We acquired data on discussions of retinoblastoma from Twitter. We accessed Twitter through its official API services [30]. We retrieved posts and subsequent discussions between January 1, 2011, and September 30, 2022. We searched for discussions relevant to retinoblastoma and imposed no limitations on language and origin. The search was conducted using the terms “retinoblastoma,” “enucleation,” or “eye cancer” and adapted to the search format for individual platforms. The search was conducted in the 6 official UN languages—Arabic, Chinese, English, French, Russian, and Spanish.

We employed the DeepL translator service to translate the texts into English from their original language [31]. Manual validation was performed with 1000 sampled non-English posts to ensure accurate translation. The translation performance of the DeepL translator was further supported by a previous benchmarking study, demonstrating outstanding accuracy and outperforming other existing solutions [32]. We excluded irrelevant records, for example, those related to nonretinoblastoma eye malignancies and enucleation for nonretinoblastoma causes, using deep learning models. We fine-tuned a BERT (Bidirectional Encoder Representations from Transformers) model to filter texts from Twitter. The training (n=4000) and validation datasets (n=1000) were randomly sampled from the entire dataset and interpreted by ESW and RWC. Texts were labeled as relevant or irrelevant, and discrepancies were resolved by consensus. We fine-tuned the base BERT model on English texts using 10-fold cross-validation and validated on the holdout set [33,34].

We performed topic modeling on the resultant relevant records stratified by platforms. Texts were cleaned by removing hyperlinks, hashtags, and user mentions and were further stemmed and lemmatized. Using the automated hierarchical Dirichlet process (HDP), we generated keywords that explained the topic distribution, which were sampled automatically from the input text word distribution [35]. Sensitivity analysis was performed using BERTopic, an embedding-based method. BERTopic uses embedding models to capture the semantic relationship for posts. Compared to HDP, it might be more resilient to short text with noise but requires higher computational power. On the contrary, HDP adopts a probabilistic Bayesian nonparametric approach. Apart from requiring less computational power, an essential strength of HDP is its ability to automatically infer the number of topics from a given dataset without prior assumptions. Hence, we compared the output from both models in this study.

Topics identified from both models were fused under several criteria. These included (1) topic keywords overlap, (2) semantic coherence and thematic redundancy, and (3) hierarchical proximity in the HDP model. For each topic, we computed the Jaccard similarity coefficient between the top 15 words for each topic pair, and those with a similarity score above 0.3 were highlighted for a manual review. In addition, topics with conceptual similarity were also reviewed manually, even if the similarity score was lower than the threshold, as semantic similarity may not be fully represented by token frequency similarity. Topic fusion was conducted independently by 2 authors and reviewed by a third author in case of discrepancies.

Similar to filtering posts, we fine-tuned a BERT model to classify posts into positive, neutral, and negative categories based on users’ acceptance of disease management. For example, posts representing the acceptance of treatment modalities while expressing sorrow for a diagnosis were labeled as positive [36]. The prediction probability was used to reflect the strength of emotions, that is, how strongly the post belonged to the respective category. We implemented 10-fold cross-validation and tested combinations of learning rate, batch size, and epoch. We calculated the strength of negative sentiment toward different treatment modalities for further analysis.

We fine-tuned a BERT model to classify posts by patient roles, that is, whether the post is posted by patients and their direct relatives or by general users. Similar to the above training approach, we labeled 1000 posts and fine-tuned the base BERT model using 10-fold cross-validation.

As only Twitter provided the detailed data on time, conversation details, and geolocation with a broad global user base among the platforms, we performed subsequent global comparisons using Twitter data only.

A fraction of posts was initially geotagged, that is, provided with the country or region of origin. We extracted their likely location for the other posts using the author’s user profile, with geotagging [37] and subsequent geocoding with the OpenStreetMap API [38]. The current approach has been validated by prior studies, showing high accuracy and reliable spatial granularity [39,40].

Users’ occupations were analyzed using their user profiles through the named entity recognition technique. Descriptions were processed using the default Stanford CoreNLP (version 4.5.1) model [41]. We manually labeled 2000 profiles to benchmark the extraction algorithm’s accuracy, sensitivity, and specificity. Subsequently, we manually processed the identified titles and aggregated them into different sectors, including clinical medicine, education and academia, media outlets and creative media, religion, politics, and public administration, and mapped them back to individual users.

Network analysis was conducted using Gephi [42]. After identifying isolated communities and nodes, we removed them to focus on the largest connected component of the network to investigate the functional component of the online discussion. In addition, this avoids complicating the calculation of path distance as disconnected nodes represent infinite path distance. We calculated network density, centrality measures—including eigenvector centrality—modularity, average path length, and other network properties for the discussion network and performed community detection. Eigenvector centrality in particular measures how influential a user is to the social network by accounting for both the number of connections and the number of important nodes that one is connected to. Therefore, we used eigenvector centrality to select those with high centrality scores for subsequent analysis [43].

We acquired the OS rate and GS rate for retinoblastoma from our group’s previous systematic review and meta-analysis. This study analyzed clinical outcomes by country or region from 2010 to 2021 [7].

Variables used for global comparisons were first aggregated to the country level by taking averages. These included sentiment strength for different topics and clinical outcomes including OS and GS. We computed associations between the sentiment distribution and clinical outcomes (ie, OS and GS) using multivariable linear regression, adjusting for the Human Development Index, a statistical composite index covering socioeconomic development, health infrastructure and outcomes, and standard of living. The subgroup analysis was conducted by repeating multivariable linear regression with data for each region. Similarly, the associations between the participation rate of various professions on social media and the resultant discussion sentiments and intensity were also analyzed using multivariable linear regression. Linear model assumptions, including normality, homoscedasticity, independence, and linearity, were tested using the Shapiro-Wilk test, Breusch-Pagan test, Durbin-Watson test, and Rainbow test, respectively. We performed multiple testing adjustments with sequential Bonferroni correction in R (Bell Laboratories) to the reported statistics with a balanced focus on reducing the false discovery rate and preserving statistical power. The analysis was conducted in Python 3.9.0 and R 4.2.2 for the complete study. The complete data flow, detailed choice of models, and configurations from the above steps are shown in Figure S1 in Multimedia Appendix 2.

This study was approved by the Joint Chinese University of Hong Kong–New Territories East Cluster Clinical Research Ethics Committee (reference number 2025.559). This study involved only the secondary analysis of publicly available social media data, with no attempt to reidentify individuals. The extracted data were obtained via Twitter Academic API with official approval and stored in encrypted storage. Only aggregated data were reported.

To our knowledge, this is the first study to characterize the role of social media in global discussions of retinoblastoma. Our study provides a comprehensive picture of the real-world application of social media for public education and support of retinoblastoma, with discussions ranging from the promotion of retinoblastoma awareness campaigns to specific discussions on treatment and patient support groups. Being a rare disease, the wide base of direct user data provided by a global social media network like Twitter enables the exploration of subjective factors such as cultural perspectives, social viewpoints, the practice of alternative medicine, gender bias, etc, on retinoblastoma treatment acceptance and outcomes at a population-based level. This study extends the message from earlier publications on the impact of social media on pediatric cancer and rare diseases and further substantiates the context for its use in retinoblastoma. Building on the published approaches from previous social media studies, we provided an integrated pipeline for the systematic and structural analysis of social media platforms that can be readily used for various conditions [40,45-49]. Our study examined the association between discussion dynamics and the real-world clinical outcomes of retinoblastoma. Locations with generally negative sentiment toward retinoblastoma in parts of Asia and Africa correlated with zones of inferior survival rates, as identified in our earlier meta-analysis [7]. Negative sentiment and lack of education toward disease diagnosis and treatment contributed to delayed diagnosis and predisposed patients to the advanced clinical stage of retinoblastoma at presentation, often leading to limited treatment options and resulting in unsatisfactory GS and OS rates in these regions [5,6,9,11]. Other important factors such as health care accessibility also contributed to clinical outcomes across the world. The vast amount of user-generated content also allowed us to probe into the perceived barriers behind this negativity, which potentially resulted in delayed presentation or impeded discussions and understanding of retinoblastoma. While negative post sentiment does not directly cause poor clinical outcomes, it served as a proxy for understanding the underlying predisposing factors that inevitably aggravated the challenges to survival and surgical success.

We observed an association between a stronger negative sentiment toward enucleation and a lower OS rate for retinoblastoma patients worldwide. A large proportion of posts focused on the cost of treatment and rehabilitation related to vision loss. This concern was especially profound in resource-limited localities but also notably impacted resource-rich locations such as Western countries. Enucleation remains an important treatment modality for advanced retinoblastoma, with delayed treatment leading to metastasis and mortality. Alongside other clinical and socioeconomic factors that affect retinoblastoma outcomes, increased enucleation hesitancy may hence partly contribute to worse survival rates. While for the general population there might be well-established health care infrastructure to cover most patients’ needs, there were still unmet individual needs for patients, particularly for enucleation, which are associated with long-term visual impairment, socket complications, and required continual specialist support [50]. Various stakeholders, including governments and nongovernmental organizations, have been working rigorously to reinforce safety nets for retinoblastoma patients [51-53]. However, the current observation on Twitter may underpin a need to match patients with such resources to enhance use and alleviate patients’ burdens. The rising trend of the lack of patient support after diagnosis observed post-2019 overlaps with the COVID-19 pandemic, with frequent closures of treatment centers and delays in follow-up appointments. This might have affected the management of retinoblastoma patients. In the post–COVID-19 era, reversing such trends and perceptions would be paramount [54-56]. In a study in India, COVID-19–related full territory lockdowns were related to delayed and interrupted treatments, temporary or even permanent defaults on follow-ups, and mortality. Specifically, 60% of active cases prior to lockdown were shown to have worsened; 12 (5%) eyes were enucleated due to poor response to restarted treatment post default [54]. In a separate systematic review, care delivery was found to be particularly disrupted in lower- to middle-income countries compared to higher-income ones. Staff shortages, interrupted supply chains, and restricted accessibility to treatment facilities were among the most worrying situations [57]. Both online and offline interventions should hence be attempted to maximize the reach of such messages.

Other reasons for enucleation hesitancy, including the unpreparedness of parents and lack of disease understanding, should also be addressed in systemic and multilevel ways. Such causes may be due to inadequate counseling by health care workers upon diagnosis, denial of diagnosis, and sex stereotyping and discrimination against girls. Parent education is of paramount importance in retinoblastoma as parents are usually the ones who provide consent for treatment options. Clinical studies have shown the propensity for parents to develop anxiety and depression when their child is diagnosed with retinoblastoma, and support for their psychosocial needs is necessary for the best benefit of the patient [58]. On the other hand, parents should also be educated on the presentation, screening, and testing for retinoblastoma [59]. Greater knowledge of the disease has been associated with earlier diagnosis. On the contrary, parental shopping for alternative medical advice may result in disease progression and delay in diagnosis and treatment [60]. Gender stereotypes and sex bias, despite not being the main reason for treatment delays, still contribute to the constellation of concerns that aggravated enucleation hesitancy. Our prior study noted that sex ratios of retinoblastoma cases were significantly higher in some parts of Asia and Africa, despite no sex predilection for retinoblastoma. (1.30 vs 1.14, P=.04). Differences in gender health and mortality in South Asian countries have been identified previously, with inequality in access to care noted as a plausible explanation [10,61-63]. Discrimination and neglect of females for treatment due to sex-based referral bias have been reported in Nepal, Bangladesh, Pakistan, and China [10,64-69]. Our user-generated contents from Twitter show persistent yet declining intensity in sentiment related to sex inequality for girls over the years. Our data also demonstrated geospatial trends in the pattern of enucleation hesitancy, which allow for multilevel targeted health care interventions by trained personnel to improve acceptance in vulnerable zones and outcomes. Matching patient needs with existing resources can optimize resource usage and facilitate improvements in retinoblastoma outcomes on both regional and global scales.

Social media also proves to be an essential platform for health care workers to disseminate accurate information. In an editorial titled “Oncologists in social media—what are the limits?” from The Lancet Oncology, it was highlighted that despite their importance and effectiveness in information dissemination, social media sites may facilitate the dissemination of misinformation and pose a mental health burden. It is essential for health care professionals to use these sites effectively to benefit daily clinical practice [70]. A significant rise in the number of posts toward treatment hesitancy was observed in 2016. This corresponds to a major new development and renewed interest in GS therapies, including intra-arterial chemotherapy and intravitreal chemotherapy, the introduction of which greatly improved GS and OS rates, especially in developed countries. Unfortunately, our study showed that only a very small fraction of online information currently originates from health care workers. The interaction quantity of posts by health care workers is overshadowed by those from media or political outlets. Not only has this limited the potential benefits for resolving treatment hesitancy and introducing newer technology that could improve treatment outcomes, but it has also allowed the circulation of alternative information with variable levels of accuracy and reliability. Increased participation by local health care agencies, particularly in regions with higher treatment hesitancy, would be beneficial in generating locally appealing contents for treatment and disease acceptance. The reach can be further improved with cooperation and assistance from media outlets, creative media creators, government agencies, and nongovernmental organizations. Of note is that health advocacy on social media sites may lead to the establishment of doctor-patient relationships and carry duty-of-care concerns. We suggest that rather than providing consultation and advice on a case-by-case basis on social media platforms, agencies, whether government or nongovernment, should standardize critical information based on proven facts or the latest research about the disease that can be conveniently accessible to medical and patient support groups. The goals would be to establish factful discussions online and promote awareness of retinoblastoma. Training sessions could be provided to interested health care professionals to improve content delivery and communication strategies specific to social media platforms and techniques for mitigating misinformation as well as tackling legal and ethical concerns. Public health agencies shall also increase coverage and research in social media spaces and provide timely action if misinformation arises while maintaining high levels of disease awareness and health literacy. With the low latency and cross-border nature of social media sites, verified and translated content can be developed to target vulnerable regions identified via research, together with surveys for pinpointing patients in need. This would be particularly important for localities with fragmented health care systems and limited resources. Alongside existing international collaborations, this could provide a high-level intervention to improve retinoblastoma outcomes.

There are limitations to this study. First, although we have analyzed a large volume of social media posts and correlated them with global data from a large cohort, the ecological study approach restricted the interpretation of results to the population level. It is important to note that the identified relationships represent associations and should not be interpreted as direct causative effects. Second, due to privacy and ethical concerns, we were not able to individually verify the authenticity and background of user profiles for sentiment analysis and progression, limiting the generalizability of sentiment analysis and its correlation with underlying disease status and treatment acceptance. Third, due to local variation in the popularity of social media platforms and the lack of publicly available data for platforms such as Facebook, Weibo, and Line, we were not able to exhaust all the information online. It is essential to compare results from multiple platforms and incorporate the analysis of popular local platforms to integrate findings in this study into local practice. Moreover, this study included posts in 6 official UN languages. This might, however, introduce selection bias against populations in Asia and South America where the official languages were not covered. The results may not fully represent the general population in those locations and may be skewed toward English-speaking or Western-educated populations in these localities. Finally, the extracted data were in multiple languages, and translation was necessary for further comparative analysis. Although we used a validated translation tool for processing, there might still be potential nuances in the cultural context or linguistic subtleties lost in translation, affecting the accuracy of sentiment analysis. Future studies may collaborate with popular social media sites, following the proven importance of social media in disease management, to provide a more comprehensive analysis of health information distribution online. Validation studies may also check the accuracy of profiles and contents mentioned by users online. Moreover, a slight time discrepancy existed between the sentiment data and the clinical data. Sentiment data were collected in September 2022, with clinical outcomes collected only up to October 2021. This short mismatch was noted and may affect the reported association if clinical outcomes from October 2021 to September 2022, that is, the period without clinical outcomes report, were notably different from earlier periods. There may be a possibility of noncausal noise in the reported association accordingly. Furthermore, despite our efforts to include posts from global users and cover parts of Africa and Asia, ecological fallacies may persist. In particular, Twitter users in these regions may represent a socioeconomically distinct population, including those living in urban areas or those who are socioeconomically privileged. These users may not fully represent the general population reflected in the clinical survival data, resulting in ecological fallacy or skewed representation. Due to privacy settings and the inherent nature of social networking sites, geotagging was only successful for 16.4% of identified users. There might hence be underrepresentation or biased association estimates for certain regions. Ecological fallacy may also arise as associations were computed on country-level aggregates and may not represent associations at the individual level. For subgroup analysis, since there are a limited number of countries in each subset, the multivariable analysis with adjustment for the development index may introduce severe overfitting. The results only aimed to supplement the global association and to reflect potential regional differences with additional caution regarding the differences in the development level. The association should hence be interpreted with caution. In addition, despite the significant associations between treatment sentiment and clinical outcomes, the adjusted R-squared values range roughly between 0.316 and 0.599. The sentiment hence only partially explained the variance in clinical outcomes, and other contributing factors should not be overlooked. Despite the potential power of social media sites in providing timely global insights for disease conditions, the ethical implications should not be overlooked. Apart from user privacy, users from regions with restricted social media or political protections may be underrepresented. Policy decisions that solely rest on social media insights may hence paradoxically aggravate such inequity.

Social media is extensively used for retinoblastoma education and patient support, particularly in North America and Western Europe. The improved understanding of parents and patients with retinoblastoma, as reflected by positive sentiment toward the disease and its treatment, and more engagement from clinicians, was associated with better OS and reduced enucleation hesitancy. Demonstrating geospatial trends in enucleation hesitancy allows for targeted information dissemination by health care workers to improve acceptance in vulnerable regions and reduce barriers to receiving timely treatment. Matching patient needs with existing resources can optimize resource usage and facilitate improvements in retinoblastoma outcomes on a regional and global scale.