The dataset for this study was collected from Naver Knowledge iN, where users post questions and receive answers from the public. Each entry on this platform consists of a user-generated question and one or more community-generated answers. As the primary aim of this study was to examine public perceptions, concerns, and information-seeking behaviors regarding palliative and hospice care, we included only the original question posts, which reflect users’ initial intentions and perspectives. Answers were excluded because they predominantly contained brief institutional referrals or promotional content (eg, advertising health care services or insurance), which were not aligned with the analytical focus of the study.

To retrieve relevant data, we used keywords such as “hospice,” “palliative care,” “palliative medicine,” “advance care planning,” and “end-of-life care.” Posts from the platform’s launch in 2002 up to June 30, 2024, were collected using Python’s BeautifulSoup library for web crawling. The extracted data were stored in a structured format that included the title, body text, and date of creation. In cases where the body text was missing, the title was used as a substitute, as it typically conveyed the user’s central inquiry.

Social media texts, such as those found on Naver Knowledge iN differ from formal medical documentation or academic articles. They often include informal language, emotive expression, and inconsistent grammar [22,23]. These characteristics required customized preprocessing procedures, such as duplicate removal, noise filtering, and lexical normalization [22,24]. This process begins with a thorough understanding of the data source, which is crucial for effective data cleaning in text-mining research [25].

In topic modeling studies, ensuring that datasets contain relevant content is critical for accurately identifying latent themes. During preprocessing, duplicate and null entries were removed from the collected dataset. Notably, some Naver Knowledge iN search results included posts where relevant keywords appeared only in the answer, rather than in the main body of the question. For example, a text such as “What certification should I pursue?” could be included if the answer mentioned “hospice nurse,” even if the main body of the question had no relevance to palliative care. To address this, we applied explicit inclusion and exclusion criteria to ensure the thematic relevance and analytical integrity of the corpus. Inclusion criteria required the presence of keywords directly related to (1) palliative or hospice care concepts, (2) specific diseases or conditions often requiring such care, or (3) life-sustaining treatment procedures. To capture diverse linguistic expressions, we used n-gram extension and ChatGPT-based term augmentation, subsequently reviewed and validated by subject-matter experts.

In contrast, exclusion criteria were derived through keyword frequency analysis and manual screening. Posts were excluded if they contained relevant keywords in non-health-related contexts, such as student assignments, pet care, dermatological concerns, academic or career guidance, insurance advertisements, or fictional writing. A complete list of these criteria is available in the Multimedia Appendix 1. Notably, due to the morphological complexity and contextual nuances of the Korean language, the n-gram and terms generated by artificial intelligence (AI) are retained in their original Korean form to preserve semantic integrity.

To analyze temporal patterns, we extracted the year from each post’s creation date and calculated annual frequencies for each identified topic and sentiment. Because the data for 2024 were incomplete, we estimated full-year values using proportional scaling based on available months. Temporal trends were visualized with line plots, enabling examination of how public discussions evolved in response to major societal and policy events related to palliative and hospice care.

To examine the emotional distribution in web-based posts related to palliative and hospice care and to explore the relationship between topics and emotions, we conducted a multiclass sentiment analysis using the Korean bidirectional encoder representations from transformers (KoBERT) model [31]. By leveraging the KoBERT model’s ability to capture the contextual nuances of texts, we acquire a more refined analysis than traditional binary sentiment classification, which typically divides sentiments into positive and negative categories. Polarity determination, widely used for opinion analysis, has often been applied in areas such as health policy or institutional evaluation [32,33]; however, our multiclass sentiment analysis allows us to uncover individuals’ varied emotional expressions in palliative and hospice care discussions.

The dataset used for developing the model was constructed by merging 3 publicly available Korean sentiment corpora from AIHub. The final dataset included 146,567 texts labeled across 7 emotion categories: “neutral” (n=48,616), “sadness” (n=27,517), “anger” (n=19,710), “anxiety” (n=15,999), “happiness” (n=14,406), “disgust” (n=5649), and “embarrassment” (n=14,670). During model development, the “anger” and “disgust” categories were merged into a single category called “hostility” to reduce confusion due to overlapping expressions. The final model, therefore, predicted 6 emotion classes. We randomly partitioned the dataset into training (75%) and testing (25%) sets. For model architecture, we adopted a BERTClassifier, which consists of the BERT model followed by a linear classification layer. Each sentence was first converted into tokens using a BERT tokenizer, which splits text into smaller linguistic units suitable for input into the model. To ensure uniform input length, all sequences were truncated or padded to a maximum of 64 tokens.

To address class imbalance, we applied Focal Loss, which adjusts the contribution of each class by introducing 2 hyperparameters: alpha, which is calculated dynamically based on class frequencies to assign higher weights to underrepresented classes, and gamma, which focuses learning on incorrectly predicted samples [34]. This approach significantly improves the model’s performance for minority classes. Training was conducted for 12 epochs with a batch size of 16. Throughout the training process, key performance metrics (accuracy, precision, recall, and F₁-score) were tracked for both the training and testing sets. The final model was selected based on the highest test accuracy.

All preprocessing and analyses were conducted using Python. A summary of the overall data processing and analysis workflow is provided in Figure 1.

This study received exemption from review by the institutional review board at Seoul National University (institutional review board number E2407/001-003). All data used in this study were publicly available posts collected from Naver Knowledge iN. In accordance with Naver’s policy, users are explicitly advised not to include personally identifiable information in their posts. Additionally, during preprocessing, we applied term frequency thresholds to remove rare terms and conducted manual screening to exclude any entries that could potentially reveal personal identity. As a result, all analyzed data were deidentified and did not contain information that could be used to identify individual users.

A total of 34,501 posts related to palliative and hospice care were collected from Naver Knowledge iN. After removing the null and duplicate entries, 29,489 posts were screened for relevance. Ultimately, 9072 posts were included in the final analysis after excluding irrelevant content (Figure 2). The token length distribution of the dataset had a mean token length of 177.01, with a standard deviation of 280.45. The IQR was 143.25, and the 50th percentile token length was 107.00.

Figure 3 illustrates the temporal growth of the dataset from 2002 to 2024, with significant events related to life-sustaining treatment decisions marked along the timeline. Over time, a steady increase in the number of records was observed, with several key legal and institutional milestones potentially influencing this trend.

The first notable event, Kim Grandmother’s Case in May 2009 (Figure 3A), coincided with a gradual increase in the growth of the dataset. This event, in which the Supreme Court ruled in favor of respecting patients’ wishes to withdraw life-sustaining treatment, may have sparked an increased public discourse on end-of-life care.

In July 2013 (Figure 3B), the National Bioethics Committee’s recommendation to institutionalize life-sustaining treatment decisions corresponded to a period of steady data accumulation. This growth continued, peaking after the issuance of advance directive registration certificates in January 2019 (Figure 3F) and the establishment of the “1st Comprehensive Plan for Hospice and Life-Sustaining Treatment” in June 2019 by the Ministry of Health and Welfare (Figure 3G). Finally, in April 2022 (Figure 3H), the announcement of regular reimbursement rates for decisions to discontinue life-sustaining treatment correlates with a sharp increase in data volume.

To identify the key topics of interest from 9072 web-based posts related to palliative and hospice care, we performed a CTM while varying the number of epochs and topics. The diversity scores for the topics consistently remained above 0.9, indicating high topic diversity across the models. For the coherence scores, the researchers closely examined models with a coherence score greater than 0.6. Based on this expert review, a model with 9 topics trained over 50 epochs was selected as the final CTM because it balanced diversity, coherence, and interpretability (Figure 4).

Table 1 presents the results of the CTM for 9072 web-based perspectives related to palliative and hospice care that identified 9 distinct themes. The largest theme, “emotional and psychological support,” accounted for 15.95% (1447/9072) of the data, followed by “ethical and legal concerns” (1170/9072, 12.90%) and “medical care in hospitals” (1105/9072, 12.18%). Other identified themes included “nutrition management” (1096/9072, 12.08%), “disease treatment outcomes and prognosis” (1078/9072, 11.88%), “financial issues” (908/9072, 10.01%), “symptom management” (904/9072, 9.96%), “administrative and volunteer services” (832/9072, 9.17%), and “health insurance” (532/9072, 5.86%). These results highlight a broad range of concerns, from emotional support to legal and financial issues, within the context of palliative and hospice care.

Figure 5 illustrates the annual distribution of the topics identified through the CTM analysis, displayed as line graphs, along with the total volume of data per year, shown as a shaded area plot. Notably, topics such as “ethical and legal concerns” and “medical care in hospitals” have demonstrated considerable peaks in recent years, suggesting growing public interest in these areas. In contrast, topics such as “nutrition management,” “symptom management,” and “health insurance” fluctuated around the 2010s but have maintained a relatively lower presence in the 2020s.

Throughout the study period, “emotional and psychological support” consistently remained among the most frequently discussed topics. Perspectives on “disease treatment outcomes and prognosis,” “financial issues,” and “administrative and volunteer services,” topics related to patients’ circumstances, financial matters, and administrative processes rather than direct care, were moderately represented.

Figure 6 shows the performance of the multiclass sentiment analysis model across several key metrics. The left plot shows accuracy trends over 12 epochs, where the train accuracy (blue line) steadily improved to nearly 98%, whereas the test accuracy (orange line) stabilized at approximately 72%. The middle plot provides an overview of the precision, recall, and F₁-score. Precision (blue) peaks early but slightly decreases in later epochs, whereas recall (orange) gradually improves and converges with precision around epoch 8. The F₁-score (green) represents the balance between Precision and Recall, followed by similar trends, and stabilized at approximately 72% toward the end of training. The best model performance was observed at epoch 5, with an accuracy of 0.72, precision of 0.73, recall of 0.72, and an F₁-score of 0.71. The plot on the right presents a Confusion Matrix depicting the true versus predicted label distribution across the 6 sentiment classes for the best-performing model. It indicates strong performance in certain classes, such as “neutral” (label 5), which shows the highest number of correct predictions.

Figure 7 illustrates the annual distribution of sentiments identified in perspectives related to palliative and hospice care using the best-performing multiclass sentiment analysis model identified in the previous stage. “neutral” and “sadness” sentiments dominate the dataset, with “sadness” showing a marked increase from 2015 onwards, peaking in recent years. “anxiety” also exhibits a steady rise over time. In contrast, sentiments such as “embarrassment,” “happiness,” and “hostility” maintain a relatively lower and stable distribution across the time frame.

Figure 8 shows the relationship between the topics identified through the CTM and the sentiments predicted by the multiclass sentiment analysis model. A notable observation was that the majority of topics, such as “emotional and psychological support,” “disease treatment outcomes and prognosis,” “medical care in hospitals,” and “symptom management,” were predominantly associated with “sadness.” All 9 topics exhibit similar connections to the “neutral” sentiment, while topics such as “administrative and volunteer services” and “nutrition management” show the strongest association with “anxiety.”

This study demonstrates that public interest in palliative and hospice care, as captured through web-based discourse, tends to rise sharply in response to major policy changes and high-profile legal cases. These spikes suggest that regulatory and institutional developments, such as the establishment of national advance directive systems or legislation governing life-sustaining treatment, function as societal turning points that prompt individuals to seek information and engage with end-of-life care issues. This pattern is consistent with prior findings by Huemer et al [35], who showed that legislative changes across Austria, Germany, and Switzerland significantly influenced public search behaviors related to palliative care, often redirecting attention away from controversial practices such as euthanasia toward more supportive care models.

The role of media in catalyzing these shifts cannot be overstated. In South Korea, the “Kim Grandmother’s Case” served as a defining moment that not only raised public awareness but also led to institutional and legal reforms. Media coverage, particularly investigative reports and emotionally resonant narratives, can demystify palliative care, correct misconceptions, and stimulate public discourse [36]. Yet, despite these impactful events and ongoing promotional efforts, national awareness of palliative and hospice care remains relatively low at 35.7% (4676/13,546) [9], suggesting that current outreach may not be sufficient to shift cultural norms or drive meaningful behavioral change.

Given these dynamics, social media platforms offer a critical, real-time lens through which evolving public concerns can be monitored and addressed. The clear temporal spikes observed in our analysis illustrate that social discourse is sensitive to policy events and ethical debates. This highlights the need for responsive communication strategies that can align with these moments of heightened attention. By leveraging timely messaging, expanding public education, and integrating awareness campaigns with broader policy initiatives, stakeholders can foster more informed, stigma-free conversations about palliative care. Doing so may ultimately improve uptake of services and better support patients and families navigating end-of-life decisions.

Building on these observations, it is important to consider the specific topics that have gained prominence in public discourse. In particular, topics such as “ethical and legal concerns,” “administrative and volunteer services,” “disease treatment outcomes and prognosis,” and “medical care in hospitals” have demonstrated considerable increases in recent years, suggesting growing public interest and need for information. Numerous studies have emphasized the importance of providing palliative and hospice care information early in the course of illness [37,38]. However, a systematic review revealed that more than half of the patients remain unaware of such services [39]. To address this gap, targeted promotional strategies and public education materials should be developed that speak not only to patients and families but also to the broader public [39]. A more inclusive and tailored communication approach may help meet diverse informational needs and promote earlier, more proactive engagement with palliative and hospice care.

Topics related to “nutrition management” were prominent, indicating heightened interest in dietary considerations within palliative care, which aligns with East Asian caregiving traditions that emphasize high-quality, condition-specific food as a key component of care [40]. This cultural focus may also reflect the influence of eastern medicine, which emphasizes maintaining health through diet and natural remedies [41].

Meanwhile, the “ethical and legal concerns” theme, encompassing keywords such as “euthanasia,” “life-sustaining treatment,” “DNR (Do-Not-Resuscitate),” “refusal of life-sustaining treatment,” and “withdrawal of life-sustaining treatment,” reflects ongoing social discussions in Korea regarding the discontinuation of futile life-sustaining treatments. This topic can be viewed with particular consideration to this study’s findings on temporal trends in web-based palliative and hospice care perspectives. In particular, the 1997 “Boramae Hospital Case” (which preceded our data analysis period) and the 2008 “Kim Grandmother’s Case” drew substantial public attention to the issue of withdrawal of life-sustaining treatment in Korea, ultimately prompting the creation of related legislation. The “Boramae Hospital Case” involved an intensive care unit patient dependent on mechanical ventilation whose family requested discharge against medical advice due to the treatment costs; the patient subsequently died at home. Another family member sued both the requesting family and the medical staff, and the court found the medical team guilty of aiding homicide through inaction [42]. Following this case, conflicts arose between health care providers and families regarding continued life-sustaining treatment for patients with no chance of recovery. These issues became starkly apparent in the 2008 “Kim Grandmother’s Case,” in which the family of an older woman in a persistent vegetative state requested withdrawal of mechanical ventilation, citing the patient’s previously expressed wish for a natural death. However, the earlier “Boramae Hospital Case” influenced the medical team’s refusal. Eventually, the family filed a constitutional complaint, leading to Korea’s first legal precedent permitting the withdrawal of life-sustaining treatment [43].

Despite these legal milestones, withdrawal of treatment remains infrequent in Korea (550/54,699, about 1%) compared with countries like Sweden (7865/12,072, 65.2%) and the global average (42.3%) [44-46]. Contributing factors include persistent misconceptions about treatment cessation, limited public awareness, and low completion rates of advanced care plans among adults, and gaps in the official system for creating and registering these plans [9,47,48]. Nevertheless, the frequent appearance of these issues in web-based discourse reflects a growing desire for compassionate, dignified death. Therefore, we propose integrating palliative and hospice care education into public and school curricula, providing a structured means of preparing for the universal experience of death while emphasizing the value of end-of-life care.

Financial concerns are prominently reflected in public discourse on palliative and hospice care, with approximately 16% (1440/9072) of analyzed posts referencing financial burdens or health insurance, comparable in frequency to expressions of emotional distress. Previous research highlights high treatment costs, limited reimbursement, and income loss from prolonged caregiving as critical barriers to care access [49-51]. For families facing financial strain, home-based palliative care can offer a dignified and more affordable alternative to institutional settings [52]. Indeed, studies have shown that early integration of palliative and hospice services contributes to substantial health care cost reductions [53,54].

In South Korea, the National Health Insurance Service reduces copayments to 0%-10% for patients with terminal cancer, cardiovascular, cerebrovascular, rare, and intractable diseases, significantly lower than the standard 20% for inpatient and 30%-60% for outpatient care. Home-based palliative care, which costs only 9% of equivalent inpatient services (225,688 KRW [US $192.24] vs 2,481,479 KRW [US $2113.70] per week; conversion based on KRW 1174 per US $1 in September 2020), offers significant savings [55]. The long-term care insurance system provides in-home care services for eligible older adults, and the Comprehensive Nursing Service System supports hospital-based care without requiring private caregivers. In addition, the National Responsibility Policy for Dementia offers financial subsidies for dementia caregivers. However, existing policies fall short in several respects. The 90-day unpaid family care leave permitted under the Equal Employment Opportunity and Work-Family Balance Assistance Act lacks wage replacement, thereby exacerbating financial burdens. Moreover, many indirect nonmedical expenses, such as transportation or respite care, remain uncovered.

Educational materials should not only emphasize the cost-saving benefits of palliative and hospice care but also provide detailed information on available financial support. Policy efforts must expand to include paid family care leave and promote equitable access to home-based services. In particular, coordination with existing public health initiatives, such as home-visiting nursing services via community health centers or long-term care insurance programs, could enhance outreach, education, and timely identification of eligible patients in need of palliative care.

A notable finding of the sentiment analysis was the high prevalence of neutral sentiments, which were relatively evenly distributed across all identified topics. This pattern reflects the informational nature of the social media platform, Naver Knowledge iN, where users often seek factual guidance rather than emotional support. Representative questions include: “Where can I draft an advance directive for life-sustaining treatment?” and “Which mushrooms are beneficial to terminal cancer patients?”. These examples point to a strong public demand for practical, reliable, and easily accessible information related to end-of-life care. In response to this need, the development of AI-driven tools, such as large language model–based domain-specific chatbots, could offer a scalable solution. By delivering timely, accurate, and empathetic answers, such tools can help reduce informational barriers, address misconceptions, and better support both individuals directly affected by serious illness and members of the general public seeking guidance.

In contrast, “sadness” emerged as the most frequent emotion, followed by “neutral” and “anxiety.” The prominence of “sadness” reflects the emotional burden commonly associated with serious illness and end-of-life decision-making. Although the identities of individual posters remain unknown due to the platform’s anonymity, it is plausible that many posts were authored by patients, family caregivers, or individuals with personal experience of palliative care. Notably, “anxiety,” “hostility,” and “embarrassment” were closely linked to topics such as “emotional and psychological support,” “nutrition management,” and “symptom management,” suggesting frequent emotional struggles in navigating disease progression. These findings align with prior research reporting high levels of psychological distress among people facing terminal illness [56,57]. Accordingly, beyond managing physical symptoms, health care providers should prioritize emotional and psychosocial support to ensure holistic care [39,58]. Moreover, these insights indicate that educational interventions must be designed not only to deliver accurate medical information but also to meaningfully engage with the emotional and ethical concerns that shape public attitudes toward palliative and hospice care.

Despite the growing public interest in palliative care, communication between health care providers and individuals facing serious illness remains a persistent challenge [18]. Misconceptions, such as viewing palliative care as synonymous with “giving up” on curative treatment, alongside cultural taboos surrounding death, often hinder open and timely dialogue [5]. Addressing these barriers requires culturally sensitive public education and sustained efforts to reframe palliative care as a proactive and empowering choice. In clinical contexts, tools such as the supportive and palliative care indicators tool can facilitate communication by providing clinicians with structured conversation guides that support early and compassionate discussions about palliative needs [59]. Embedding such tools into clinical workflows and linking them with public education initiatives may improve patient-provider communication, reduce stigma, and promote earlier access to supportive services.

Beyond clinical tools such as the supportive and palliative care indicators tool that support provider-initiated conversations, broader cultural change is also necessary to normalize discussions about death and dying across society. In South Korea, while the concept of dignified dying (also known as well-dying) has gained limited attention in the context of an aging society, public discourse around death remains marginal. Cultural taboos rooted in Confucian and Buddhist traditions often associate such discussions with misfortune, inhibiting timely communication among health care providers, patients, and families [60,61]. A national survey of Korean adults revealed that 52.7%-84.0% rarely or never discuss death, and only 17.9%-32.2% feel confident about their end-of-life preferences [62]. To address this, public education, community engagement, and school-based death education programs should be expanded to reduce stigma and promote timely, values-based decision-making about palliative and hospice care.

This study has several limitations, particularly in the context of broader frameworks of serious illness and care. First, the analysis of Naver Knowledge iN posts lacked contextual details regarding the author’s background as well as clinical information such as diagnosis, disease stage, or palliative care status. These limitations stem from the anonymous and unstructured nature of the platform. As a result, the dataset may not fully reflect the diversity of serious illness experiences, including non-cancer conditions such as chronic obstructive pulmonary disease or congestive heart failure [63]. Future studies should incorporate alternative data sources that allow for the identification of author roles and longitudinal care trajectories to better account for structural and contextual influences on palliative care needs. Second, we excluded answer posts from the discussion threads during preprocessing. While many of these entries were promotional or duplicative, their exclusion may have narrowed the range of perspectives included in the analysis. Future research could explore these responses using preprocessing and filtering techniques to extract informative content. Third, although our multiclass sentiment analysis achieved robust performance (accuracy, precision, recall, and F₁-score all above 0.70), class imbalance remained a challenge. For instance, low-frequency emotions such as “embarrassment” may still be underdetected [64]. This issue could be addressed in future work through advanced data augmentation techniques or by collecting larger and more diverse datasets. Fourth, the perspectives of bereaved families may have been underrepresented due to the keyword-based data collection strategy. Given that author identity could not be determined, it is likely that posts written by bereaved caregivers were unintentionally excluded. Refining search strategies and terminology to better capture bereaved experiences would improve the inclusivity and depth of future research. Last, although our approach was grounded in large-scale computational analysis, it did not examine whether individual posts referenced specific policy changes or widely mediatized cases. Future research could build on our findings by conducting a secondary analysis of a subset of our dataset, focusing on posts written by specific groups such as patients and family caregivers. Qualitative methods, such as content analysis or reflexive thematic analysis, could help uncover how these groups interpret ethically complex issues such as the withdrawal of life-sustaining treatment. Such analysis would also enrich our understanding of the narratives and concerns that are unique to these groups and how they respond to social and institutional changes.

This study offers novel insights into public discourse on palliative and hospice care in South Korea by analyzing large-scale social media data. Our findings reveal that emotionally negative sentiments, particularly sadness and anxiety, are prevalent and closely tied to key concerns such as symptom management, ethical dilemmas, and financial burdens. Notably, approximately 16% (1440/9072) of posts addressed concerns related to caregiving stress or economic hardship, reflecting the multifaceted challenges of palliative care as expressed by individuals participating in web-based discussions. The prevalence of neutral, fact-seeking queries underscores a strong public demand for accessible and reliable information. At the same time, widely documented misconceptions, such as viewing palliative care as giving up on treatment, continue to hinder effective communication, reinforcing the need for clearer communication. However, the anonymous and unstructured nature of platforms like Naver Knowledge iN limits their potential for sustained, interactive dialogue. To address this gap, user-centered tools, such as culturally tailored educational materials, AI-powered chatbots, or moderated web-based forums, could provide timely and emotionally responsive support. These tools may also bridge public needs and clinical systems by identifying emerging concerns and facilitating early intervention. In parallel, policy and education initiatives, including the integration of death education into school and community curricula, are essential to normalize conversations about dying and promote informed, compassionate decision-making. By demonstrating the utility of social media analysis for capturing real-time public concerns, this study contributes to global efforts in end-of-life care advocacy, underscoring the importance of culturally responsive communication strategies and emotional support infrastructures.