–[95], managing symptoms [72], [96], participating in social activities [97], problem-solving [98], working with healthcare professionals [99], [100], managing medications [99], and a combination of these. (see Table I for examples)

Table I.

Examples of Conversational Agents Attributes for Chronic Conditions Self-Management

Chronic Condition	Example Agent	Target Area	Theory or Evidence/Evaluation	Communication Mode	Techniques
Alzheimer	N/A By Griol et al	Disease monitoring	N/A Ease of use	Multi-mode (voice & visual)	AI-based: NN, ML, ASR, NLU, NLG, TTS
Asthma	mASMAA by Rhee et al [82]	Disease monitoring and treatment adherence	N/A User satisfaction	Text	AI-based: NLP
Breast Cancer	Vik By Chaix et al. [81]	Education, &Treatment adherence	N/A User satisfaction	Multi-mode (text & visual)	AI-based: NLP, ML
Cardio-vascular	Personal Health Management Assistant by Ferguson et al. [85]	Symptom monitoring	N/A Feasibility	Multi-mode (text & voice)	AI-based: SR, NLP
COPD	Avachat by Easton et al. [83]	Info seeking, problem solving, & motivation	N/A User Acceptance	Multi-mode (text, voice & video)	AI-based: SR, NLP
Depress-ion	N/A By Ly et al. [94]	Symptom monitoring	Cognitive behavior therapy / Perceived Usefulness	Multi-mode (voice & video)	AI-based: Emotion algorithms, ML, NL
Diabetes	Healthy Coping by Cheng et al. [79]	Symptom monitoring	N/A User satisfaction & Ease of use	Voice	AI-based: SR, NLU, ML
Heart failure	CARDIAC by Galescu et al. [86]	Monitoring & Self-care	N/A Perceived Usefulness	Voice	AI-based: NLU, SR
Parkinson	Harlie by Ireland et al. [88]	Education & Disease monitoring	N/A Ease of use	Multi-mode (voice & text)	AI-based: NN, ML, ASR, NLU, NLG, TTS
Psycholo-gical disorders	Pocket Skills by Schroeder et al. [96]	Goal-setting, skill practice & self-tracking	Dialectical behavior theory / Symptom reduction	Multi-mode (text, voice & video)	N/A
HIV /AIDS	N/A by van Heerden et al. [100]	Disease monitoring	CDC guidelines for HIV counseling and testing / Sentiment toward the agent	Multi-mode (text, voice, & muti-media)	N/A

Notes: AIML: artificial intelligence markup language; ASR: automatic speech recognition; API: application programming interface; DL: deep learning; NLG: natural language generation; NLP: natural language processing. NLU: natural language understanding; NN: neural network; ML: machine learning; SR: speech recognition; STT: speech-to-text; TTS: text-to-speech.

Conversational agents for self-management are discussed as suitable, personalized, and affordable solutions to react to challenges of self-management and slow down individual disease deterioration. Studies report that conversational agents have the ability to simulate a therapeutic process such as cognitive behavioral therapy or motivational interviewing [101] which goes beyond traditional healthcare sites and reaches personal settings [102], [103]. This could promote a range of self-management skills, including goal setting, provision of feedback, education, behavioral changes, self-care, illness monitoring, and personalized treatment [5]. Some agents exclusively focus on particular self-management skills such as monitoring [86], [91] and support [104], [105]. Some agents support multiple self-management skills [79], [87].

B. Acceptance and Efficacy

Despite limited, evidence suggests conversational agents are useful and effective for engaging users, and acceptable to those living with chronic conditions [75], [76], [106]. Conversational agents were found to be effective in a range of aspects, including treatment outcomes, adherence to self-care, and health literacy. To illustrate, the agents were found to help alleviate negative emotions or symptoms of mental problems by forming an alliance or rapport with patients through conversations and showing empathy [12], [107]. The agents are also capable of supporting natural health routine tasks when in-person health care is not available [108], [109]. Furthermore, the agents can support the understanding of the health status of the users through rich linguistic information using conversational syntax and semantics [12]. Its embodiment and design also affect users’ attitudes toward conversational agents. For example, text-based chatbots are found to engage users about sensitive or stigmatized topics which alleviates negative symptoms or emotions [12], [110]. Voice-based conversational agents were found to facilitate users with low literacy or with visual [111], intellectual [112], linguistic, motor, or cognitive disability [113]. AI-based agents also demonstrated positive feedback for chronic condition self-management regarding its helpfulness, satisfaction, and ease of use [76].

Several studies reported the efficacy of conversational agents in a range of aspects [72], [73], [114]–[117]. To illustrate, the engagement was found beneficial to treatment outcomes [118], [119]. The effects of utilizing conversational agents in treating chronic conditions are not superior to active intervention with human psychological therapy [98], [120]. However, in some controlled studies such as studies by Freeman et al., [108] and Pinto et al., [109], the agents were found relatively effective, showing small-to-large effects on chronic condition intervention. Study participants also reported satisfaction with accessibility, availability [94], [115], interactivity [95], agents’ ability to form a relationship and show empathy [72], [94], [97] and learning from input [72], customizability of its gender and appearance of embodied conversational agents [94], and the option to tailor the session to the users’ own needs [94].

There are, however, challenges in using conversational agents to provide interventions. The noticeable reasons include repetitive content [72], [94], [97], [106], [116] and limitations in understanding user inputs or generating responses appropriately [96], [98] and tailoring content and response to the users’ needs [117]. Concerns for privacy invasion and reduction in independence were also reported [75].

C. Attributes of Conversational Agents

A range of types of conversational agents have been developed and designed for chronic condition management. Besides, various technologies are utilized in those agents to make them efficient and useful in managing their target conditions. To illustrate, the conversational agents are developed as chatbots [72], [92]–[96], [99], [106], [122], voice-based [79], [79], [87], [88], [91], [105], [123], and embodied as robots or virtual human-like agents [101]. Their dialogue management strategies are mostly either finite-state or frame-based, letting either the users or the systems initiate the conversation [72], [94]–[98]. These agents have mobile, web, text messaging, and robot interfaces. Some conversational agents have multiple interfaces such as smartphones and the web. The conversational agents found in the literature utilized various techniques including speech recognition, NLP, text-to-speech, speech-to-text, AI techniques including machine learning and deep learning.

Some of the conversational agents for chronic condition self-management are developed based on theories of behavior change or contained evidence-based content. To illustrate, an agent developed by Elmasri and Maeder [92] for substance use was developed based on the alcohol use disorders identification test, which is a method of screening for unhealthy alcohol consumption or any alcohol use disorder. Another chatbot to assist with substance use developed by Kazemi and colleagues [93] was based on ecological momentary interventions (EMI) and the transtheoretical model of change. EMI moves to intervention from clinical settings to an individual’s daily life [124]. The transtheoretical model construes health behavior changes are an intentional process that takes multiple stages [125]. A chatbot for depression management developed by Gaffney and colleagues [106] was also based upon a behavioral theory, i.e., perceptual control theory. This theory posits human behavior varies in a way of keeping their perception of the world [126]. The task-oriented agent developed by Watson and colleagues [127] employed behavioral and social cognitive theory to promote exercise behavior change to manage overweight. Several chatbots were based on cognitive behavioral theory. For example, the chatbots developed by Fitzpatrick et al., [72] and Ly et al., [94] both for mental health management used cognitive behavioral theory (CBT). Another chatbot developed by Stein and colleagues also employed the same theory for diabetes management. CBT indicates a range of psychological treatments grounded on the theory that our thoughts, emotions, sensations, and behavior are all linked, therefore, our thoughts and actions affect the way we feel [128]. Another chatbot developed by Schroeder and colleagues [96] adopted dialectical behavior theory which is the modified CBT to apply CBT to broader mental health conditions by developing healthy ways to cope and regulate emotions and stress [129]. Some other chatbots utilized established evidence. For example, van Heerden and colleagues utilized CDC guidelines for HIV counseling in non-medical settings to assist symptom and healthy lifestyle management and collaboratively work with healthcare providers [100]. Wang and collaborators used the PubMed medical information database for their conversational agents for healthy lifestyle management and social activity participation [97].

D. Evaluation Measures

Conversational agents that target chronic condition management were assessed in three dimensions: technical performance, user experience, and health-related aspects [130]. Technical performance was assessed with accuracy [80], [90], [116], [131], specificity [80], [116], [132], [133], and task completion rates [80]. To illustrate, Azzini and colleagues [90] evaluated the task completion rate and the accuracy of their voice-activated agent. Rehman and colleagues [80] calculated Cohen’s D to assess task completion and the accuracy of their agent for diabetic conditions. Bickmore et al [133] and Philip et al [134] assessed the specificity of their agents that were developed for depressive disorder.

As for user experience, commonly evaluated features found in the literature are helpfulness, satisfaction, and ease of use [130]. According to literature reviews [75], [130], the majority of studies included in their reviews reported high to moderate to high positive user experiences. However, some studies reported dissatisfaction. For instance, in Baptista et al.’s study, some participants reported that the embodied agent for diabetes coaching was annoying and boring whereas the majority of other participants reported the agent is helpful and friendly [135]. Schroeder et al.’s participants also addressed the difficulty in engaging with the agent [96].

Regarding health-related measures, the most common measures found in the literature are symptom relief [30] and improvement in self-management skills. These studies used methods like quasi-experiments and randomized controlled trials.

One issue with these evaluations of conversational agents in healthcare is inconsistency and invalidated questionnaires [15], [75]. Inconsistency found in the reporting of design methods includes the number of study arms, method of assignment, allocation ratio, outcome measurement, and attrition [15]. To illustrate, the studies that tested the engagement of patients using highly variable metrics for engagement from 30-minute sessions over 2 weeks [114], a single session [98], [136], daily sessions for 2 weeks [72], [94], to unlimited access for 30 days [115]. To improve the consistency in reporting and the validity, Laranjo et al [15] suggest using the followings: Consolidated Standards of Reporting Trials of electronic and mobile health applications and online telehealth [137], the Transparent Reporting of Evaluations with Nonrandomized Designs statement [138], and the Standards for Reporting Diagnostic accuracy studies [139].

V. Conclusion

Despite mostly being preliminary and limited, the findings in the literature suggest conversational agents show advantages for self-management of chronic conditions. The overall acceptance of conversational agents for the self-management of chronic conditions was found promising. Particularly the recent advances in NLP and AI present multiple benefits for healthcare providers and those living with chronic conditions by supporting them with behavior changes and completing specific tasks for managing their conditions. These features of conversational agents make scalable, less costly medical support solutions for chronic condition management. However, there are limitations; firstly, more reliable and comparable evidence is needed to determine the efficacy of conversational agents for chronic conditions due to inconsistent reporting of technical implementation details, users’ attitudes towards agents, and medical efficacy.

This survey paper is similar to others in terms of outlining the definition, architecture, and underlying NLP techniques of conversational agents. But this survey paper extended prior similar work by examining the applications of conversational agents for chronic condition self-management including agent examples of agents, medical context, and theoretical framework applied in the development of those agents with specific purposes.