A mixed methods study was conducted from September to December 2022, adhering to the GRAMMS (Good Reporting of a Mixed Methods Study) guideline. Quantitative analysis alone could identify who used or did not use the chatbot but could not explain the underlying reasons for these behaviors. To understand the contextual factors influencing adoption and engagement, qualitative inquiry was incorporated. Quantitative surveys and user logs provided objective data on participant characteristics, chatbot use, and usability, whereas qualitative interviews explored reasons for use or nonuse, perceived benefits, and barriers to use. This study used a sequential mixed methods design, in which quantitative surveys were conducted first, followed by qualitative interviews.

Patients with cancer and their caregivers were recruited from the Cancer Education Center of Samsung Medical Center in Seoul, South Korea. All patients scheduled to undergo PICC insertion were required to attend a face-to-face education session prior to the procedure. We consecutively contacted and recruited potential participants following each education session. As some patients with a PICC line were not scheduled to receive chemotherapy, an oncology nurse screened eligible patients and their caregivers prior to recruitment. We used convenience sampling, and individuals who voluntarily agreed to participate were enrolled in the study. Those who were 19 years or older, scheduled for PICC insertion (patients only), owned a smartphone, and had ever used KakaoTalk (most popular instant message service in Korea) were eligible for the study. However, those with any cognitive or physical problems with using a smartphone and who were scheduled for PICC line removal within a month were excluded. Following the provision of written informed consent, participants received approximately 15‐20 minutes of training on how to access and use the chatbot. The session covered the basic user interface, methods for interacting with the chatbot, troubleshooting technical issues, and general precautions for use. Participants practiced using the chatbot under the guidance of research staff.

Participants were asked to voluntarily use a chatbot prototype to support PICC management for 1 month. The chatbot was a rule-based system that provided structured responses along prespecified dialog trees. It was technically implemented by Skelter Labs Corp (Figure 1) and integrated into KakaoTalk, a popular instant messenger in Korea [26]. The main purpose of the chatbot was to provide structured information regarding PICC management according to user inquiries. In the welcome message, the chatbot told the participants what the chatbot can and cannot do. The chatbot clarified that it has not been designed to replace clinical practice or make clinical decisions.

Users could ask the chatbot questions by tapping buttons listed according to prespecified dialog trees or typing their queries directly. Regarding the users’ free text, the chatbot was enabled to understand the intent based on natural language processing. Dialog trees were constructed according to 6 main themes: catheter management, adverse events with the catheter, coping with emergencies, symptom management, precautions during daily life, and heparin-related inquiries. These themes were determined by reviewing records of PICC consultations in which patients had previously asked questions. The most frequently asked questions for each theme were displayed in a carousel-type list, allowing users to navigate through the available inquiry options with minimal scrolling. The chatbot used the health information about PICC management developed by advanced practice nurses in charge of patient education for PICC management. Considering the role of chatbots as social actors, health information in the chatbot was delivered through “Seulgi” (“wise” in Korean), a virtual character with the appearance of a nurse and a friendly tone. However, some information too complicated to convey through text alone was delivered via photos or videos. In addition, for irrelevant or ambiguous inquiries, the chatbot delivered fallback messages that reminded users of the chatbot’s original purpose and highlighted the contact number of the cancer education center for additional help.

Chatbot feasibility was evaluated using triangulation with quantitative usability survey, user logs, and qualitative interviews. The investigated topics included questions such as who voluntarily used chatbots, why they used them, what the most frequently asked questions were, and how satisfied users were.

Descriptive statistics were used to compare characteristics between chatbot users and nonusers, identify the most frequently asked questions, and determine user satisfaction. Continuous variables were presented as means and SDs and categorical variables as frequencies and percentages. Group differences between users and nonusers were examined using independent t tests for continuous variables and chi-square tests for categorical variables. Graphical depictions were presented to show the main reasons for using (or not using) the chatbot. Regarding user satisfaction, all responses were dichotomized into “positive” (agree and strongly agree) and “negative” (all other responses) and summarized in frequencies. All quantitative analyses were explanatory in nature. Statistical tests were 2-sided, and statistical significance was defined using a 95% CI and a P<.05. Participants with missing values for a given variable were excluded from the corresponding analysis but retained in other analyses. Because the proportion of missing data was small, missingness was assumed to be missing completely at random, and no imputation was performed. All statistical analyses were performed using R (version 4.1.2; R Foundation for Statistical Computing).

For the qualitative data, we conducted thematic analyses based on Braun and Clarke’s method [28]. Three researchers (BJ, ML, MJK) independently read the transcripts, generated initial codes, and organized them into categories and themes related to participants’ motivations for chatbot use, perceived benefits, unmet needs, and contextual factors of PICC self-management. A shared coding framework and codebook were developed through iterative discussions. Using investigator triangulation, at least 3 researchers coded each transcript independently, compared results, and resolved discrepancies by consensus. The final themes were reviewed jointly to ensure coherence and distinction and by a principal investigator who audited the overall analytic process. Representative quotations were selected to illustrate key themes.

Quantitative and qualitative data were integrated during the interpretation phase using a narrative approach [29]. Qualitative findings contextualized quantitative results by explaining the reasons for chatbot use or nonuse. Integration discussions were conducted by the research team to ensure coherence between datasets.

The study was reviewed and approved by the Institutional Review Board of the Samsung Medical Center (IRB No. SMC 2022-07-059). Written informed consent was obtained from all participants. All collected data were deidentified, and access was restricted to authorized personnel. Participants received remuneration of 30,000 KRW (approximately US $25) upon completion of the study.

During the study, 5‐10 individuals received PICC-related education each day, and 78 individuals initially consented to participate. Among those, 56 who completed the study were included in the final analysis. Among them, 28/56 (50%) used chatbots (Table 1). Chatbot users were more likely to be younger (51.1 y vs 59.6 y; P=.02) and female (24/28, 86% vs 15/28, 54%; P=.02) compared with nonusers. The chatbot user group tended to have a higher educational level, but there was no statistical significance (19/28, 68% vs 13/28, 46%); P=.17). No significant differences were observed between chatbot users and nonusers in terms of the type of participants (patients vs caregivers), working status, family living together, comorbidities, income, and residence area. In the patient subpopulation, chatbot users tended to have breast cancer (13/17, 76% vs 10/20, 50%; P=.28), present with early-stage disease (I or II; 11/17, 65% vs 8/20, 40%; P=.18), and undergo curative chemotherapy (17/17, 100% vs 17/20, 85%; P=.29; Table 2).

Regarding the reasons for using the chatbot, 58.3% (n=14) were for inquiries about symptoms after PICC insertion, followed by learning how to manage daily life (n=12, 50%) and understanding how to deal with abnormalities at the insertion site (n=9, 37.5%; Figure 2). Additional reasons mentioned in the qualitative interviews were repetitive learning to acquire information and the management of anxiety about medical services at local hospitals (Multimedia Appendix 1).

Among the chatbot nonusers, 14/27 (52%) did not have any issues due to their PICC (Figure 2). Additionally, 10 nonusers mentioned that they did not know how to use the PICC (n=5) or that it was difficult to access (n=5). Additional reasons mentioned in the qualitative interviews were a lack of trust in chatbots and low expectations for chatbots (Multimedia Appendix 2).

A total of 25 users agreed to share their logs. During the study period, 347 inquiries were identified and categorized into 4 main themes: catheter care (126 observations), managing daily life (85 observations), symptoms (72 observations), and heparin use (55 observations; Table 3). Additionally, 9 cases were requests to contact the oncology nurse directly (Multimedia Appendix 3).

Regarding catheter care, the most frequently asked question was about the timing for changing dressings (22/126, 17.5%), followed by buying dressing materials (17/126, 13.5%) and taking other intravenous medications through the catheter (15/126, 11.9%). Regarding the management of daily life, workout-related inquiries (24/85, 28%) were the most common, with additional concerns including showers (14/85, 16%), travel (11/85, 13%), driving (11/85, 12.9%), and wearing clothes (11/85, 13%). In terms of symptoms following catheterization, fever (38/72, 53%) was the primary concern, followed by bleeding (10/72, 13.9%) and pain at the insertion site (10/72, 14%). The timing of the heparin injection (44/55, 80%) was the most frequent heparin-related inquiry. During the study period, there were 705 interactions with the chatbot, and we identified a total of 347 user inquiries. Of these, 100 were free-text inputs, and 69 (69%) were successfully recognized by the chatbot. The 31 failed cases occurred because users did not provide sufficient context in their inquiries, entering only single terms, such as “injection,” “not detachable,” or “heparin.” As the fallback scenario was designed with limited functionality, the chatbot could not prompt users to clarify vague inputs.

Among the 28 chatbot users, 23/28 (82%) completed the questionnaire, with 22/23 (96%) agreeing that the chatbot was easy to use and 21/23 (91%) mentioning that it was useful for managing their PICC catheter (Table 4). Furthermore, 17/23 (74%) felt like they were conversing with a person during interactions with the chatbot. All participants agreed that the chatbot’s responses were trustworthy, 21/23 (91%) expressed the intention to continue using the chatbot while the catheter was in place, and 22/23 (96%) were satisfied with using the chatbot. All participants (23/23, 100%) recommended the use of chatbots for other patients using catheters.

While users recognized the benefits of the chatbot, several limitations highlighted areas for improvement to enhance user experience and functionality. Through the qualitative analysis, 3 main themes emerged regarding the chatbot’s limitations: conversational issues, user experience issues, and lack of personalization.

The aim of this study was to evaluate the feasibility of a chatbot designed to support PICC self-management by examining chatbot use rate, usability, and qualitative satisfaction among patients with cancer and their caregivers, using data from survey, usage logs, and interview. Regarding chatbot adoption for PICC management, half of the participants used the chatbot during the study period. The most popular topic among chatbot users was PICC line management (eg, timing of dressing changes and purchasing dressing materials), followed by questions about adjusting to daily life, symptoms after PICC insertion, and heparinization. In terms of user experience, the 3 main benefits of the chatbot were information accessibility, effective guidance, and psychosocial support. Despite these promising findings, there were 3 key challenges related to the chatbot: conversational issues, user experience issues, and a lack of personalization.

In terms of chatbot adoption for PICC management, at least half of the participants used the chatbot during the study period. In previous studies, the adoption rates of health care chatbots ranged from 35% to 80% [30,31]. The heterogeneity in adoption rates across studies might depend on specific topics addressed by chatbots or the availability of alternative resources. For example, a study reported that 80% (n=48) of 60 patients with pancreatic cancer accessed the chatbot providing genetic information [30]. In another study examining a chatbot for chemotherapy care, 58% (n=20) of 34 patients with gastrointestinal cancer used the chatbot at least once [31]. Conversely, a chatbot for postoperative guidance after ureteroscopy was used by only 35% (n=7) of 20 patients [32]. Their lower activation rate may be attributed to the availability of multiple alternative information sources [32]. Compared with other studies, this study included both patients with cancer and their caregivers and provided nurse-led education for PICC management prior to offering the chatbot. These specific conditions might have contributed to participants experiencing a lower need to obtain PICC-related information via the chatbot. The primary motivation for chatbot utilization was related to individual needs for acquiring knowledge or information to prepare for or cope with PICC-related symptoms or issues. Notably, most nonusers reported experiencing mild or no PICC-related problems, diminishing their perceived need for the chatbot. Therefore, when assessing chatbot feasibility, it may be essential to consider whether potential users are experiencing the specific circumstances related to their perceived needs for health information.

The most popular topic among chatbot users was PICC line management (eg, timing of dressing changes and purchasing dressing materials), followed by questions about adjusting to daily life, symptoms after PICC insertion, and regular heparinization. A study on PICC management in nursing facilities reported communication gaps, such as uncertainty about the last dressing change date or PICC removal timing, even between hospitals and nursing facilities [33]. Another study on adults with a PICC at home found that showering was the most common challenge, followed by difficulties with exercising, sleeping, and dressing [34]. The chatbot in this study was designed and implemented based on patients’ information needs on PICC management in a real-world setting. The chatbot allowed users to access reliable information whenever they wanted. Although the questions and answers in the chatbot were prespecified, the topics the patients inquired about were comparable to the findings from previous studies. This suggests that chatbots could be feasible for providing informational support to patients with PICC lines. However, given the rule-based nature of chatbots, continuous expansion of the information the chatbot covers would be necessary to ensure sustainability.

In terms of the chatbot’s usability, users mentioned that it was an accessible channel for effective PICC self-management guidance. In this study, the chatbot was designed to provide practical tips or information regarding PICC self-management in a structured format with succinct and plain language. The implementation incorporated patient health literacy considerations, adhering to evidence-based recommendations for health care chatbot interfaces derived from previous studies [35,36]. Users particularly appreciated the convenience and rapid access to relevant information provided by chatbots, which may have enhanced their perceived usefulness in this study [17]. In addition, the chatbot was helpful for them to receive proper psychosocial support regarding PICC management. Patients with cancer, who face disease and treatment uncertainties, seek reliable information and justify their feelings through social support networks [37]. Moreover, PICC management is unfamiliar to most patients with cancer, requiring high competency in its long-term maintenance [38]. These features of patients with cancer can increase their anxiety regarding potential complications and coping strategies for cancer-related problems [38,39]. The chatbot in this study allowed users to access reliable information about PICC-related issues without temporal or spatial constraints, addressing unmet self-care needs and providing psychosocial support. During the COVID-19 pandemic, when hospital visits and communication with health care providers were often limited, the chatbot served as a clinically meaningful alternative by offering timely, trustworthy guidance. This support helped reduce patient anxiety and might have contributed to better treatment adherence. Furthermore, although the chatbot was designed to mimic human conversation, participants did not perceive its interaction as equivalent to real human dialogue. Future research should quantitatively evaluate these supportive functions to better understand their impact on patient outcomes.

Despite the promising findings, the chatbot encountered 3 main challenges: conversational issues, user experience issues, and a lack of personalization. A natural language understanding engine was integrated into the chatbot, but it occasionally failed to accurately interpret users’ intentions. The system was primarily optimized for button-click interactions rather than free-text interactions to facilitate an intuitive user experience. However, these rule-based, static responses seemed to diminish conversational naturalness and had limitations in capturing the various nuances of individual expressions for the same PICC-related issues. A systematic review found that rule-based health chatbots often limited user engagement owing to constrained input and predefined dialogues [19]. To resolve these unmet needs, incorporating large language models (LLMs), such as GPT or Claude, into rule-based systems would be necessary [40,41]. These models may provide better situational awareness and more human-like communication, which allow for personalized support in PICC self-management [41]. Given patients’ documented preference for direct clinician interaction, the strategic implementation of LLMs may be required for health care chatbots to approximate human-like communication efficacy [42]. Nevertheless, LLMs may generate plausible but inaccurate responses (“hallucinations”), necessitating rigorous reliability assessment and empirical validation prior to the clinical implementation of LLM-oriented health care chatbots [43]. Independent of preceding considerations, swipe-based navigation in the chatbot impeded user engagement, particularly among older users. Swipe-based navigation was difficult for them, and they often preferred icon-based menus to reduce cognitive load [44,45]. To assist user navigation, the implementation of subtle visual affordances indicating interface functionality and hierarchically structured list-based layouts is recommended [46].

First, although patients or their caregivers are required to attend PICC education, those who voluntarily consented to participate in this study may be more interested in using digital technologies or more motivated to self-manage. While this study did not collect demographic data of those who did not consent to participate, further understanding this in future research would provide more insight into scalable implementation. Second, user experiences were assessed after approximately 1 month of chatbot use. While PICCs are often maintained for 3-6 months, these findings primarily reflect the early phase of use rather than long-term engagement. Thus, the 1-month study period may not fully capture sustained engagement patterns. However, research has shown that many complications and difficulties tend to arise within the first few weeks following insertion, making this period particularly critical for patient education and support [47]. Hence, the data collection during this initial phase offers important insights into the most vulnerable period of PICC self-management. Nevertheless, future studies should incorporate longitudinal follow-up to the association between chatbot use and PICC-related complications, the sustainability of chatbot engagement, and its long-term effects on self-management. Third, the absence of a widely adopted, validated Korean-language instrument specifically designed for assessing chatbot usability limited comparability with existing literature. At the time this study was conducted, chatbot-specific usability instruments were not widely adopted; therefore, similar to prior studies, usability was assessed by adapting items from existing general usability questionnaires to reflect chatbot-specific characteristics [48,49]. Subsequent studies published after the completion of our research have begun to adopt chatbot-specific instruments, such as the Chatbot usability questionnaire; however, a validated Korean-language version or culturally adapted chatbot usability measure remains unavailable. Future research would therefore benefit from the development and validation of culturally appropriate, standardized Korean-language chatbot usability instruments to enable more rigorous and comparable assessments across studies. Finally, the small sample size and single-site design limit the generalizability of the findings and the ability to draw definitive conclusions. Future studies with larger and more diverse samples will be necessary to validate quantitative findings and support more robust statistical comparisons. Moreover, because participants in this study received structured onboarding and brief training, uptake in routine clinical practice where such support may not be systematically provided could be lower. This possibility further underscores the need for implementation strategies tailored to diverse clinical environments.

In this study, the feasibility of a chatbot-based supportive care system for managing PICC lines in patients with cancer and their caregivers was explored. The chatbot had potential in enhancing information access, providing guidance, and offering psychosocial support. In particular, during situations such as the COVID-19 pandemic when patients were unable to visit hospitals or easily contact health care providers, the chatbot served as a valuable alternative by delivering reliable information and independently supporting the resolution of minor issues, thereby improving access to medical guidance. However, limitations in conversational quality and personalization were identified. Further quantitative research is warranted to strengthen the evidence base for the chatbot’s effectiveness. Incorporating LLMs may enhance future chatbot performance and user experience.