We searched 8 databases (IEEE, ACM, Springer, ScienceDirect, Embase, MEDLINE, PsycINFO, and Google Scholar) to collect studies relevant to the topic. For Google Scholar, we only used the first 100 results from each search string, as Google Scholar returns the most relevant results belonging to each search query first. The search was conducted between September 9, 2019 and September 13, 2019. For the forward reference list checking, we used the cited-by functionality of Google scholar. We also checked the reference list of the included studies to review the backward reference list.

We used 2 different sets of search terms to search the databases. The search term set was decided based on the type of studies indexed by the database. For databases that mainly indexed studies in the medical field (Embase, MEDLINE, PsycINFO), we relied on keywords that are strictly related to the intervention (eg, chatbot, chatterbot, conversational agent, conversational bot). For databases with no specific application domain, we resorted to using keywords that are related to the medical domain (eg, health, medical, illness, disease, disability) in addition to the intervention-related words. The search strategy used for searching the databases is presented in Multimedia Appendix 2.

The chatbots in the included studies consisted of 4 main modules: text understanding module, dialog manager, text generation module, and database layer that holds the various types of information needed for chatbot training and function. Figure 2 shows a high-level architecture of chatbot development and the relationship between the different modules that constitute it. The following subsections highlight the findings in terms of how each of the previously mentioned components is implemented.

The text understanding module is the module with which the user directly interacts. The function of the text understanding module is to extract meaning from the user input before a specific answer is generated. In the included studies, 7 methods were used for text understanding. Pattern matching methods were the most common, used by 18 of the 45 chatbots (40%). Many text understanding approaches can fall under pattern matching. The different pattern matching methods reported in the included studies are as follows: keyword matching or string matching, where specific keywords or strings of the user input are matched to scripts stored in the database, which was used in 11 studies [10-15,32,33,46,48,50], or Extensible Markup Language (XML) variants, such as Artificial Intelligence Markup Language (AIML), which were used in 7 studies [16,17,34-37,47]. The second most popular approach was the use of machine learning methods for text understanding, which was used in 6 studies. Most of these studies used supervised machine learning algorithms, including decision trees and random forests. These include groups of decision trees, where each tree gives a classification to the text, in the form of voting, while the forest chooses the classification having the most votes over all the trees in the forest. Machine learning training materials were sometimes based on real health data extracted and anonymized from hospital information systems and electronic health records. In other instances, chatbots were trained on billions of words extracted from Google News datasets [18,19,38,39,49,51]. Another 6 studies relied on web services such as Dialogflow by Google. Web services are usually provided through a computer server that responds to requests through ports over a network, such as the internet, to solve specific domain problems over the web. They can also use a mixture of machine learning and rule-based methods to produce a customizable chatbot implementation. Three studies used natural language processing–based approaches, such as named entity recognition, to extract meaning from user input. To develop the chatbot, 2 of the studies used a rule-based approach, where the chatbot is operated by either a set of IF-THEN rules or a state-based system. A few of the studies relied on a fixed input method, where the user selects the input from a list of possible inputs. Fixed input chatbots do not technically need to go through a text understanding module as the system does not need to interpret the user input. The input needs only to be directed to the dialogue manager. Some studies proposed a few hybrid text understanding approaches. Hybrid approaches use a combination of 2 or more of the previously mentioned methods for text understanding. Table 2 summarizes the different text understanding methods proposed and the studies that fall under each method.

In the reported studies, the input to the dialog management module is the processed user free-text input provided by the text understanding module. The dialog management module controls the different aspects of the conversation and links each user input to an appropriate output. In the included studies, 2 main types of dialog management techniques were reported: static dialog management and dynamic dialog management. In the case of static dialog management, user input is matched directly to the appropriate output using a pattern matching algorithm (25/45) or using a set of rules (7/45). In the case of dynamic dialogue management, the context of the conversation changes based on specific user input characteristics. The context switching can be done by training machine learning algorithms to identify the context from the user input (5/45) or using a web service for intent identification (4/45). Table 3 summarizes the different dialogue management methods and the studies that correspond to each method.

Most of the chatbots in the included studies contained one or more of the 3 data repository types identified. Most of the included studies kept a medical knowledge repository (28/45, 62%). The medical knowledge repository contains medical information related to the application domain of the developed chatbot. The medical knowledge source can be local, such as obtaining it from medical personnel, or it can be collected from online sources such as Wikipedia or other websites. The medical knowledge can be presented to the user in the context of educational chatbots (17/45) [12-14,19,23-25,27,32,34, 41,46,48-51,53], or it can be used to train machine learning algorithms (5/45) [39,40,43,44,51]. Many of the developed approaches store users’ data and use the data to customize the chatbot response and improve its functionality (11/45). Conversation scripts are the third-data repository type, and they are usually kept by chatbots that use pattern matching as a text understanding or text generation modality (5/45). A few studies did not mention the type of data stored (9/45). Table 4 summarizes the different database types and studies that reported keeping each database type.

The text generation module provides output to the user. Text generation in the included studies was done using one of 2 methods: fixed output or generated output. The fixed output methods search the database for the most appropriate output to a user input and present it to the user. The generated output method relies on machine learning to generate original natural language output that is produced by the machine learning algorithm. Chatbots in most of the included studies provide fixed output that is extracted from the database with the exception of those in [25,39,51], which provide output that is generated using machine learning and deep learning methods, and [43], which proposes a hybrid approach that can provide fixed and generated outputs. A few studies did not report the output generation method, or it was not applicable as the output modality was not text [23,24,41,42,45].

This review focused generally on chatbots in the medical field, without specifying the field of application, which makes it more comprehensive than previously conducted reviews [55]. Other reviews that included different medical applications [56] presented a general taxonomy of conversational agents, while we presented a more granular description of the development techniques of each component. Searching more libraries from different application domains allowed us to include more chatbots in our study than that of Montenegro et al [56]. This is why we excluded Wizard of Oz studies, while they were used in the previous reviews. Two similar systematic review studies were conducted earlier in 2018 and 2019 [57,58]. The first study reviewed the applications and evaluation measures of chatbots and conversational agents, while the second paper provided a critical review of the tasks involved in natural language understanding and machine learning of chatbot systems used in the medical domain. Neither of the 2 papers discussed the technical aspects and development methodologies of the chatbots used in the medical domain.

This review only focused on text-based chatbot applications, where either the input or output modality is written. This excludes studies where the input or output modalities are spoken or visual, as well as robotics and telephone-based methods. This choice was made because we wanted to focus on text processing techniques rather than image or voice processing, as speech-to-text technologies can also introduce errors and another layer of complexity to chatbot development.

We enforced some constraints on the type of publications that were included in the current review. These constraints might have led to missing a portion of developed chatbots that have been published in other research venues, such as workshops, book chapters, and conference abstracts. Furthermore, limiting the search to papers published in English could also have led to missing some chatbots that were developed for communication in other languages and published in their own language. For example, we did not include papers published in Chinese or Arabic that discuss chatbots communicating in these languages.

This review focused on the development process of chatbots without considering the impact of these methods on patients. For this reason, some of the implementations in some of the included studies might be feasible from a technical point of view, but this does not necessarily mean they are effective from a medical point of view.

This paper reports the technical aspects of developing chatbots in the medical field. This review can be used to identify the most common development approaches by specialists to help them narrow down their options and make a decision on which development approach is the most appropriate for their applications.

The reported results show that most of the developed chatbots communicate with users in the English language. While a few attempts to design and develop chatbots in other languages exist, more work needs to be done in this regard, especially for languages that are spoken by a large portion of the world’s population, such as Chinese and Arabic.

Even though dynamic dialogue management provides a more natural user experience, most developed systems rely on static dialogue management methods. Changing the dialogue context based on user emotions or by detecting topic changes in user input are important aspects to be considered in chatbot development.

The use of machine learning and artificial intelligence methods in the development of conversational agents in different application areas has recently increased. The rate of adoption of machine learning–based methods in developing chatbots is still relatively low, even though it has been increasing in recent years. Supervised machine learning algorithms seem to better suit the development of chatbots for special medical conditions and diseases, while rule-based methods are being used more for developing chatbots used for general medical purposes. Machine learning methods allow the development of more intelligent agents that can provide a more realistic user experience, by providing a better text understanding experience, including more dynamic and flexible dialogue management, and generating a wider range of responses.

As the purpose of this review was to survey the technical aspects of chatbot development, the clinical results of performing clinical trials were not considered. Further reviews linking the different development techniques used to the clinical outcomes of the chatbot developed are possible and recommended.

More openness and a wider adoption of state-of-the-art methods in dialogue management [4], text understanding [5], and text generation [6] methods in the literature can really benefit the development of conversational agents in the medical field.

It is worth noting that the technical aspects of developing chatbots were not always clearly mentioned in the studies. The devised architecture is a general one that does not necessarily apply to every developed chatbot. One or more component might be omitted, and the chatbot might still function properly.