The data sources comprised 2 datasets: ADHD-related FAQs were obtained from publicly available content on the “ADHD Baroalgi (Understanding ADHD)” website operated by the Korean Medical Association [27] and the “National Center for Mental Health Portal” provided by the Ministry of Health and Welfare [28]. The second dataset consisted of publicly available questions and answers authored by anonymous users on “Naver Jisik-in (Naver Knowledge iN),” a widely used online question and answer (Q&A) platform [29]. Both datasets were publicly accessible and unrestricted. To distinguish between the 2 datasets, the official FAQs provided by the Korean Medical Association and the National Center for Mental Health Portal were referred to as provider-generated content, whereas the questions collected from Naver Knowledge iN were referred to as consumer-generated content. In this study, the term “provider” refers to official institutions that provide health information on ADHD. However, because members of the general public search for and consume health information in online community environments, they are referred to as “consumers” or the “online community.” These terms collectively refer to anyone who accesses information through online communities, including patients, caregivers supporting patients, and members of the general public who suspect they may have a condition.

All provider-generated ADHD FAQs and their corresponding answers were collected manually, consisting of 88 Q&A pairs. Online community questions and their corresponding responses were collected by accessing the open API of the online community platform using Python (version 3.10.11; Python Software Foundation) after obtaining an API key through registration with the platform provider (ie, Naver).

A 2-step approach was used to identify questions highly relevant to ADHD in the online community. We first retrieved and reviewed the 1000 most recent questions on “ADHD” posted between December 5, 2023, and November 28, 2025. A total of 1000 ADHD-related questions were tokenized, and a word frequency analysis was conducted to identify commonly occurring terms. The terms, listed in descending order of frequency, were as follows: “symptoms,” “treatment,” “child,” “adult,” “attention,” “pediatric,” “concentration,” “diagnosis,” “hospital,” “examination,” “early stage,” and “inquiry.” Among these terms, “attention” was excluded from further analysis because it is inherently embedded in the disease name, ADHD. Similarly, “pediatric” was excluded during the data retrieval process because it semantically overlaps with “child.”

A second round of search and retrieval was conducted using the query format “ADHD AND high-frequency keyword,” with 100 questions extracted per keyword across the 10 identified terms (yielding 1000 questions). After removing duplicate entries and conducting preprocessing, 4988 online community questions and their corresponding answers were retained for final analysis.

This study was reviewed and approved by the Institutional Review Board of Seoul National University (IRB E2508/002-008).

This section sequentially addresses the 3 RQs. First, RQ1 was examined by analyzing the threshold optimization for semantic similarity and the subsequent coverage assessment of online community questions by provider-generated FAQs. Next, RQ2 was examined by identifying and characterizing online community questions that remained uncovered by FAQs, highlighting unmet information needs. Finally, RQ3 was explored by comparing response styles between provider-generated FAQ answers and online community responses using structured LLM-assisted content analysis.

For each word embedding model, a similarity matrix with dimensions 4988×88 was constructed to calculate the pairwise similarity between provider-generated FAQs and online community questions. The initial distribution-based criterion was the 90th percentile of the similarity score distribution. The resulting threshold values were 0.7522 for MBERT, 0.5912 for LaBSE, and 0.8879 for mE5L (Figure 1).

As the second threshold criterion, cumulative distribution curves were constructed by listing questions similar to question B to determine the final similarity values across a range of questions. The knee point, where the rate of change became notably more gradual, was identified from these curves and used as a threshold. This knee point was not determined by a simple slope measurement but corresponded to the point at which the coverage performance reached its maximum value. Thus, this threshold indicates the point beyond which further increases in similarity do not significantly enhance the FAQ’s coverage of online community questions. The Kneedle algorithm, implemented using the Python Kneed library, was used to identify the knee points in the cumulative similarity distributions [38]. The Kneedle algorithm allows the adjustment of inflection point detection by setting the sensitivity, which was set to the default value of 1. The resulting knee point thresholds were 0.4742 for MBERT, 0.4567 for LaBSE, and 0.8738 for mE5L (Figure 2).

As the third threshold criterion, the threshold optimization step was validated manually. Optimization was performed in 2 phases (calibration and confirmation), and the results are presented in Tables 2 and 3. The Opt threshold consistently yielded higher F₁-scores than the heuristic threshold selection methods across the evaluated embedding models. Under the Opt threshold, MBERT achieved the best overall balance between recall and precision, resulting in the highest F₁-score. Therefore, the optimized threshold identified using the validation dataset was adopted as the reference threshold for subsequent automated semantic similarity classification.

For the calibration phase, the calibration validation set consisted of 90 question pairs for each of the 3 embedding models (mE5L, MBERT, and LaBSE), totaling 270 pairs. The Cohen κ coefficient for interrater reliability between the 2 independent reviewers was 0.8596. Among the 270 reviews, 11 discrepancies were observed. The author subsequently provided comments and feedback to the external reviewer, leading to a consensus. The external reviewer accepted 6 of the 11 discrepancies, with the author agreeing with the external reviewer’s assessments. After the calibration validation, the performance metrics for each model are presented in Table 2. In the information retrieval evaluation literature, in situations where relevant items are sparse or exploratory searches are required, precision and recall are considered key metrics rather than accuracy, and the F-measure, which reflects the balance between these 2 indicators, is widely used. Accordingly, this study set the F₁-score as the primary metric for model and threshold selection. The final embedding model was selected by calculating the sample size for the additional review dataset based on the F₁-score identified during the calibration validation process [42]. The 95% CI for the F₁-score was constrained to a width of 0.20 or less. Applying this criterion to the provided code, the minimum required sample size was determined as 246. Consequently, 270 samples were selected for the confirmation set using the same method as in the calibration phase. Subsequently, to finalize the model’s threshold, each reviewer independently conducted and adjusted a manual review of 270 questions. Cohen κ coefficient for interrater agreement was 0.8282. In the second validation, there were 14 differing opinions out of 270, which were conveyed to the external reviewer in the same manner as in the first validation. The external reviewer accepted 10 of the 14, and the researcher also agreed; therefore, the final decision was made. The interrater agreements are presented in Multimedia Appendix 1. After confirmatory validation, the performance metrics for each model are presented in Table 3. By comparing the threshold criteria for each model on the test validation set, the optimal performance in terms of the F₁-score was observed with an MBERT manual grid search threshold of 0.7660 (F₁-score=0.4500). Under these conditions, the precision was 0.3913, and the accuracy was 0.7556, both of which were the highest, and the overall balance of performance was also the best. In contrast, under the F₂-score criterion, the Kneedle threshold of MBERT yielded the highest value (F₂-score=0.5859; recall=0.8824). This suggests that adopting a lower threshold may be advantageous in the exploration of unmet needs. However, as the precision was lower and the risk of false positives increased, the F₁-score was used as the primary metric in the final model selection, with the F₂-score and recall considered auxiliary indicators.

When the similarity score, calculated from the closest semantic match among the FAQs for a consumer’s question, is above 0.766, it indicates that the health information provider understands the consumer’s intent. The author expressed this as FAQ covering questions from the online community. With the best similarity threshold for MBERT set at 0.766, 2598 of the 4988 online community questions were matched to at least one FAQ item, achieving a coverage rate of 52.09%. To examine how individual FAQs contributed to this coverage, the FAQs were ranked according to the number of online community questions that matched each item. As a single FAQ could be matched with multiple online community questions, the ranking reflected the relative contribution of each FAQ to overall coverage. Cumulative coverage across the ranked FAQs was analyzed using the Kneedle algorithm [38]. The knee point was determined at rank 18, signifying that FAQs constituted approximately 43.89% of the matched online community questions (Figure 3).

Table 4 presents a list of community questions that most closely align with the top 10 FAQs, as depicted in the cumulative distribution curve above. The FAQs were ranked according to the number of community questions that received judgments surpassing the threshold of 0.766.

To determine whether it is meaningful to assume that FAQ questions with a similarity score above a certain threshold cover consumers’ information needs, the author specifically examined the responses corresponding to the 18 FAQs up to the knee point. The author examined whether these answers were appropriate responses to the intent reflected in the questions most similar to the 18 FAQs. Of the 18 answers, 16 (88.9%) were deemed to adequately address the consumers’ inquiries, while 2 (11.1%) were deemed insufficient. One question appeared likely to be answered by describing the drug’s effects, while the intent of the other was misjudged. Therefore, we concluded that comparing question similarity was a feasible and practical approach for quickly identifying consumers’ unmet information needs.

According to the similarity comparison analysis, the proportion of questions in online communities not covered by the FAQs was 47.91%, totaling 2390 questions. By analyzing these, we identified the information needs of consumers that were not addressed in the FAQ. To explore the unmet information reflected in these questions, the researchers conducted an analysis using the LimTopic framework [43]. LimTopic used UMAP and BERTopic, followed by GPT-4o-mini. The parameters for these methods are accessible via publicly available codes. LimTopic identified 118 topics, although numerous instances of duplicate content were observed. To elucidate unmet needs, the researcher initially categorized the data and sought a review from the same external expert. Subsequently, the researcher integrated the feedback to finalize the study results. The final unmet need outcomes are presented in Table 5.

We conducted a structured LLM-assisted content analysis to compare expert-generated FAQ answers with online community responses. The purpose of this analysis was not to evaluate the correctness of the responses but to explore the differences in the stylistic and communicative characteristics of responses. To this end, among the 88 FAQs, the answers to 62 questions for which the highest similarity with any consumer question exceeded the similarity threshold, along with the corresponding answers to the online community questions, were designated as analytical data. A predefined codebook (Table 1) was used for the analysis. Figure 4 shows the distribution of response classifications across the 4 dimensions.

In this study, chi-square tests were used to explore differences in the distribution of questions and responses across 4 dimensions (communicative function, evidence and authority signaling, outcome framing, and actionability), between provider-generated FAQs and online community responses. The chi-square statistics and P values for each area helped determine whether there were significant differences between the 2 groups. First, in the communicative function dimension (Figure 4A), the results showed no significant difference (χ²₂=2.44; P=.30) between the 2 groups. This means that both FAQs and community responses showed similar patterns in providing information, emotional support, and experience sharing. Second, the evidence and authority signaling dimension (Figure 4B) showed a significant difference (χ²₂=15.85; P<.001). This means that expert-generated FAQs present research evidence and professional authority more clearly, whereas community responses often include fewer expert signals and more personal opinions than the FAQs. Third, for the outcome framing dimension (Figure 4C), the chi-square test showed no significant difference (χ²₃=2.83; P=.42). Both groups had similar distributions of gain-, risk-, or balance-focused framing outcomes. Finally, the behavioral actionability dimension (Figure 4D) showed a significant difference (χ²₃=11.68; P<.001).

This study examined the extent to which existing provider FAQ on ADHD content addressed the information needs reflected in consumer-generated questions collected from an online community. The findings revealed meaningful gaps between the questions consumers asked and the information provided in the provider FAQ, indicating that some information needs remained insufficiently addressed. In addition, this study demonstrated the feasibility of using an embedding-based approach to systematically assess the extent to which existing FAQ content addresses consumers’ information needs. By combining automated semantic matching with manually validated, model-specific thresholds, the proposed approach offers a scalable means of identifying potentially unmet information needs in large-text datasets.

The provider-generated ADHD FAQ addressed only 52.09% of questions from online communities, leaving 47.91% of consumer needs unmet. Most online community questions were related to only 18 of the 88 FAQ items, with the remaining 70 contributing minimally to overall coverage. These findings suggest that although FAQs are intended for consumers, they only partially meet consumer needs. Therefore, rather than maintaining a large number of FAQs based on existing materials, it is advisable to adopt an information provision strategy that improves FAQ relevance by incorporating topics aligned with consumer needs through continuous monitoring. The proposed directions for information restructuring are as follows.

Health information on ADHD should transition from a disease-centric perspective to one that more effectively supports real-life decision-making. Table 3 outlines fundamental, standardized topics such as diagnosis, treatment, prognosis, parental roles, school adaptation, and genetics. Conversely, Table 5 reveals consumers’ unmet informational needs, which encompass more practical and comprehensive issues. These include challenges in accessing diagnostic institutions, uncertainty about diagnostic reliability, financial burden of testing and treatment, concerns about disadvantages stemming from diagnostic history, anxiety regarding side effects, insufficient information on nonpharmacological alternative therapies, and management of comorbidities. Consumers in online communities expressed a desire for systematic information on access to care and associated costs, including ADHD diagnostic pathways, available medical institutions, estimated expenses, insurance coverage, and financial support systems.

Furthermore, practical strategies should be incorporated to assist consumers in applying medical knowledge for effective self-management. Given that children are the main group affected by ADHD, expanding the focus to encompass details about assistance from educational institutions and the community is crucial. Providing information to ensure adolescents’ access to treatment is essential. We should address the demand for information on whether adolescents can independently access diagnosis and treatment due to social stigma and a lack of parental understanding. Additionally, information on social support systems is necessary. For adolescents facing challenges such as parental opposition to treatment or lack of symptom recognition, accessible counseling services should be offered, along with age-appropriate procedures and support. This highlights the necessity of providing information capable of addressing intricate and context-dependent inquiries about issues such as diagnostic delays, academic assistance, and reductions in daily capabilities, among other matters. Furthermore, concerning adult ADHD, there is a requirement for information on employment limitations due to diagnosis, management of social maladjustment, and treatment strategies for late diagnosis.

Given the concerns about drug side effects, which are seen as a barrier to treatment, and the growing interest in nonpharmacological interventions, such as traditional Korean medicine, play therapy, and social skills training, it is crucial for health information to thoroughly cover the effectiveness, evidence, and appropriate use of these treatments. Referring to the actionability domain results in Figure 4, if FAQ responses—already known for providing actionable advice from official sources—could further develop the expertise of relevant organizations and broaden the information to include formal processes and reliable social support services, it is anticipated that consumers would receive practical information that aligns with their needs.

Automation techniques are necessary to rapidly identify consumers’ health information needs and provide customized information. However, this suggests that approaches involving expert validation are essential. The 3 embedding models used in this study (MBERT, LaBSE, and mE5L) each display different similarity distributions and characteristics, and manual validation confirmed that it is inappropriate to apply the same absolute similarity threshold across models. When 2 experts reviewed the similarity determinations produced by the embedding models, they found that clear judgments could not be made in the following cases: when there are overlapping words within sentences but the intent of the question is not clearly captured (eg, when the same symptom is described but the distinction between pediatric and adult cases differs); when it is the same case but the subject seeking the answer is different (eg, answers differ depending on whether the parent or the child is asking about the same issue); when the key point is simple yet accompanied by excessive explanation; or when names of specific times or unique locations are included. In such cases, it was difficult to rely solely on the results of similarity analysis. Therefore, to rapidly analyze consumers’ unmet needs embedded in large datasets, selecting the optimal similarity threshold through manual validation based on the results of the embedding model’s similarity analysis is an important process in the pipeline for analyzing consumer information needs.

The findings of this study emphasize the existence of various informational gaps at the social and psychological levels, including systemic factors related to ADHD diagnosis and treatment, social stigma, and the disadvantages associated with being diagnosed. These findings align with the need to address stigma and unmet needs [14], correct misconceptions for timely diagnosis and treatment [15], and provide accessible and empathetic online information [16]. The experiences of individuals describing delayed diagnosis due to social stigma, parental conflicts, and social disengagement support previous conclusions [58] that educating people to dispel misconceptions about ADHD can reduce psychological distress and encourage timely diagnosis and treatment. When individuals elaborate on their situations in their inquiries or use phrases such as “I'm struggling,” “I’m so upset. It’s because I raised it wrong,” or “I'm not the only one like this, right?,” it suggests a need for information that provides emotional support alongside factual details. For those expressing such sentiments, responses from the online community group that acknowledge the emotional state of the questioner and show empathy, such as “You must be very worried about your child’s symptoms,” “I hope this answer helps ease your concerns,” and “I fully understand the difficulties you are facing as a parent,” can enhance the effectiveness of the information provided. Furthermore, it was identified that there is a necessity to include details about the diagnostic process, treatment protocol, and multidisciplinary management of common comorbid conditions such as depression and anxiety that frequently occur with ADHD [59,60]. In particular, in cases where individuals have negative views of pharmacological treatment, it is crucial to offer suitable information about ADHD treatments along with explanations of situations where medication is necessary.

Additionally, prior work has verified that consumers generally trust health care professionals because of their roles as physicians, and physicians use this trust to spread health information [61,62]. In light of this, we recognized the importance of being cautious about exposure to marketing information rather than information about care services among consumers in the online community. In the context of the evidence and authority signals shown in Figure 4, while the FAQ provider group strives to enhance their credibility through research evidence, the online community’s response group portrays themselves as physicians or Korean traditional medicine practitioners and subtly mentions the hospitals where they practice, thereby attracting consumer interest.

This study had several limitations. First, the analysis used data from a single large Korean online community. Although this platform offers extensive information on ADHD, the findings may not comprehensively reflect the information needs of other communities, platforms, and cultural contexts. Second, although structural imbalances in existing FAQs were identified and directions for content restructuring were proposed, user-based evaluations are necessary to verify whether the revised information effectively meets consumer needs. Future work will involve restructuring health information and evaluating consumer satisfaction based on actual user assessment. Additionally, a pipeline should be established to systematically identify consumer health information needs.

The ADHD FAQs created by institutional information providers covered half of the semantics of consumer questions posed in online communities. FAQs were concentrated within a limited number of FAQs, and many failed to address consumers’ unmet health needs. Twelve unmet needs were identified, offering a more comprehensive and detailed perspective than the current scope of information provided. Furthermore, these research findings indicate that using model-based similarity analysis and LLM-assisted analysis allows for accurate and scalable measurement of how well provider-generated FAQs address consumer questions. This methodology also aids in identifying significant gaps in the information provided. The restructured information produced through this approach is expected to enhance the clarity, accessibility, and relevance of FAQs, thereby more effectively addressing consumer needs.