To develop a chatbot for patients with lung cancer, we used a sequential mixed methods approach (Table 1) [9]. We adopted this approach because of the lack of sufficient nationwide data on the categories of frequently asked questions (FAQs) asked by patients with lung cancer and their caregivers. We also decided to adopt an initial qualitative approach to generate hypotheses. We adopted the lens of health care providers to provide palliative care. This study was conducted at a tertiary care hospital in Japan.

Descriptive statistics were used to summarize the results. We performed an unpaired t test for univariate analysis. We used Stata (version 16.1, Stata Corp) for statistical analysis.

The study protocol was approved by the institutional review board of the Hyogo Prefectural Amagasaki General Medical Center (number 1-31). We obtained individual consent from the participants through a web form.

We obtained 246 questions from 15 health care providers. We identified 5 major categories (home care: n=50, 20%, changes in health condition: n=54, 22%, prognosis: n=20, 8%, lifestyle: n=82, 33%, and medications: n=39, 16%), and 75 supplementary categories (Multimedia Appendix 1). We identified 7 additional categories by screening resources (Multimedia Appendix 2).

We formulated responses for 82 categories identified in phase 1. The responses contained brief information intended to teach patients and their caregivers to deal with symptoms and direct them to various weblinks to information resources that are maintained by the public budget.

Based on the FAQs identified in phase 2, we developed the first version of the chatbot and integrated it in the LINE social media network. Figure 2 shows a sample interaction shared with the chatbot.

A total of 14 medical staff participated in the α test. They asked 71 questions to the chatbot during this period. Of these, 11 (15%) questions did not match the appropriate categories. Therefore, we added 3 new categories. Some of the categories had corresponding responses because the questions included the names of products associated with drinking and smoking. Thus, we added keywords for the corresponding categories to match responses.

From among 309 patients who received either oral or intravenous chemotherapy at our hospital, 11 patients and 1 caregiver participated in the β test. Of them, 9 patients were aged <60 years, 1 was aged between 60 and 70 years, and 2 were aged >70 years. All of them used computers or smartphones daily. The participants were asked 60 questions, of which 8 (13%) did not match with the appropriate categories. Furthermore, 2 (3%) questions had corresponding responses, and 6 (10%) questions did not. Additionally, 3 (5%) questions based on daily conversation did not have corresponding responses (eg, “I played golf today in the morning”).

After the test, 7 participants responded to the postexperimental questionnaire. The mean satisfaction score was 2.7 (SD 0.5) points out of 5. We used two free feedback comments: “those who have a connection with a patient group, the answers were not too much to hope for” and “there were many times when I did not get a proper answer.” We added 3 categories based on the unmatched questions. Finally, 91 categories and responses were curated [14].

This is the first study to develop a chatbot and evaluate its early feasibility to improve the knowledge of symptom management among patients with lung cancer and their caregivers. Based on the experience of medical professionals, we developed an autoresponsive chatbot, which could respond to most of the FAQs. However, it was not adequately used, and user satisfaction was low.

The categories identified in this study could be used as a reference for patient education in future. Currently, only one chatbot is available to empower patients with cancer and their caregivers. This chatbot was developed only for patients with breast cancer [6,15,16]. Given that patients with breast cancer and those with lung cancer differ in their age of onset, chemotherapy regimens, and cancer-related symptoms [17], it is important to increase the number of categories to improve their knowledge of symptom management based on the FAQs identified in this study.

However, since the generation of the chatbot alone did not yield adequate questions from participants, the low response rate potentially indicates that our chatbot may not be acceptable in its current form. Further studies are required to activate the interaction between patients and chatbots. A previous systematic review reported that web-based interactions between physicians and patients or those among patients may improve the patients’ quality of life [18]. Thus, chatbots can be incorporated into web-based networks for increased interaction. Indeed, one of our patients involved in the β test (phase 5) preferred interactions with other patients rather than the chatbot. Therefore, further studies are required to generate a platform that initially forms a text messaging–based online support group involving medical professionals and patients and then gradually reduces the need for these professionals to respond to patient queries by involving chatbots instead.

This study has two major limitations of note. First, 8 (13%) questions did not match suitably with responses in phase 5, and a patient complained of not receiving appropriate responses. Although the ability of the chatbot to match questions with categories was not a major problem, questions for nonexistent categories remained unmatched. Therefore, it is necessary to add educational categories and responses through further discussion. Furthermore, the integration of existing chatbots for daily conversations may improve the proportion of matches [19]. Second, this was a single-center study, and further studies are needed to evaluate the applicability of our findings in other hospitals.

Medical staff who provide care to patients with lung cancer may be able to educate them using a chatbot containing the categories of FAQs curated in this study. Nonetheless, further studies are required to improve this chatbot in terms of interaction, especially considering its low usage in this study.