This was a retrospective study that involved analyzing texts collected from social media platforms and online forums. The study proceeded through 4 primary methodological phases: (1) collecting textual data, (2) cleaning and processing the extracted texts, (3) performing and optimizing topic modeling, and (4) conducting thematic analysis. The study process workflow is illustrated in Figure 1.

This study exclusively utilizes anonymized data from publicly available sources. The quotes presented in this report are modified or paraphrased to prevent any potential identification and direct linkage to the original posts/quotes. All identifiers or pseudo identifiers have been removed from the report. This study was approved by the Survey and Behavioral Research Ethics Committee of the Chinese University of Hong Kong (reference SBRE-20-864).

For this study, textual data were collected from online forums and social media platforms. For consistency and minimal translation concerns, only English-language texts were included. The posts from these platforms dated before November 15, 2022, were included, with varying start dates. For social media platforms, posts were extracted from Facebook, Twitter, and Reddit. These platforms were chosen for their popularity and the substantial textual content they contain [19]. For cancer-related online forums, the selection criteria have been detailed in a previous study [20]. To describe briefly, we first used the top search engines (Google, Microsoft Bing, Yahoo, and Yandex) to search for health forums by using different combinations of search terms (cancer, tumor/tumour, patient, and forum) [21]. The search results were extracted and examined by the investigators. Web forums were included if they (1) appeared in more than one search result, (2) were open access (ie, no membership or passwords were needed to access the messages), (3) were active for at least the past 5 years, (4) had at least 10 messages posted to the group within the past 30 days from the date of the search, and (5) enabled web scraping of posts or feeds using Python (Python Software Foundation) or R (R Foundation for Statistical Computing). Finally, 16 different online cancer-related forums, including Cancer Chat, Cancer Survivors Network, Cancer Council Online Community, and other forums, were included for the analysis (Table 1). Search terms related to immunotherapy were identified from multiple authoritative websites, including the American Cancer Society, Cancer Research UK, and the National Cancer Institute (Table S1 in Multimedia Appendix 1). Site-provided search engines were utilized to identify related posts.

Spark NLP, a natural language processing library for Python built on top of Apache Spark, was utilized to preprocess the textual data [38]. The texts extracted from the forums and social media sites were subjected to a thorough cleaning process to enhance their suitability for further analysis. The cleaning process involved multiple steps, including checking and correcting spelling errors, lemmatization, removal of stop words, and elimination of short entries (Figure 1). Each text message was first spell-checked, and the vocabulary was reduced to its base root form by using Spark NLP. After that, each word vocabulary was reduced to alphanumeric characters that contained only letters from “a to z,” “A to Z,” and “0 to 9.” For a precise and semantically meaningful text corpus, any message composed of fewer than 4 words as well as commonly occurring words (eg, the, a, and, in) that carried little or no meaning in the message were removed. Finally, the cleaned text messages were processed using the medical language models from Healthcare Spark NLP, a pretrained pipeline to recognize medical terminologies within the messages [39].

The next step involved the extraction of topics from the data. Topic modeling, an unsupervised machine-learning approach, was employed to unveil hidden semantic structures and extract distinct topics from the extracted textual data [40]. In this study, we first evaluated 10 common topic modeling techniques, namely, Latent Dirichlet Allocation, Latent Semantic Analysis, Non-Negative Matrix Factorization, Principal Component Analysis, Random Project, K-Means, Top2Vec, BERT + K-Means, BERT, and Latent Dirichlet Allocation + BERT to determine the appropriate topic modeling approach. Topic diversity (indicating the model’s ability to capture differences between generated topics), coherence (the frequency of descriptive words of the topic within each cluster), and quality (topic diversity multiplied by topic coherence) were used as metrics to evaluate the effectiveness of isolating topics from the texts [41,42]. Finally, we used BERTopic to extract quality topics from the data based on the modelling performance. All texts were passed through the BERT model to create embeddings that converted text messages into numerical representations by using BERTopic. The dimensions of the embedding vectors were reduced due to the “Curse of Dimensionality” [43]. The bisecting K-means algorithm was used to group the reduced-dimension embedding vectors into clusters until optimal silhouette scores were achieved [44]. Various numbers of topics, ranging from 5 to 25, were tested to determine the optimal metrics.

Following the identification of clusters with the highest topic quality, the 20 most frequent words from each topic were extracted. BERTopic modeling was conducted using the Worcester Polytechnic Institute’s Turing cluster, which has high computing power for large volumes of data processing by using multiple physical computers simultaneously. The Turing cluster consists of a 4-node hyperconverged head that controls 79 compute nodes and is located at Worcester Polytechnic Institute. The total central processing unit/random access memory/graphics processing unit counts across all computer nodes were 5224, 49 TB, and 84, respectively [45].

Sentiment analysis was performed using 2 popular libraries, VADER (Valence Aware Dictionary and Sentiment Reasoner) and TextBlob [46,47]. Both libraries were used to enhance the accuracy of the results and minimize potential biases that may arise from relying on a single library. The sentiment scores obtained from each library were then averaged to derive a final sentiment score, which ranged from –1 (indicating a very negative sentiment) to 1 (indicating a very positive sentiment). Based on the topics generated through BERTopic modeling, thematic analysis was conducted by 2 investigators (XW and CSL) to identify themes associated with immunotherapy. The thematic analysis in this study involved the following steps: familiarization with the posts, inductive analysis, theme generation, and theme review among domain experts in the research team (including oncologists and oncology pharmacists) [48]. First, 2 investigators (XW and CSL) familiarized with the posts and encoded them independently. Then, the codes were compared, the same codes were merged, and any discrepancies were addressed through discussion with a third investigator (YTC). The themes were generated through the identification and refinement of the codes, which were grouped by perceived thematic similarities. The final themes were identified by achieving consensus among domain experts. The consensus process involved multiple rounds. The experts rated their level of agreement for each theme: themes with 75% agreement were retained, and themes without consensus were discussed, adjusted, and rated in the next round until 75% agreement was reached.

A total of 4.9 million posts were gathered from online forums and social media platforms. After data cleaning, 3.6 million posts remained for further analysis (Figure 1).

In comparison with other topic modeling techniques, BERTopic demonstrated the most satisfactory performance, with a topic diversity score of 81.11%, a topic coherence score of 80.48%, and a topic quality score of 65.28% (Table S2 in Multimedia Appendix 1). Table S3 in Multimedia Appendix 1 presents BERTopic’s performance across different numbers of topics to determine the optimal value. Fifteen topics achieved the highest overall topic quality, scoring 70.47% (with the highest topic diversity at 87.86% and a high topic coherence of 80.21%). We also used these metrics to assess the quality of individual topics generated by the model, yielding topic quality scores ranging from 78% to 88% (Figure S1 in Multimedia Appendix 2). One topic (Topic 15) was excluded due to the mismatch between its extracted words and the topics related to immunotherapy. Based on the sentiment analysis results, the average sentiment score of around 0.3 across the 14 topics preliminarily suggested generally positive sentiments toward immunotherapy within the online community (Figure 2). A more comprehensive interpretation of the specific topics is presented in the subsequent sections.

A total of 6 themes were identified (Table 2). The quotes in Table 2 have been modified to prevent any potential identification and direct linkage to the original posts. Theme 1 centered on the feelings of hope and positivity among patients treated with immunotherapy. This theme highlighted the belief that immunotherapy offers hope for patients with cancer to lead normal lives and provides a potential cure for cancer with minimal side effects. Theme 2 described immunotherapy as a more effective treatment option after other treatments have failed. Patients may have tried other cancer treatments such as chemotherapy and radiotherapy but only found successful results with immunotherapy. Besides, patients believe combining immunotherapy with other conventional therapies, particularly chemotherapy, can lead to higher efficacy in treating cancer than using a single treatment approach. Theme 3 suggested that patients receiving immunotherapy also explore diets or complementary medicines to manage their symptoms or side effects from treatment or the cancer itself. For example, patients may employ mind-body practices such as massage and meditation or use herbal medicines to alleviate certain physical and psychological side effects caused by immunotherapy. Probiotics were frequently mentioned by users as a potential way to boost the effectiveness of immunotherapy. In Theme 4, the potential ineffectiveness, worse outcomes, and high cost of treatment were frequently mentioned among patients, which acted as hindrances to the use of immunotherapy. Theme 5 suggested that immunotherapy can disrupt patients’ normal daily lives, particularly for those enrolled in immunotherapy trials. Additionally, the treatment may negatively impact their activities of daily living, such as taste changes, dietary restrictions, and time consumption. Theme 6 indicated that immunotherapy may cause physical and psychological symptoms and side effects such as rash, inflammation, and nephrotoxicity. These adverse effects may sometimes pose greater challenges compared to those arising from other treatments or the cancer itself. Some patients acknowledged that these symptoms may be unavoidable alongside the curative effects of immunotherapy.

We analyzed a large dataset comprising 3.6 million posts from 3 social media platforms and 16 online cancer forums, providing a comprehensive overview of perceptions regarding immunotherapy within English-speaking online cancer communities. A cutting-edge natural language processing technique, BERTopic, was used to generate interpretable topics, ensuring robust data analysis. The identification of different themes underscores the diverse attitudes toward immunotherapy across various aspects. The posts show the positive attitudes of the patients toward immunotherapy and its perceived effectiveness and benefits. However, there are also concerns about the side effects of immunotherapy and potential disruptions to patients’ lifestyle. These themes may offer valuable insights into the facilitators and barriers to the adoption of immunotherapy. In addition, the results showed that using machine learning techniques to analyze patient-generated online textual data can enhance our understanding of perspectives on newer cancer treatments, and such information may facilitate communication between patients and clinicians when discussing immunotherapy as a therapeutic option.

The positive factors that encourage the utilization of immunotherapy included its comparatively mild side effect profile and the potential to instill hope in patients by offering a viable treatment option when other methods have failed. Our findings also indicated that online cancer communities seem to be well-informed, with some patients well-versed in the benefits of immunotherapy. Laypeople now access medical information on the internet independently and acquire firsthand information about advances in cancer treatment. Access to online health information assists patients in making health-related decisions, leading to more professional consultation, improved treatment compliance, and better self-care [49]. However, the internet should not be viewed as a replacement for professional health information sources [50]. The guidance and advice of health care professionals remain indispensable in helping patients make complex medical decisions [51]. Online health information cannot provide detailed diagnoses or personalized treatment plans for patients; moreover, the presence of misinformation on websites might mislead patients about treatment [52]. Taken together, our results highlight the rising health literacy of online cancer communities, especially those involving younger generations. Future efforts should focus on enhancing the accuracy of online health information by using reliable rating tools, effective search engine ranking, and progress in crowdsourcing websites.

Although immunotherapy has shown clinical benefits in several trials, there are still potential barriers to its utilization. The high cost associated with immunotherapy emerges as a frequently discussed concern among patients with cancer, emphasizing the need for revising drug valuations and reimbursement models [53]. Many quotes were related to patients’ descriptions of their physical symptoms induced by immunotherapy and how it has impeded their activities of daily living and disrupted their normal lives. In addition to physical irAEs, the long-term use of immunotherapy has also been associated with psychiatric side effects such as fatigue, insomnia, anxiety, and depression [54,55]. Hence, it is important for oncology practitioners to manage patients’ expectations and communicate with patients regarding the potential side effects of immunotherapy before initiating treatment. This proactive approach allows patients to better understand the benefits and risks of the treatment, which may prevent them from discontinuing treatment because of known adverse reactions.

Several clinical practice guidelines have been proposed to address the evaluation and management of immunotherapy-induced toxicity [56] and the diagnosis and treatment of toxicity associated with immune checkpoint inhibitor therapy in specific organ systems [57]. The provision of high-quality supportive care from a multidisciplinary team of health care professionals can help identify and manage irAEs [58]. In practice, patients may turn to self-management strategies to cope with these symptoms. We found that some patients mentioned in online posts that they used traditional, complementary, and integrative medicine modalities such as probiotics, herbal remedies, and meditation to self-manage the toxicity induced by immunotherapy. Recently, some studies have evaluated how complementary medicines may enhance the efficacy of immunotherapy and alleviate immune toxicity [59,60]. For example, the treatment of irAEs often necessitates the long-term use of high-dose corticosteroids [61]. Integrating complementary modalities into the management of irAEs may reduce corticosteroid use. Our group has previously reported that patients with cancer frequently seek advice on the use of complementary medicines on social media platforms [20]. At this juncture, clinicians should focus on initiating effective communication regarding the use of complementary medicines and help patients establish realistic expectations while being well-informed about the limited evidence supporting these approaches.

Our study has several limitations. First, the data used in this research were collected from English-language online cancer communities, which may limit the generalizability of the findings to data posted in other languages. Social media and cancer forums may be used distinctly in different countries, and the types of immunotherapeutic approaches may also vary among regions. Future studies could conduct a thorough investigation into whether distinct concerns regarding immunotherapy exist among patients from different countries and regions of practice. Second, perceptions regarding immunotherapy on social media and cancer forums are dynamic and may be influenced by current news or events. Therefore, future research can consider exploring the fluctuations in posts on these platforms over varying time periods. Third, although the performance metrics (topic diversity and topic coherence scores) quantified from our models suggest that BERTopic is effective at unveiling latent patterns within large datasets, this approach requires validation. Nevertheless, our study demonstrates that applying artificial intelligence to analyze social media data can facilitate our understanding of patients’ perspectives as well as the effectiveness of machine-learning techniques in processing substantial volumes of data retrieved from social media platforms and online forums.

This study captures comprehensive insights and perspectives of online communities regarding immunotherapy through BERTopic modeling. The identified themes and topics may facilitate the understanding of immunotherapy utilization and serve as valuable supporting information when making treatment decisions. The posts have shown the positive attitudes of the patients and their perceived effectiveness and benefits of immunotherapy. Although immunotherapy presents a beacon of hope and a viable treatment option, its side effects and the accompanying potential lifestyle disruptions cannot be ignored and are major concerns perceived by the online communities. It is essential for clinicians to inform patients of the undesirable effects of immunotherapy and establish realistic expectations, thereby enabling patients to make an informed decision considering the benefits and risks associated with the treatment, which may ultimately lead to more optimal treatment outcomes with immunotherapy.