Recent advancements in natural language recognition and synthesis have resulted in the adoption of conversational agents (CAs) in many fields. CAs can be defined as systems that support conversational interaction with users by means of speech or other modalities [1]. The rising popularity of conversational technologies has been facilitated by a renaissance in Artificial Intelligence, the development of powerful processors supporting deep learning algorithms, and technological advancements, making a large amount of computationally accessible knowledge available [1].

One emerging area in which conversational technologies have been increasingly used is health care. A recent systematic review in this area examined technical performance, user experience, and health-related outcomes and found that most studies had not employed standardized evaluation methods or had failed to address aspects of patient safety [2]. There have also been other recent reviews on health care conversational agents [3-5]. This study differs from them in that it focuses on the implementation of personalization in health care conversational agents.

Personalization is:

When applied specifically to digital technologies, personalization can be defined as:

A recent interdisciplinary review study proposed a framework to characterize personalization along three dimensions: (1) what is personalized (ie, content, user interface, delivery channel, and functionality); (2) for whom is it personalized (either a specific individual or a user group, eg, elderly women); and (3) how automated is the personalization (how the information needed for user modelling is collected) [7]. The personalization process involves user models containing characteristics, preferences, interests, and needs of users as the basis for providing adaptive information and services. Depending on the degree of automation, two types of personalization can be distinguished: implicit and explicit. In implicit personalization, information needed for user models is obtained automatically through the analysis of observed user activities and interactions with the system. In explicit personalization, information needed for user models requires users’ active participation in obtaining the required information.

One of the earliest applications of personalization in a conversational system was Grundy, a virtual librarian that delivered book recommendations [8]. To build a user model for personalization purposes, Grundy asked questions at the beginning of an interaction and associated users with predefined stereotypes. After the initial user provided information, the user model was updated implicitly over time during conversations. It was a hybrid system bringing together both explicit and implicit personalization. This foundational work on personalized CAs has been followed by a range of works focusing on dialogue management [9], personalized messages [10], recommender systems [11], and adaptive systems [12].

Personalization in CAs can be achieved implicitly by processing past interactions with users [11,13] or explicitly by user-entered information at the set-up time [8] or using ongoing confirmation style input [14]. The messages presented to users [10], or the conversational style of systems [15], can be personalized. Personalized and adaptive system behavior in conversational systems can improve user comprehension [16], user satisfaction [17], task efficiency [18], and the likelihood of behavior change [19]. Furthermore, personalization can be an essential system feature for voice interfaces due to the limitations in presenting large amounts of information through a voice-only modality [20]. The effects of personalization have been evaluated in various ways by measuring aspects like efficiency in terms of the number and duration of interactions [11,20], user satisfaction, relevance, and understandability [20], information quality presented [21], and appropriateness of system responses [10].

Studies of personalization in health care and medicine have been increasing in number since the early 2000s [22], with growing evidence showing their effectiveness [23-26]. One important limitation in the health care personalization literature is equating it to genomics-supported efforts in medicine [27]. Genomic markers are only one dimension of personalization that helps to recognize the uniqueness of individuals and make their medicine personalized [27,28]. There are other factors that affect this personalization of health care, such as people’s lifestyle choices, their socioeconomic context and living environment, and other health care services that can be personalized like health education and therapies [29].

A review of behavior change interventions characterized four intervention groups according to their degree of personalization in the messages delivered to individuals: generic (one-size-fits-all messages), personalized (messages with the person’s name), targeted (messages specific to a subgroup of the general population), or tailored (messages specific to an individual’s characteristics) [30]. The review found that 78% (11/14) of the tailored and 95% (22/23) of the targeted interventions reported improved outcomes, with 54% (6/11) of the tailored and 68% (15/22) of the targeted interventions being statistically significant.

Dialogue systems can offer fine-grained possibilities to personalize the information to be delivered:

Learning from a history of previous conversations plays a key role in ensuring the continuity of health communications that take place over multiple interactions over time [31].

Informed by the recent theoretical developments in personalization [7], a broader understanding of personalization in health care [22,29], and an increasing interest in health care CAs [3-5], this study aims to review the use of personalization in health care CAs and characterize the methods that have been applied to implement this personalization. Aligned with the rapid advancements in natural language processing technologies used in CAs [1] and the increasing adoption of CAs using unconstrained natural language [32], this review focuses on agents with unconstrained natural language input capability. These agents include chatbots, which can engage in small talk or casual dialogues, embodied conversational agents, which feature computer-generated visual virtual characters capable of both verbal and nonverbal communication, and commonly available smart conversational interfaces such as Apple’s Siri, Google’s Google Assistant, Samsung’s Bixby, and Microsoft’s Cortana [1,33,34].

This review uses the search protocol of an earlier systematic review that was performed between April 2017 and February 2018, with a focus on technical performance, user experience, and the health-related outcomes of CAs in health care [2]. The current review has: (1) focused on the use of personalization features in CAs that were not examined previously; (2) used the same inclusion and exclusion criteria as the review by Laranjo et al [2] with an additional criterion on personalization (ie, the studies with no personalization features were excluded); and (3) performed a new search in March 2019.

We searched in the PubMed, Embase, CINAHL, PsycInfo, and the ACM Digital Library databases, and did not restrict by the publication year or language. The search terms included “conversational agents”, “dialogue systems”, “relational agents” and “chatbots”. The complete search strategy is available in Multimedia Appendix 1. In addition, the reference lists of relevant articles and grey literature identified in those databases were also included for screening.

The identified publications were included if they: (1) were primary research studies that focused on consumers, caregivers, or health care professionals; (2) involved a conversational agent; and (3) tested the system with human users. The studies were excluded if they involved: (1) user input by clicking or tapping an answer amongst a set of predefined choices, or by using the telephone keypad (eg, interactive voice response systems with dual tone multi frequency); (2) output not generated in response to what it received from the human user (eg, predefined and preprogrammed messages that are not dependent on the information obtained from or about the user); (3) question-answer type interactions; (4) asynchronous communication technology such as email; or (5) no personalization features. Furthermore, studies evaluating only individual components of a conversational agent, like automatic speech recognition, or using Wizard of Oz methods were excluded.

Screening was conducted independently by two researchers to extract data from each study. Cohen kappa was used to measure intercoder agreement between the researchers. Any disagreements between the assessments of two researchers were resolved by consensus agreement. To identify the relevant information, the researchers used the personalization definition presented in the introduction section. In addition, the following keywords were used as a guide to identify personalization-related information within the studies: personalizing, adapting, customizing, tailoring, configuring, individualizing, modifying, changing, altering, transforming, modelling, tuning, setting, preference, and profile. The data extraction process was guided by an assessment scheme based on the personalization framework offered by Fan and Poole [7]. In addition to these dimensions, we included three more dimensions to provide further details on the included studies: purpose of personalization, methods to evaluate personalization, and outcomes in relation to personalization. Table 1 summarizes the final assessment scheme for personalization.

The first search found 1513 papers, and the updated search found an additional 445 papers (Figure 1). After the subsequent title, abstract and full text screenings, 13 studies were included in this review [35-47]. The first search’s kappa statistic for the title and abstract screening was 0.45 (fair agreement) and for the full-text screening it was 0.53 (fair agreement). For the updated search, the kappa score was 0.77 for the title and abstract screening (substantial agreement), and 0.61 for the full-text screening (substantial agreement). The list of excluded studies, their major themes, and the reasons for their exclusion are available in Multimedia Appendix 2.

Table 2 and Table 3 summarize the personalization features of CAs in the included studies [35-47]. For both tables, studies evaluating the same conversational agent were grouped together. Since the delivery channel and functionality were not personalized by any of the studies, they were not included in Table 2. Out of the 13, 8 studies supported patients and 5 studies supported both patients and clinicians. Regarding the target of personalization, all the studies implemented individuated personalization (targeting an individual user). However, one study employed categorical personalization (targeting a group of people) to differentiate novice and expert users in addition to the individuated one [41].

Information needed for personalization was provided explicitly by the users in seven studies [35,37-40,44,45], and obtained implicitly by the system in one study [36] where the conversational agent analyzed users’ audio-visual features, such as facial expression and head position, to determine its feedback. A mix of implicit and explicit methods was employed by five studies [41-43,46,47]. Across all the studies, data explicitly entered by the users included personal goals [35,37,46,47], symptoms and medications [37,44,45], measurement of vital signs [39,40], knowledge level on a specific topic [38], and daily practices [38]. User data implicitly obtained by the systems primarily involved the analysis of conversation history. Differently from the rest, one study analyzed users’ voice and nonverbal facial gestures to determine narrative skills of the users [36].

Personalization was primarily used for tailoring the content to be delivered. Personalized content included: (1) feedback on mood states [35], narrative skills [36], symptom summaries [37], meditation practice [38], and current progress towards the goals set [39,40,46,47]; (2) reminders [37,47], warnings, and alerts [39,40,42,43]; (3) multimedia [35,46]; and (4) questions on pain [41], physical activity [46], and health status [42,43].

Two studies personalized the user interface through changing conversational styles according to users’ motivation state, users’ level of expertise with the system, and dialogue history [38,41]. For example, one study used either didactic, relational, or motivational conversational styles based on the user profile and progress [38]. While the didactic style was used for training-related conversations, the relational style was used at the beginning of sessions to improve user engagement based on the answers received from the user. The motivational style was employed to gather progress-related information and then to provide customized responses to support users. In a simpler implementation, another study used shorter question formats for follow-up sessions [41].

The purposes of providing personalized content and conversations were to: (1) improve user engagement [35,37,38] and dialogue quality [42,43,54]; (2) provide timely feedback [39,40], adaptive user support [41], and adaptive training [36,38]; and, (3) support self-reflection [45,46].

Only two studies directly assessed users’ perceptions of personalization via custom-built questionnaires with questions on adaptive features [38] or via interview questions on tailored feedback [47]. One study employed a virtual coach to teach mindfulness and meditation [38]. The intervention group participants found the experience more rewarding, enjoyable, beneficial, and engaging than the control group participants. The coach’s ability to provide customized feedback was the most successful feature, but this was only rated neutral to mildly positive (0.3 on a –2 to +2 Likert scale). Another study evaluated a relational agent to promote exercise and sun protection [47], with a total of 85% (29/34) of the study participants finding the tailored feedback helpful for achieving their behavior change goals [47]. The remaining studies did not directly evaluate the personalized features. Rather, they focused on evaluating factors that could be associated with personalization, such as user satisfaction, user engagement, and dialogue quality, or effects of personalization, such as improved skills, self-management, and awareness of the user’s health status. One study conducted user interviews in which the users made positive remarks on personalization features [40].

The use of CAs with unconstrained language input in health care is still limited, but there has been a notable increase in the number of studies in recent years. Almost half of the papers included in this study were published in the last two years. While most studies used quasi-experimental study designs, only two used randomized controlled trials [35,44]. Considering the recent emphasis on the role of replication in health informatics [55], the lack of technical details on conversational systems used in the studies is a major obstacle impeding replicability. In terms of personalization, our review found only 13 studies with personalization features. The studies provided various forms of personalized content, however, they were implemented without being supported by any prior evidence showing their effectiveness or any theoretical frameworks underpinning personalization [7,56,57]. Only three studies explicitly mentioned utilization of user profiles or user models to support personalized and adaptive features [42,43,46]. Similarly, only two studies directly assessed the personalization features [38,47]. The effects of the chosen personalization methods (either implicit, explicit, or a mix of the two) on user engagement and health outcomes received little attention.

While personalization of content to be delivered was common across all the studies, personalization of conversational style was implemented by only two studies [38,41]. The lack of conversational adaptation can be an impediment to improving usability and user experience, since different users may have different conversational preferences and needs that require different conversational strategies to be applied. Previous research has shown that adaptive conversational strategies can improve system performance, usability, and efficiency [58,59]. In this review, examples of conversational adaptation included using shorter questions for follow-up sessions [41] or using didactic, relational, or motivational conversational styles according to the user models [38]. Although such adaptive behaviors are useful steps towards accommodating the needs of various users, there are advanced implementations of conversational adaptation that can be applied to health care CAs, such as implicitly detecting users’ level of expertise and thus adjusting the complexity of the terms used and the dialogue path to be taken [60], or configuring the level of system initiative and confirmation strategies when a user faces difficulties in performing a task [12].

Only two studies evaluated personalization as a distinct factor [38,47]. The direct assessment of personalization, involving how users perceive the extent of personalization, is an important element in the evaluation processes. When the effects of personalization were evaluated, it was not possible to determine whether any of the outcomes were attributed to the availability of personalization features or to other factors. Therefore, new conversational agent studies with carefully controlled conditions are needed to understand the relationships between personalization features and other evaluation factors such as user satisfaction and user engagement. To guide the direct assessment of personalization, a theoretical framework such as the one developed by Fan and Poole [7] may prove useful for systematically considering various dimensions of the personalization process.

The implications of different implementations of personalization were not addressed by any studies. For example, a recent research paper drew attention to the limitations of implicit and explicit personalization [61]: while implicit personalization with its often-imperceptible user models and hidden assumptions can result in biased decision-making [62], over-reliance on system suggestions [63], and filter bubbles [64], explicit personalization may involve very formulaic and superficial choices for users who may not be well-equipped to customize the presented choices in a satisfactory manner [65]. The study employed a reflective personalization approach, allowing users to reflect on their own goals and priorities when making or modifying choices [61]. This approach demonstrated an implementation of personalization that recognizes the complexities associated with human choices, preferences, and agency when using interactive technologies. Out of all the studies in our review, one study implemented a reflective approach to personalization by using adaptive mini-dialogues to support users’ self-reflections on their goals [46]. These dialogues were successful in supporting discussions on awareness, goal accomplishment, self-tracking data, and trends in behavior.

Using CAs with unconstrained natural language input can be risky [66]. Thus, it is important for such CAs to include patient harm considerations into their study protocols. None of the included studies reported any personalization-related harms., but there was a lack of attention to the safety implications of CAs, as evident in the absence of patient safety as an evaluation dimension. In addition to patient safety, future studies need to consider the effects of different personalization methods on patient privacy. In particular, implicit methods used for gathering user information need to be clearly communicated to the users, since such methods often run automatically in the background, not being noticed by users. To this end, the model of informed consent for information systems may prove useful for considering various factors involved in collecting personal information [67].

Overall, most of the reviewed papers did not focus explicitly on personalization. Little attention was generally paid to the complexities associated with implementing personalization features and measuring their effects.

In line with our study, a recent scoping review of psychology-focused embodied conversational agents reported that only a few studies employed user models to personalize user-system interactions [68]. Another recent mapping study on health chatbots for behavior change found personalization to be one of the most appreciated technical enablers [69]. In terms of the implementation of personalization features, most of the studies in our review implemented personalization features without being informed by the advancements in other domains of personalization (eg, more automated personalization methods [12,60] or the implications of personalization on privacy, safety, and decision-making [61,62,70]).

Our results are based on the presence of personalization features of health care CAs in the studies that do not necessarily have an explicit focus on personalization. Therefore, the results are limited by the extent to which the included studies reported on their personalization features. In addition, our review focused on CAs using unconstrained natural language input. Therefore, the results may not be extended to agents using constrained natural language input (eg, multiple-choice of utterance options). Since the conversational systems used in the reviewed studies involved multiple components, the reported outcomes were attributable to the systems rather than only the personalization features. Our paper recommended using a theoretical framework of personalization to support a more systematic treatment of personalization features. However, it may be possible to implement personalization features effectively with no theoretical support. Moreover, other theories not specific to personalization may prove useful for personalization purposes, such as the Theory of Planned Behavior [71]. Various contextual factors such as location and time may also be integrated with user models to support more adaptive information and services [72].

Future research can focus on incorporating a theoretical framework [7] and an evidence-based approach to implement personalization features in the domain of health care CAs. Another line of research could investigate the relationships between personalization features in conversational systems and health processes, and outcome measures such as treatment adherence or management of chronic health conditions. Future work can also focus on the use of the unique characteristics of the conversational medium for personalization purposes, such as capturing prosodic features in users’ speech to automatically detect changes in mood or speech pathologies and thus provide adaptive information and services.

The use of personalization in health care CAs with unconstrained natural language interfaces has been limited and is not evidence based. While the CAs with personalization features were reported to improve user satisfaction, user engagement, and dialogue quality, little evaluation was performed to measure the extent of personalization and its role in improving health outcomes. Future research in health care CAs could evaluate the impact of personalization on health outcomes and its potential implications on privacy, safety, and decision-making.