Kidney diseases primarily manifest in 2 main forms: chronic kidney disease (CKD) and acute kidney injury (AKI), which are 2 significant kidney conditions [1,2]. CKD is characterized by a progressive decline in kidney function lasting >3 months, whereas AKI refers to a sudden decrease in kidney function that occurs over a period of hours to days [1,2]. Both conditions can arise from various underlying causes, including primary glomerular diseases (eg, glomerulonephritis), inherited disorders (eg, polycystic kidney disease), and secondary complications of systemic conditions such as diabetes and hypertension [2,3]. These conditions collectively affect a significant proportion of the global population, with CKD alone estimated to affect 13.4% of people worldwide [4]. Both CKD and AKI, regardless of their underlying etiology, increase the risk of other medical complications and significantly impact patients’ quality of life and health care costs [2,5]. Managing these conditions can require substantial lifestyle changes, particularly in dietary habits, which can be challenging for patients [6]. However, maintaining a healthy lifestyle can reduce the risk of kidney disease progression and lower mortality rates [7].

Despite the critical role of lifestyle modifications in managing kidney diseases, patients often report receiving ambiguous or conflicting information from health care providers, leading to confusion and a sense of disempowerment [8]. Time constraints in clinical settings further limit the extent of guidance that health care providers can offer, leading many patients to turn to online health communities for information and support.

The use of web-based platforms for health-related information has become increasingly prevalent, with approximately 80% of patients using the internet to exchange information and over one-third of adults using social media for health information and social support [9,10]. Unlike traditional research methods such as surveys or interviews, social media analysis provides access to authentic, unprompted patient discussions, offering unique insights into their daily challenges and evolving needs [9,10]. This approach has been widely used to understand patient experiences within chronic disease communities. For example, studies have used social media analysis to capture patient narratives about cancer [11] and examine the social media behaviors of patients and health care providers across various conditions, including asthma, mental health disorders, diabetes, and gout [12,13].

While topic modeling has been widely used to extract health-related topics from social media, studies specifically examining public discourse on kidney diseases are scarce, particularly in South Korea. Despite the country’s high internet penetration and active online health communities, research that systematically analyzes patient experiences with kidney disease through social media remains limited. Successful applications of social media analysis in South Korea include studies on diabetes-related communication on social networking services [14], evaluations of physician-created thyroid cancer content on YouTube [15], the use of topic modeling for pancreatic cancer information [16], and investigations into cross-cultural perceptions of urological conditions [17]. However, comprehensive studies of kidney disease communities are notably absent. These communities often reflect unique cultural challenges and interactions with the health care system that are specific to South Korean patients.

Recent advances in natural language processing (NLP) have transformed how researchers analyze and interpret unstructured health-related text data. Deep learning approaches in NLP have shown significant success in processing various types of health care text, including clinical narratives, electronic health records, and social media content [18,19]. These computational methods are particularly effective in uncovering patterns and insights that are not readily apparent through traditional analytical methods. Applications of NLP in health care range from extracting medical entities and relationships from clinical texts to analyzing patient sentiments, behaviors, and experiences shared in online health communities. Such approaches are instrumental in bridging the gap between raw data and actionable insights in health care research.

While topic modeling allows for quantitative analysis of discussed topics, it does not fully capture the emotional and experiential dimensions of patients’ lives. To address this limitation, our study combined topic modeling with sentiment analysis, using advanced NLP techniques to provide both quantitative insights and a qualitative understanding of patients’ emotional challenges. This dual approach bridges the gap between unstructured patient narratives and actionable health care insights. By focusing on the experiences of South Korean patients with kidney disease, our findings can inform the development of culturally appropriate, patient-centered interventions, enabling health care providers to address patient concerns more effectively, design impactful educational materials, and enhance support systems tailored to these patients’ unique needs.

The main objectives of this study were to (1) analyze social media data using topic modeling to better understand the experiences, challenges, and concerns of patients living with various kidney diseases and their caregivers in South Korea; (2) explore the ways in which online communities can provide assistance, specifically in relation to disease management for different kidney conditions; and (3) examine the sentiment surrounding disease management among individuals with kidney diseases.

By achieving these objectives, this study aimed to provide valuable insights that can inform health care providers, policy makers, and support organizations in better addressing the needs of patients with a spectrum of kidney diseases and renal conditions.

The goals of this study were to identify the challenges faced by patients and determine the key topics related to kidney disease management. We used topic modeling for quantitative analysis and sentiment analysis for qualitative analysis to achieve this. Relevant posts were gathered from online communities discussing kidney disease, preprocessing was conducted, and the topics and sentiments were analyzed. Figure 1 provides an overview of the research methodology.

We collected data from KidneyCafe, a prominent South Korean online community for patients with kidney disease and caregivers hosted on the Naver platform [20]. This community has approximately 183,545 members and facilitates discussions across various themed boards, each concentrating on specific kidney-related topics. For our analysis, we focused on the disease-specific boards within the Member Participation section, which is dedicated to active discussions among engaged community members. These boards address topics such as kidney transplantation, dialysis, and related medical conditions, enabling us to capture in-depth conversations pertinent to a wide array of kidney disease experiences and management strategies.

The Selenium WebDriver was used to extract 124,211 posts from 10 disease-specific boards, dating from the community’s inception on December 15, 2003, to the date of data collection (April 20, 2024). The collected data included post titles and content.

Data preprocessing was conducted using Python (Python Software Foundation), with the KoNLPy library used for Korean NLP [21]. This step involved removing punctuation and noninformative words as well as extracting nouns from the text to prepare for subsequent analysis. To handle informal language patterns in user-generated content, we used a custom-built stop word list to exclude frequently appearing, noninformative terms that do not contribute to semantic analysis. This stop word list was created by reviewing existing lists from similar Korean health-related studies [22] and further customizing it to focus on discussions surrounding kidney disease. These preprocessing steps were essential in ensuring the robustness of our analysis and the interpretability of the results.

We used latent Dirichlet allocation (LDA) topic modeling [23], an unsupervised machine learning model, to uncover latent topics in the collected posts. Topic modeling is a technique widely used in NLP for discovering topics and extracting meaning from large, unstructured documents. LDA is the most popular topic modeling method and is based on statistical probability inference. It assumes that each document can be represented by a probabilistic distribution over latent topics, and each topic is characterized by a probabilistic distribution over the words [24].

We imported LDA models from the Gensim library for this study. When analyzing topics using LDA, the number of topics is typically determined using coherence scores such as U_mass and perplexity [25,26]. The U_mass coherence score, which is based on word co-occurrence statistics within documents, has been validated to closely correspond with human labeling results [27,28]. This score has been effectively used in analyzing health-related social media topics, including obesity, COVID-19, and vaccines [27-30]. Similarly, perplexity has proven to be a valuable metric in recent topic modeling studies [31,32]. While a higher coherence score typically indicates better topic quality and interpretability, recent research shows that statistical optimization alone may not guarantee optimal interpretability of topics [32-34]. To tackle this issue, several studies have adopted validation methods that integrate statistical metrics with qualitative assessments. For instance, Min et al [31] combined perplexity and coherence metrics with qualitative evaluations in their analysis of occupational accidents, whereas Lyu et al [33] and Li et al [32] emphasized the significance of semantic coherence alongside statistical validation.

Following established approaches [32-34], we first calculated perplexity and U_mass coherence scores across a range of topics (3-40) using 10-fold cross-validation. As shown in Figure 2, both metrics demonstrated better performance with fewer topics, leading us to concentrate on the range of 3 to 10 topics. Within this range, we conducted 10 iterations of the model to ensure consistency and, ultimately, selected 6 topics. This decision was based on a combination of qualitative assessments of topic interpretability and the high U_mass coherence scores observed at both 3 and 6 topics (Figure 3).

Using the optimal cluster size, we ran the LDA model to obtain lists of representative words for each cluster. The topic-labeling process involved 4 researchers, each with diverse expertise in nutrition and food science, who collaboratively reviewed and interpreted the word lists. We followed established methodologies in social media analysis [33,35,36] using an iterative, discussion-based approach. Initially, each researcher independently analyzed the word lists generated by the topic modeling algorithm, focusing on the most significant terms and their relative weights within each topic to identify potential themes. Next, they reviewed representative posts selected for their strong association with each topic to contextualize the keywords and understand their relevance to the experiences and concerns of patients with kidney disease. This independent review enabled researchers to form initial interpretations and propose tentative labels for further discussion.

After the independent review, the researchers gathered to compare their initial interpretations and collaboratively refine the proposed labels through discussion. When differing interpretations emerged, the team reviewed additional example posts from the topic clusters to gain more context. This process included examining word clouds and representative posts that distinctly exemplified each topic while also considering the specific experiences and information-seeking behaviors of patients with kidney disease. Final topic labels were assigned once all 4 researchers reached a consensus through comprehensive discussion and contextual analysis.

The contribution of each word to a cluster was evaluated using the probability metric derived from the model results. Representative example sentences for each cluster were identified based on the probability that each post belonged to the topic.

The Bidirectional Encoder Representations From Transformers (BERT) method, a machine learning technique for NLP, was used to identify the sentiment of posts about dietary management—positive, negative, and neutral—to uncover which aspects of dietary management were the most challenging. While word embeddings can be derived from large and unannotated corpora using word co-occurrence statistics, they do not consider the context when creating word vectors [37]. However, BERT is context dependent; it produces dynamic and context-specific representations of the word by accounting for neighboring words, which helps interpret the semantics along with the grammatical structure [38].

The sentiment analysis in this study used the Naver CLOVA Sentiment platform, which uses a pretrained BERT model specifically adapted for analyzing Korean-language texts. This adaptation enables the model to effectively capture context-specific nuances, particularly in the informal and colloquial language common in social media discussions. While the specific parameters and fine-tuning details of the application programming interfaces are proprietary to Naver and cannot be disclosed, the effectiveness of this methodology for Korean sentiment analysis has been validated through several academic studies [39-41]. The platform provides sentiment classification results (ie, which sentiment—positive, negative, or neutral—is best represented in the texts), along with the expected probability of each sentiment.

The disease management topic was investigated by conducting a detailed analysis. One author manually classified the posts categorized under disease management by the LDA results into questions and nonquestions. In total, 3 additional authors independently annotated a subset of 300 posts to establish the annotation rules. The posts were reviewed for seven key sentiments: (1) general dietary management methods, (2) food recommendations, (3) inquiries about the safety of consuming specific foods, (4) requests for dietary evaluations, (5) water management, (6) exercise management, and (7) weight management. After discussion, the annotators agreed on a finalized set of annotation rules and guidelines.

Next, 2 independent annotators labeled a random set of 1000 posts from the overall dataset. The Cohen κ coefficient for interrater reliability was 0.77, indicating substantial agreement. These annotated data were used for transfer learning. We fine-tuned the Efficiently Learning an Encoder That Classifies Token Replacements Accurately (ELECTRA) model, a pretrained language model that has demonstrated state-of-the-art performance across various Korean-language tasks [42]. Compared to BERT and the A Robustly Optimized BERT Pretraining Approach model, ELECTRA offers both computational efficiency and improved performance on downstream classification tasks [43], making it particularly suitable for handling large-scale datasets such as ours. In addition, ELECTRA’s generator-discriminator training mechanism has been shown to outperform masked language models in tasks that require nuanced contextual understanding, as validated in previous Korean NLP studies [42]. We used 1000 manually annotated posts split into training and test sets at a ratio of 80:20. The evaluation metrics on the test set yielded an F₁-score of 0.85, indicating high performance. This fine-tuned ELECTRA model was then used to annotate the question posts related to the topic of disease management.

This study used preexisting publicly accessible data with no identifiable information and therefore does not require institutional review board review per Federal Regulations for the Protection of Human Research Subjects (45CFR 46.104(d)(4)) or patient consent [44]. In compliance with these guidelines and established research practices, all identifying information, such as usernames and specific identifiers, was anonymized or removed before analysis. This approach is consistent with methodologies used in similar international studies using publicly accessible data [33,45,46], as well as previous research analyzing public Naver Cafe posts [47-49].

We gathered a comprehensive dataset that included 124,211 posts from various discussion boards in the KidneyCafe community to better understand the experiences, challenges, and concerns of patients living with kidney disease and their caregivers in South Korea. These posts were written by 34,472 unique users, indicating a highly engaged community. Of these users, 34.92% (12,039/34,472) wrote multiple posts, with an average of 3.6 (SD 7.3) posts per user. This high level of repeated engagement suggests that many users found ongoing value and support within the community. Figure 4 illustrates the monthly distribution of posts, showing a significant increase during the beginning of the COVID-19 pandemic in South Korea.

One of the notable advantages of this community is the categorization of discussion boards according to specific diseases. An examination of the number of posts collected per board (Table 1) showed that the highest number of posts was related to kidney transplantation (39,221/124,211, 31.58%), followed by kidney dialysis (27,692/124,211, 22.29%).

We analyzed the distribution of topics across different boards to gain insights into the specific focus areas in each board (Figure 5). Our analysis revealed that the Kidney Transplant board had the highest concentration of posts related to the family health and support topic and also showed a significant focus on the medication and side effects topic. Of note was that the Hemodialysis & Peritoneal Dialysis board had a higher proportion of transplant-related posts than of dialysis-specific ones. The Nephrotic Syndrome & Glomerulonephritis and Proteinuria & Hematuria boards showed a high concentration of posts related to the examination and diagnosis topic. The Pediatric Kidney Disease board notably focused on the medication and side effects topic, whereas the disease management topic was particularly prominent in the Hypertension & Diabetes board.

Posts on each topic cluster were analyzed to determine their sentiment, which was then categorized as positive, neutral, or negative. The results indicated that posts on the medication and side effects topic contained the highest proportion of negative sentiments (Figure 6). This cluster often included discussions about adverse reactions to prescribed medications, concerns about the impact of other drugs on kidney health, and general anxiety about the side effects of medications used to manage kidney disease. The following are a few posts from the community expressing these sentiments:

While many posts described challenges, some shared positive experiences, such as temporary relief from side effects:

Similarly, the surgery for dialysis and examination and diagnosis topics mainly comprised posts with negative sentiments. These posts often discussed difficulties with dialysis, dissatisfaction with diagnostic outcomes related to kidney disease, and other related issues. The following are some examples from each topic:

Positive sentiments were observed when participants shared their favorable test results:

The family health and support topic also comprised posts with many negative sentiments, primarily expressing worries regarding kidney transplantation among family members:

However, alongside these concerns, high proportions of posts about physical improvements after transplantation were also notable among the positive sentiments:

While the disease management topic also had significant negative sentiments, they were relatively fewer compared to the negative sentiments in other topics. Negative posters often expressed frustration with managing the disease or feeling overwhelmed by uncertainties in disease management:

In contrast, many posts inquiring about disease management strategies were classified as expressing neutral sentiments:

In the costs and insurance topic, negative and neutral sentiments were expressed in similar proportions. Negative posts often focused on concerns regarding delays or cost issues in insurance paperwork, such as the following:

Satisfaction with the health care system was occasionally noted:

In contrast, neutral posts mainly shared information regarding the handling of costs or insurance, such as the following:

Given the complexity and importance of self-management for patients with kidney disease, we conducted an extensive analysis of the disease management topic. We carefully sorted the posts in this category into 2 main groups: information sharing and questions, as shown in Figure 7. We used an ELECTRA-based model to further categorize the questions into 6 specific subtopics. This process allowed us to qualitatively examine which issues were being actively discussed. We also analyzed the sentiments associated with each specific subtopic, which was important for determining which subtopics were causing more difficulties for the patients (Figure 8).

The number of posts under the questions category was >4 times higher than that of posts under information sharing (Figures 7 and 8). The information sharing category included valuable resources such as Recommended information for post-transplant patients and Nephrotic syndrome management methods extracted from the Korean Dietetic Association. Most posts in the questions category were related to dietary management (Figure 7), and these subtopics also comprised posts with more negative sentiments (Figure 8). Specifically, 29.5% (4276/14,495) of the inquiries were about the safety of specific foods, 20% (2899/14,495) sought guidance on dietary management strategies, and 10.9% (1580/14,495) requested dietary recommendations. Examples of inquiries about the safety of specific foods included the following:

Examples of inquiries seeking guidance on dietary management strategies included the following:

Examples of inquiries requesting dietary recommendations included the following:

Other popular subtopics included water intake management (1145/14,495, 9%), weight management (1015/14,495, 7%), and exercise management (971/14,495, 6.7%). Examples of these questions included the following:

Our study findings provide valuable insights into the concerns and challenges faced by patients with different kidney diseases and conditions. The analysis of online community discussions revealed several key points with important implications for patient care, education, and support.

Despite the active discussions in kidney disease communities, a notable paucity of research analyzing these communities exists. To our knowledge, the only study that has previously analyzed a kidney disease community focused on the National Kidney Foundation’s online community for CKD [73]. However, this study only qualitatively analyzed approximately 20,000 posts without using topic modeling techniques. In contrast, our study used text mining methods to analyze a larger dataset, capturing more comprehensive insights into the perspectives prevalent in kidney disease communities. Another key advantage of our study was the ability to analyze posts based on disease state, which was not possible in the National Kidney Foundation’s online community due to the lack of separate boards for different disease states. The previous study did not conduct a sentiment analysis either.

Other studies exploring the challenges faced by patients with kidney disease have relied on surveys or interviews [67,76]. Analyzing online communities offers practical and voluntary insights into consumer sentiment [77], as highlighted by previous studies [78,79]. Our approach captured a more mainstream perspective due to the vast amount of user-generated content available on the web over extended periods [79]. Traditional survey methods are limited by resource constraints in analyzing data within specific spatial and temporal boundaries. In contrast, diverse individuals contribute extensive posts on web-based platforms, making text mining particularly valuable for the qualitative analysis of such large datasets.

Our findings resonate with the themes identified in previous studies. For instance, previous studies [73,76] have highlighted themes related to managing CKD and symptoms, medical tests and results, and navigating health care and clinical needs, which translated to medication and side effects, tests and diagnostics, and disease management in our study. In addition, family health and support and dialysis surgery have been similarly categorized as themes such as considering dialysis and family and transplantation, reflecting the dynamics of CKD status. Support on the forum, a significant theme in previous research [73], was classified in our study as sharing information within disease management, emphasizing the role of information exchange and communication in disease management.

While our study provides valuable insights into the discussions and challenges in online kidney disease communities, it is not without limitations.

First, the data were limited to posts from the KidneyCafe community, which may not represent all patients with kidney disease. Individuals who do not actively participate in online communities—due to factors such as digital literacy, access to technology, or personal preferences—were likely underrepresented. This limitation may particularly affect perspectives from older adults and individuals in socioeconomically disadvantaged groups. In addition, the specific context of Korean patients with kidney disease, such as differences in race, lifestyle, economic status, diet, and health care systems, may not be fully generalizable to other countries. These factors play a crucial role in shaping patient experiences and discussions. Conducting comparative analyses that use data from various web-based platforms and formal health care databases can offer a more thorough understanding of how health care systems and societal attitudes affect patient conversations. Furthermore, cross-platform comparisons would help validate our findings and reveal any potential biases specific to certain platforms in patient narratives. Expanding future studies to encompass diverse communities and geographical regions would improve the generalizability and comprehensiveness of our findings, providing deeper insights into the experiences and needs of patients with kidney disease worldwide.

Second, while the topic modeling method (LDA) was effective in extracting main topics, it does not inherently determine the optimal number of topics and may not fully capture semantic relationships between topics. The ELECTRA model, while efficient for our classification tasks, could benefit from domain-specific pretraining on medical corpora to enhance its performance on health-related text. These methodological constraints may limit the depth and precision of the topic modeling and classification results.

Third, the manual annotation process for classifying question posts related to disease management achieved substantial interrater agreement (Cohen κ coefficient of 0.77). However, the potential for bias cannot be entirely ruled out. To mitigate this, clear and detailed annotation guidelines were developed and thoroughly discussed by all annotators before the process began. Regular discussions were held to resolve disagreements and ensure consistency in interpretations, considering additional context from posts when necessary to reach consensus. While these measures contributed to the high interrater reliability observed, the limited number of annotators may have introduced some residual bias.

Fourth, the Naver CLOVA Sentiment platform’s BERT-based analysis provides significant advantages for Korean text analysis, particularly its ability to accurately interpret nuanced, context-dependent expressions in kidney disease discussions. However, this approach also has its limitations. Domain-specific lexicon-based methods, widely used in English-language sentiment analysis, provide greater transparency by enabling researchers to trace sentiment classifications back to specific words or phrases [80-82]. These methods have demonstrated effectiveness and explainability in both medical [80,81] and public health research [82], achieving meaningful results. However, their application to Korean-language analysis is constrained by the limited scale and domain-specific coverage of existing Korean sentiment lexicons [83]. This limitation highlights the trade-offs inherent in our methodology—while the BERT-based approach effectively captures contextual nuances, it is less interpretable compared to lexicon-based methods. In addition, the Naver CLOVA Sentiment platform’s computational intensity and token limit constraints may restrict its ability to process longer posts comprehensively. These choices reflect our decision to prioritize accuracy in contextual understanding, which was crucial for analyzing complex and nuanced patient discussions in Korean medical contexts.

Future research could address these limitations by incorporating data from multiple online communities, formal health care databases, and private discussions through surveys or interviews. Combining qualitative and quantitative methods, such as incorporating expert review processes or using correlated topic models, could further refine topic selection and analysis. Expanding the scope to include diverse communities from different countries would provide broader perspectives, enhancing the comprehensiveness and generalizability of the findings. These efforts would contribute to a deeper understanding of the experiences and needs of patients with kidney disease across varied contexts.

The findings of this study have important implications for clinical practice, emphasizing the crucial role that health care providers play in addressing the emotional concerns and information needs of patients with kidney disease, as is prominently reflected in web-based discussions [84].

The high prevalence of negative sentiments in posts about medication side effects and dialysis surgery underscores the need for health care providers to proactively address patients’ fears and anxieties. Strategies may include providing detailed explanations of what to expect during treatments, offering practical coping mechanisms, and regularly monitoring patients’ emotional well-being. Research indicates that such proactive support can enhance treatment adherence and improve overall clinical outcomes [85].

Our findings underscore the necessity for clearer and more consistent dietary recommendations as diet-related inquiries dominated the disease management topic. Dietary adherence is particularly challenging for patients with kidney disease due to the complexity of nutritional restrictions and their impact on daily life [86]. Patients often express confusion and frustration with conflicting or vague dietary advice, which hinders effective disease management. Health care providers should prioritize offering patient-specific, practical, and easy-to-understand dietary recommendations tailored to the disease stage, comorbidities, and individual preferences. Personalized dietary counseling has been shown to improve adherence and clinical outcomes [87]. Regular follow-up consultations are essential for adjusting these recommendations based on patients’ experiences and evolving health needs.

In addition, posts expressing frustration, confusion, or anxiety underscore the importance of empathetic communication and psychosocial support from health care providers. Integrating emotional support into routine care can effectively address these concerns. By fostering empathetic patient-provider relationships and offering clear, actionable guidance on disease management—especially regarding diet—health care providers can empower patients to actively participate in their care, ultimately improving their quality of life.

Our study offers an in-depth analysis of web-based discussions among patients with kidney disease, highlighting key concerns and informational needs. The findings underscore the importance of online communities in providing support and education to patients and caregivers. These insights can inform health care providers and policy makers when developing targeted interventions to address the specific needs of patients with kidney disease. Precise and tailored patient education regarding disease management from health care providers is clearly needed and can be significantly enhanced through the insights gathered from online community discussions.