Development of a Chatbot for Pregnant Women on a Posyandu Application in Indonesia: From Qualitative Approach to Decision Tree Method

Abstract

With the widespread application of digital healthcare, mobile health (mHealth) services are also developing in maternal and child health, primarily through community-based services, such as Posyandu in Indonesia. Patients need media for consultation and decision-making, while health workers are constrained in responding quickly. This study aimed to obtain information from pregnant women and midwives in developing a decision tree model as material for building a semi-automated chatbot. Using an exploratory qualitative approach, semi-structured interviews were conducted through focus group discussions (FGD) with pregnant women (n = 10) and midwives (n = 12) in March 2022. The results showed 38 codes, 15 categories, and 7 subthemes that generated 3 major themes: maternal health education, information on maternal health services, and health monitoring. The decision tree method was applied from these themes based on the needs of users, evidence, and expert sources to ensure quality. In summary, the need to use a semi-automated chatbot can be applied to education about maternal health and monitoring, where severe cases should be provided with non-automated communication with midwives. Applying the decision tree method ensured quality content, supported a clinical decision, and assisted in early detection. Furthermore, future research needs to measure user evaluation.

1. Introduction

The internet and smartphones are increasingly popular both as means of accessing healthcare information and as tools for healthcare management [1,2]. Mobile health (mHealth) is defined as the use of mobile phones to support the achievement of health objectives or to deliver versatile and functional healthcare services [3,4]. In Indonesia, mobile health utilization has increased as internet users have grown from 66.2 million to 74.55 million between 2016 and 2020 [5]. With the widespread application (app) of digital healthcare, mHealth services also thrive in maternal and child health [6]. For example, a recent systematic literature review indicates that pregnant women need information access using mHealth to increase their knowledge and their adherence to the regularity of antenatal care [7]. Moreover, previous research also showed the high need for accurate information to support health decision-making among pregnant women [6,8].

Given that the priority agenda of the Indonesian government is reducing the high Maternal Mortality Rate (MMR) of 305/100,000 live births [9], strengthening primary health services through information, communication, and technology (ICT) innovative strategy is one of the efforts to solve this problem [10]. One such effort involves community-based activities, e.g., an integrated health service center/pos pelayanan terpadu (Posyandu), a socio-cultural embodiment of the primary health care approach [9,11]. In this case, one of the Posyandu-based digital health interventions has been realized through ‘iPosyandu’, an android-based mobile app by Rinawan et al., which is currently being extended in its features [12,13,14]. The previous version was mainly used for community health workers (CHWs) to store and manage Posyandu data. It also has a feature to share information about child development and maternal health. This app aims to build a “bridge” between the community and the national level [12]. The extended version is intended to optimize family services focusing on pregnant women.

The need for communication is still increasing among patients, especially in having remote consultations with health workers to make decisions [15]. To optimize the service, health workers need help responding promptly to patients [16]. In this case, Rinawan et al. added chatbots for educational purposes on the extended iPosyandu app, focusing on pregnant women in Indonesia. A chatbot, known as an intelligent bot, interactive agent, digital assistant, or artificial conversational entity, is a form of Artificial Intelligence (AI) used for simulating human conversations [17]. Chatbots can be a tool for providing health consultation and education [18]. Recently, chatbots have been widely used in healthcare for such purposes as improving mental health [19], promoting physical activities [20], supporting patients with chronic diseases [21], and education [22]. The reason is that chatbots can help patients manage their health and assist healthcare professionals in handling patient care, particularly decision-making [16,23]. Managing knowledge-based content containing a list of questions and answers in designing a chatbot is essential to support decision-making [24]. In this matter, content quality is considered in the development process to encourage pregnant women to adopt and use medical chatbots [25]. The chatbot’s knowledge system should be based on users’ needs and literature/evidence, and should then be compiled by health and informatics experts [26,27]. A recent systematic review pointed out that ensuring the content quality of a health chatbot can be realized using a decision tree method [28].

A decision tree is mainly used for classification and prediction using tree-shaped figures [29,30]. The practice of decision trees has been widely used in the health sector [31], such as for early detection of health conditions [32], predicting disease severity [30], determining telehealth services [33], and assisting admissions to intensive care units [34]. Overall, it is intended to help healthcare professionals make clinical decisions by maximizing effectiveness and minimizing harm [35]. It is also one of the most effective methods for developing chatbots in healthcare [24,30]. For example, in an AI system of a health chatbot, the decision tree method will be used in the back-end system for servicing a question from pregnant women through the chatbot [36,37]. From the front-end side, it can help users predict their health conditions using chatbots [38]. Thus, the accuracy of the information in this decision tree method can strengthen the quality of content in health chatbots [39]. A recent systematic review shows that the decision tree is one of the tools for consideration in assisting the patient/provider in decision-making. External validation is also needed to ensure that decisions are made with high-quality evidence [28]. In our app, the chatbot for pregnant women using a decision tree method was built by involving information on user needs and health experts as one of the strategies to support the decision-making.

Even though the decision tree chatbot is a form of AI designed to simulate human conversations in health education, it still has limitations [16,37,40]. Education using a chatbot is acceptable to the patient [41,42], but many healthcare professionals and experts have emphasized that fully-automated chatbots are not mature enough to diagnose a patient’s condition or replace the judgment of a healthcare professional [16,19,40,41,42]. These findings highlight the perception that chatbot technology can be helpful for promptly responding in less complex roles. On one side, it frequently affects self-diagnosed patients [16,23]. On the other side, it must be carefully considered when providing a fully automatic chatbot focusing on health education instead of self-diagnosis. For this case, a semi-automated chatbot approach with both automation and human conversation can be a strategy to overcome this problem [37,43]. The importance of combining automated chatbots and non-automatic chats directed to health workers is to provide rapid health education about patient problems that can be validated by health workers [16,40,41]. A recent systematic review identifies limited studies examining semi-automated chatbots, mainly for handling pregnant women’s health [44]. Therefore, this research aims to explore the need of pregnant women and midwives to develop a semi-automated chatbot through a user-centered approach, applying a decision tree method that increases the accuracy of clinical decisions among pregnant women and assists in early detection of their health problems [25].

2. Materials and Methods

2.1. Study Design

An exploratory qualitative approach was used to explore the needs of pregnant women and midwives in developing a semi-automated chatbot on the extended iPosyandu application. A focus group discussion (FGD) through semi-structured face-to-face interviews was conducted, with the topic guide designed based on a review of the existing literature that included chatbot competencies for online healthcare [18]. The qualitative data were used to apply the decision tree method in developing semi-automated chatbot mockups. The stages in this research process can be depicted in Figure 1.

2.2. Participant Selection

Sampling was determined using a purposive sampling method, where the selection of participants was based on the researchers’ judgment about which potential participants would be most informative, such as midwives and pregnant women [45]. Collecting qualitative data was achieved through face-to-face FDGs conducted with 22 informants: pregnant women (n = 12) and midwives (n = 10). The number of participants per focus group ranged from 3 to 21, with a median of 10 [46].

The eligibility criteria of pregnant women respondents consisted of women in all trimesters who can operate smartphones and be willing to participate in the interview. Required midwife informants were working in the areas of the three public healthcare centers/Pusat kesehatan masyarakat (Puskesmas) and had attended iPosyandu training. Pregnant women with complications and mental disorders were excluded from this study. Excluded midwives were midwives known to have not carried out their duties for more than the last month. This means these midwives were not actively working and did not follow the dissemination and training about the iPosyandu app. No participants refused this interview. For each participant, the first interview began with questions about their socio-demographic data. Pregnant women’s characteristics are listed in

Table 1. Characteristics of Pregnant Women Respondents.

Code	Age	Trimester	Parity	Occupation	Educational Level
RI1	22	3	Multiparous	Housewife	Junior High School
RI2	27	3	Primiparous	Housewife	Senior High School
RI3	24	3	Primiparous	Housewife	Senior High School
RI4	23	2	Primiparous	Housewife	Senior High School
RI5	33	1	Multiparous	Housewife	Junior High School
RI6	30	3	Multiparous	Housewife	Elementary School
RI7	28	1	Multiparous	Housewife	Senior High School
RI8	27	2	Multiparous	Housewife	Senior High School
RI9	24	1	Primiparous	Housewife	Senior High School
RI10	30	2	Multiparous	Housewife	Senior High School

, describing age; trimester; parity, which was divided into the categories of primiparous (women who had given birth once) versus multiparous (more than once) [47]; and educational level. Midwives’ characteristics are listed in

Table 2. Characteristics of Midwife Respondents.

Code	Age	Midwifery Education	Participated in iPosyandu Training
RB1	32	Bachelor Degree	Had attended training
RB2	31	Diploma Degree	Had attended training
RB3	33	Diploma Degree	Had attended training
RB4	34	Diploma Degree	Had attended training
RB5	32	Diploma Degree	Had attended training
RB6	36	Diploma Degree	Had attended training
RB7	36	Diploma Degree	Had attended training
RB8	32	Bachelor Degree	Had attended training
RB9	32	Diploma Degree	Had attended training
RB10	33	Diploma Degree	Had attended training
RB11	34	Diploma Degree	Had attended training
RB12	35	Diploma Degree	Had attended training

, including age, educational level, and participation in iPosyandu app usage training.

2.3. Setting

Data were collected in March 2022 and taken at 3 selected Puskesmas out of 20 Puskesmas in Purwakarta Regency, West Java Province, Indonesia. The selection of the Puskesmas followed the inclusion criteria in which health workers and CHWs had received dissemination and training on the app use. They also actively used it in the study areas (the Purwakarta, Koncara, and Pasawahan Puskesmas).

2.4. Interview Guide

The interview guide was compiled based on the content of conversational agents in previous research [18], including (1) treatment and monitoring, which is the implementation of treatment, management, support, and monitoring in health service; (2) health care service support, which includes connecting patients to health care services; (3) education, is providing information related to health care; (4) lifestyle behavior change, which supports users in overcoming various modifiable health risk factors, and (5) diagnosis, which identifies the nature of the disease or condition. First, we asked about the familiarity of informants with using chatbots on various platforms. Then, we proceeded by asking about the needs of pregnant women and midwife users regarding five chatbot functions, including chatbot for treatment and monitoring, health care service support, education, lifestyle behavior change, and diagnosis. The interview guide can be seen in Appendix A. Interviews lasted for 30–60 min and were audio recorded, and field notes were used to observe the data in the field. After the data reached data saturation, they were transcribed, and the researchers performed a member check with the informants. The theme references chatbot content or answers prepared as chatbot-based AI conversation materials. The structure of the interview can be seen in Figure 2.

2.5. Data Analysis and Decision Tree Method

Interviews were transcribed verbatim with the thematic analysis used to understand experiences, thoughts, or behaviors across a dataset [48]. The NVivo Release 1.6.1 (1137) software QSR International, Burlington, Massachusetts, United States of America tool was used to organize, obtain an overview of the data, and compile the coding. The analysis began with reading field notes and transcripts to receive a general sense of each transcript. Next, researchers read the transcription results repeatedly to understand the content clearly, and identified significant statements from the transcript to find codes. The codes were then grouped to represent categories as the basis for generating a small number of themes. The authors checked for the correctness of these processes and consistency. Subsequently, those categories were used to develop sub-themes and grouped into themes representing the main findings [49]. Then, the authors combined all the sub-themes and constructed meanings to describe the merged themes. Finally, the authors worked together iteratively to ensure that the narrative was representative of the data, which helped establish a credible interpretation.

The codes, categories, sub-themes and themes were then compiled into a conceptual map that shows the flow of ideas in the “findings” section. The flow can represent the theme from a more general picture to a more specific picture [49]. Concept maps were created alongside qualitative data analysis to capture all thoughts visually. In this study, the authors created a concept map using five different-colored markings to facilitate the process of generating the main findings. First, the authors determined codes marked in blue boxes representing the material for the question-and-answer items needed in developing semi-automated chatbots. Secondly, collected codes were grouped into categories marked with pink oval shapes representing the topics that would be the basis for the principal finding. In the third step, the writer connected the categories into sub-themes marked by orange boxes, then analyzed and merged them into the main findings or major themes. These themes became material for the development of the main topic. The illustration of the conceptual maps can be seen in Figure 3.

Afterward, a decision tree was built by synthesis of the qualitative data, collecting evidence and literature to strengthen the knowledge base system for creating a chatbot. The decision tree we made was in the form of questions with yes or no answers. It aimed to screen general conditions, where further complaints would be directed to the final decision to chat directly with a midwife. The function of the decision tree was also the most straightforward and feasible modeling for interventions in brief and mild conditions but less appropriate for addressing severe and long-term conditions [50]. The decision tree was generally constructed with two types of nodes: decision and terminal. A simple illustrative decision tree is presented in Figure 4.

In Figure 4, the tree is a hierarchical rule directed from top to bottom, starting with a decision node (green diamond shape) as the root or initial decision. Then, each management strategy is followed by yes or no choices, representing certain events. Finally, the endpoints of decision trees are characterized by a leaf as a terminal node (parallelogram shape) consisting of blue for the education needed on general conditions and red as the option of talking to a midwife, representing severe conditions at the end of the tree. The outcome measures are generally attached to these endpoints [50]. Subsequently, the authors discussed this arrangement in realizing the creation of a chatbot.

Two main elements of AI: a knowledge base and an inference engine, shown in Figure 5, were performed to develop the chatbot. The knowledge base contained facts, theories, thoughts, and relationships [24]. The knowledge base was responsible for summarizing the system’s intelligence, consisting of keywords/phrases and responses that must be associated with each sentence. Furthermore, the chatbot design would be implemented using a chat script approach. This approach defined topics, objects (concepts), and rules. In addition, there was an inference engine that could draw conclusions based on knowledge.

3. Results

The results obtained comprised 38 codes, 15 categories, 7 sub-themes, and 3 major themes, (1) Maternal Health Education, (2) Information on Maternal Health Services and (3) Maternal Health Monitoring. The codes in the qualitative results were then compiled into a conceptual map showing the flow of ideas in the findings section [49], shown in Figure 6.

Figure 6 shows the conceptual map as synthesized by the qualitative data. First, the researcher determined the coding, marked with blue boxes. This coding became the basis for categorizing, marked with a pink oval shape. Second, these categories were then grouped into sub-themes marked with orange boxes. Subsequently, they were merged into three major themes marked with green boxes as the main topics needed in the semi-automated chatbot.

3.1. Maternal Health Education

Chatbot-based semi-automatic teleconsultation could be one form of media for maternal health education. The semi-automated interaction education requires the design of stored and adapted answers to the information needs of pregnant women and midwives. The interviews showed that this education consisted of physical and psychological complaints and lifestyle changes. Those were critical educational topics and could be accessed quickly using automated chatbots.

3.1.1. Physical Education

The number of physical complaints experienced during pregnancy was one of the reasons for the importance of education and information about valid and fast answers. Semi-automated chatbot-based teleconsultation was considered a medium for health education, offering an immediate response to physical complaints experienced by mothers. Information on physical complaints summarized from the interview results included the danger signs of pregnancy, discomfort in pregnancy, and labor symptoms.

“…about complaints of nausea, vomiting and current condition … Yes, so we can consult quickly”

(Pregnant Women’s FGD)

In addition to opinions from pregnant women, midwives also recommended providing education related to general physical complaints and early detection of labor signs.

“General danger signs, nausea and vomiting, signs of bleeding or heartburn maybe yes, dizziness, stomach pains like that. How to drink iron tablet (Fe), vitamins, all shown.”

(Midwives’ FGD)

“Many people complain of back pain, I said because the baby is getting bigger and want to give birth”

(Midwives’ FGD)

3.1.2. Psychological Education

Psychological complaints were one of the essential topics for pregnant women in preparation for childbirth and readiness to become a mother. According to informants, complaints about the mother’s mental condition during pregnancy could be served and supported through educational chatbots by providing answers automatically. The reason was that the mental preparation of pregnant women for childbirth needs to be carefully prepared. Mental health screening was also necessary. The informant hoped that chatbots could be a medium to overcome the mental health of pregnant women quickly.

“Because we are also often given mental screening for pregnant women in the health care programs, such as filling questionnaires like do you feel, for example, that you have not slept for three days or not? I think it’s important to give mental education by chatbot”

(Midwives’ FGD)

“Ee.. that’s not all. The mental preparation before giving birth needs to be supported and educated”

(Pregnant Women’s FGD)

3.1.3. Lifestyle Change Education

Changes in lifestyle needed in the discussion topics through this semi-automated chatbot service include information about daily activities, husband–wife relationships, changes in women’s appearance, changes in nutritional patterns, and personal hygiene.

“The way husband and wife may have sex in the first pregnancy. How many months or how many weeks is the inspection? Eating patterns, resting patterns, sleeping patterns”

(Midwives FGD)

Some patients said pregnant women’s diet changes were significant as educational materials through chatbots. Besides, the lack of awareness of mothers in changing their appearance was also essential information conveyed. Good appearance changes supported and facilitated the pregnancy process.

“Avoiding unhealthy foods.. Not staying up late, sleeping regularly”

(Midwives’ FGD)

“Yes, how do you do it every day to stay healthy. Dietary habit. How do you keep it clean.”

(Pregnant Women FGD)

“Clothes. Some pregnant women still wear little pants. The pants are strict jeans normally. Then sandals are usually high heels for pregnant women”

(Midwives’ FGD)

3.2. Information on Maternal Health Service

One of the topics frequently asked by pregnant women is that of maternal health services. According to pregnant women and midwives, this topic comprises general information that can be quickly responded to and does not need to wait for validation from health workers to provide it via an automated chatbot.

3.2.1. Mother’s Health Checkup

It consisted of general and supporting health checks, such as the schedule of visits, the type of service (pregnancy, family planning, repeat visits), and the number of controls.

“Schedule an appointment or consul meet with the midwife”

(Pregnant women’s FGD)

“Pregnant women do not schedule repeat visits once a day, for example, once a month. Trimester from this to this. For those who are 8 months and over 2 weeks, once. So someone asked later to check again when yes. Later, if I control it again, I’m afraid I’ll forget again. But usually, if it’s like you’ve been to the health center, I’ll let you know the return visit”

(Midwives’ FGD)

3.2.2. Mother’s Service Support

Teleconsultation media should provide information on health insurance, birthplace, and health workers’ recommendations to prepare for delivery in a health facility. According to pregnant women and midwives, information on health services during pregnancy was needed because there were still pregnant women who did not give birth to health workers. Supporting examinations also needed information given to pregnant women to increase knowledge on early detection.

“Lab checks. Ultrasound, screening”

(Midwives’ FGD)

There was still trust in ‘paraji’ or traditional birth attendants, encouraging women to give birth with paraji instead of in a health facility. The reason was the hesitation of the cost of giving birth in health facilities. According to the midwife’s interview, information about the importance of giving birth to health workers, financing, and health insurance could be answered automatically through chatbots.

“There are also deliveries (by parajis), so maybe the education is lacking, yes, the education is like that. Ee..yes, the economy is too. The average is the economy. There are still some, but not many. It’s just… there are. The second first child is the same as the paraji, huh”

(Midwives’ FGD)

Health insurance information is a solution in providing awareness about health care support that mothers can access.

“Sometimes, people who are economically dependent are thinking about costs. At least we should be educated about BPJS (national health insurance) problems from the start”

(Midwives’ FGD)

3.3. Health Monitoring

3.3.1. General Health Monitoring

Chatbots were recommended to be able to answer complaints of a general condition. This aimed to avoid self-diagnosis, which may lead to misdiagnoses. Thus, the available monitoring information could be a solution to increase knowledge and support early detection.

“You can remind us that there are several programs, which are like a class for pregnant women, can you not include them so they can join. Even though we don’t…”

(Pregnant Women’s FGD)

The midwife recommended helping to answer general complaints with a chatbot, but health workers must examine further complaints.

“For example, from the pregnant woman, there is this complaint. For example, all complaints are accommodated by the chatbot. Answered by chat, how is it? Later, they will ask again, if the chatbot has not been able to accommodate the patient’s answer, maybe they can be directed to the midwife as soon as possible”.

(Midwives’ FGD)

3.3.2. Advanced Monitoring

Further management of pregnant women’s condition was required to be confirmed with health workers. Pregnant women considered getting information through a validated chatbot rather than looking for information on their own through the internet, whose source accuracy was unknown. Pregnant women preferred communicating with health professionals as experts to support decisions about their health conditions for severe cases.

“…about complaints… Yes, so we can consult with the person concerned, the health workers as well so that we know that we are not driven by google. So go straight to the experts…”

(Pregnant Women’s FGD)

The midwife argued that the complaints of pregnant women could not be automatically controlled by communication machines such as chatbots. Clinical decisions on further patient complaints should be handled directly by the midwife.

“Well, for example, in my opinion, everything should be accommodated first. If there are 10 patients, try to be answered by the chatbot first. Well, later, if the patient has been accommodated with the chatbot’s answer, thank God. If the patient still feels anxious and wants to ask further questions, it can be brought up to communicate with the midwife”

(Midwives’ FGD)

4. Development of the Chatbot

From the theme found in the qualitative data, the decision tree was developed to structure questions and answers along with decisions that will be realized on the chatbot. Figure 7 shows the chatbot development process from the initial process (knowledge engineering) to its user interface development. The chatbot consists of two main parts: (1) a user interface that deals directly with users and (2) an inference engine to seek answers to consultations or questions via chat. Software engineers built the first part to accommodate the needs of users (pregnant women) and stakeholders (midwives). These needs comprised three themes: (1) maternal health education, (2) health monitoring, and (3) information on maternal health services. Subsequently, when the user started a consultation using a chatbot, the inference engine sought answers from the consultation based on the knowledge acquired by the knowledge engineer. The knowledge was developed based on the latest literature and experts’ validation through the knowledge engineering process. The acquired knowledge was stored in a knowledge base with a decision tree structure.

The decision tree started with a conversation about physiological complaints and then continued with pathological complaints with chat directions for professionals. It aimed to help patients obtain information to fill their knowledge gaps before visiting health workers [23]. The following decision tree for building a semi-automated chatbot for pregnant women is presented in Figure 8.

From the decision tree method in Figure 8a–c, it can be seen that many codes were compiled, such as the Q codes, marked with a green oval shape as questions that are listed in Appendix B. Then, Code A, which means the answers are marked with a blue box, can be seen in Appendix C. From the decision tree method concept, the semi-automated chatbot is presented in Figure 9.

In Figure 9, the chatbot appears in the extended iPosyandu app in Bahasa Indonesia. Aside from screening general questions at the beginning, shown in Figure 9a, health education according to patients’ queries and conditions is shown in Figure 9b. Afterward, the conversation ended with the recommendation of options to chat or talk to the midwife to strengthen health decisions mentioned in Bahasa Indonesia, “Silahkan konsultasi ke bidan”, which means “Please consult the midwife” Figure 9c.

5. Discussion

Qualitative data exploration methods have benefits and have become the basis for determining needs based on human perspectives and experiences [51]. As the intervention is intended for users, it is necessary to involve the target in developing research. The reason is that a health intervention tool can appropriately answer the requirements in its implementation if it concerns the user’s need [52]. A user-centered approach has become an effective method to determine the success of the intervention field [26,53]. User perspective can also help provide insight to researchers, application designers, and health providers in developing and evaluating health applications [15,53,54,55]. Other research explained that user evaluation through quantitative analysis after conducting qualitative methods could be used to strengthen the success of health intervention development studies [54,55,56]. However, previous studies have shown that several health applications successfully developed with qualitative exploratory methods [57,58]. Therefore, using qualitative design to explore information with a user-centered approach can be critical in application development research [12,59,60].

This qualitative research demonstrated the needs of pregnant women in the chatbot’s contents through three major themes: (1) maternal health education, (2) information on maternal health services, and (3) health monitoring. The content developed is in the form of question and answer, applying the decision tree method. Since the content is one factor that attracts clients to chatbots [22], the quality of the content must be considered according to the user’s needs, evidence, and expert recommendations [28]. All themes of chatbot content in this research were based on the user’s perspective and were related to the function of a chatbot as a preventive tool to support self-care and decision-making [61,62].

The first theme is maternal health education. In this theme, a chatbot is expected to be able to provide education about physical and psychological complaints and lifestyle behavior changes. This implies that chatbot technology is a good innovation with the capabilities of improving teaching and learning and all other aspects of education [22,57,63]. Furthermore, a recent systematic literature review study stated that chatbots for education have highly regarded benefits due to the integration of contents, quick access to educational information, motivation, and engagement because of their attractiveness, allowing multiple users simultaneously and immediate assistance [22]. Chatbots for education can also help deal with the physical complaints of pregnant women through a simple and exciting question-and-answer system [25]. Consequently, developing a chatbot that provides treatment for pregnant women’s physical complaints can be a strategy to increase knowledge and support early detection of their health problems.

Regarding the maternal health education theme, psychological issues often need to be one of the topics since they can become a problem during pregnancy and after giving birth. A systematic review found that anxiety symptoms reported during pregnancy had a combined prevalence of 22.9% across trimesters [64]. Chatbots can efficiently deal with increasing mental health problems [19,65,66,67], especially in motivating and overcoming anxiety or depression [68,69,70]. Patients feel comfortable using chatbots since it can reduce the fear of stigma when conducting health consultations using chatbots compared to humans [25,66]. Apart from psychological issues, lifestyle behavior changes for pregnant women should also be a maternal health education topic using mHealth [71,72]. A woman’s behavior during pregnancy impacts her mother and baby [72]. Previous research identified that chatbots could manage healthy lifestyle changes [20,73,74]. In this instance, chatbots can be virtual coaches, promoting physical activities and increasing human competence for healthy lifestyle education [18,65].

The second theme is information on maternal health services, which consist of routine and supporting examinations, as essential knowledge during pregnancy. The role of chatbots in increasing access to health services is shown in the previous research. Chatbots do not participate in providing health services but are an entry point for using health services [43]. Chatbots can deliver administrative-related benefits, especially in scheduling doctor appointments, finding health clinics, providing reminders for medication adherence, providing medication instructions, and answering commonly asked medication questions [16,75]. The main objective is to increase the efficiency of health services by reducing the effort and increasing the speed of access [16,75]. For that reason, chatbot providers can support maternal services by connecting patients with healthcare professionals, finding medical drugs online, or providing healthcare customer service tasks.

The third theme needed is health monitoring, another essential function of the health chatbot. Previous research states that chatbots can monitor symptoms to encourage healthy behavior [65]. However, ethically, the chatbot is not considered to be mature enough to perform health diagnostics [40,42]. A recent systematic literature review indicated that patient safety concerning chatbot consultation services needed to be observed since it might not recognize or respond appropriately when questioned about a serious health concern [59]. Chatbots are also associated with the risk that patients may self-diagnose too often and not understand the diagnosis, or patients may not feel connected to their primary physician [23,42]. This technology cannot substitute creditable health information sources from a health professional [16,41,42,76]. The chatbot designed on the extended iPosyandu application is only used for general complaints, where further or particular complaints are directed to non-automated consultations with midwives (semi-automated).

A primary step in developing a chatbot is strengthening the knowledge base. Building a decision tree can help to create a semi-automated chatbot with prediction and classification inside the knowledge base [77]. Since the benefits of decision trees method in the health sector include predicting disease, early detection, analyzing disease severity, and clinical decision-making [30,32,35,78], this method is suitable to help midwives and pregnant women to make health decisions through chatbots [38,62]. However, previous research has stated that a challenge in the use of the decision tree is the problem of overfitting. Overfitting usually occurs when the tree has too many nodes relative to the amount of training data available in machine learning [79]. To avoid this challenge, the semi-automated chatbot development in this study used forward chaining, or an expert data-based system that works with a set of known facts and applies rules to generate new facts whose premise matches the known facts to achieve the previously determined goal [80,81]. Forward chaining is also a natural solution to some problems, starting with information gathering. The decision tree in this study was applied to strengthen the knowledge base by exploring the needs of patients and health workers and reviewing literature or evidence, whose content was then validated by experts [28]. This method’s effectiveness and accuracy can help support clinical decisions with various suggestions and reasons [34,35].

Strengths and Limitations

One of the strengths of this research is that there is still rarely study on the development of semi-automated chatbots (a combination of automated chatbots and health workers’ communication), particularly for pregnant women. Besides, this semi-automated chatbot was developed through a decision tree method involving experts and using a user-centered approach to support the clinical decision. This research also highlighted strengthening knowledge-based in the development process using forward chaining to support quality content. Another strength is that this study used a focus group discussion to explore various opinions to provide a broad perspective regarding user needs and enrich the content of the educational chatbot for pregnant women.

Nevertheless, this study also has limitations. The research method for developing a health intervention through the chatbot only used qualitative methods from the user’s perspective. However, previous studies on developing new applications used a mixed-method approach, from qualitative to quantitative, to strengthen the results [12,38]. For this reason, future research needs to measure users’ evaluation quantitatively.

6. Conclusions

Information needed by pregnant women and midwives from semi-automated chatbots includes maternal health education, maternal health services, and health monitoring. In health monitoring, an automated chatbot can provide general health screening. At the same time, for further complaints, patients are advised to direct chat or communicate with a midwife to avoid self-diagnosis (non-automatic). To ensure the quality of semi-automated chatbot development, the decision tree method, with the exploration of a user-centered approach, collecting literature, and expert recommendations, can be alternatives to support clinical decisions and assist in early detection for pregnant women and midwives.