IntroductionBackground

Breast cancer is 1 of the most common forms of cancer, with an estimated 2 billion people being affected worldwide in 2020 (according to statistics released by the World Health Organization [WHO]), and is consequently a disease familiar to many people. It is a chronic disease with a high mortality rate, which poses a serious threat to human life []. For this reason, breast cancer is often viewed negatively, and new diagnoses often trigger sadness, fear, and even psychopathological comorbidities, such as depression []. In recent decades, the number of new diagnoses has continued to rise, despite important improvements in medical technologies worldwide []. In China, more than 400,000 people were diagnosed with breast cancer in 2020, with approximately 100,000 deaths (according to WHO) []. Behind these diagnoses are numerous stories emerging from the experiences of patients or the people around them who are closely intertwined []. Therefore, it is not unusual for one to come across discussions on breast cancer in daily life—be it learning about the diagnosis of a loved one or acquaintance or coming across news on a celebrity with breast cancer or even struggling to accept the diagnosis of a close relative. Therefore, a lot of these breast cancer–related narratives take place on social media–lived experiences of people who may have been diagnosed with or who know of someone struggling with breast cancer.

Social media is indispensable in the daily life of billions worldwide; almost everyone is a user of a social media platform []. On these platforms, people can share snippets of their lives with other people around them, which double as autobiographical records of their life events. As a social tool, one can smoothly interact and communicate with one’s friends and family over the internet, be it synchronously or asynchronously [,]. Such activity leaves digital traces all over the internet, and researchers have since begun using social media posts as resources for uncovering social phenomena []. Particularly in the medical field, social media analyses have also been used to great effect, for example, in examining and predicting the epidemiological spread of infectious diseases, such as seasonal influenza and COVID-19 [,]. Recently, researchers have also analyzed social media to learn about the perspectives and needs of patients with certain diseases. For example, Kamba et al [] analyzed a Japanese social media forum (Yahoo Japan) for posts relating to breast cancer and found that the most frequently mentioned concerns pertain to symptoms, screening, and lack of knowledge, to name a few (see also Refs. [,]).

However, much of this research has been conducted on Western social media platforms, such as Twitter and Reddit, which have limited penetration in the Chinese market. Chinese internet users have their own social media ecosystems and platforms: Sina Weibo is 1 of the most widely used and popular social platforms in China and has been called by some as the “Chinese version of Twitter” []. Given our research interest in Chinese social media users, we focused our paper specifically on Weibo. As a widely used platform, the number of monthly active users reached 511 million in 2020; Weibo is known by almost everyone in China [], and posts are known to reflect the diversity of opinions and perspectives by everyday Chinese []. Often, users discuss and post about all kinds of topics on Weibo, including topics pertaining to breast cancer. With the large number of users and the diversity of content, Weibo data appear to be a valuable corpus for research on Chinese perspectives from the bottom-up.

Sentiment Analysis on Social Media

To accommodate the large volume of data on the internet, conventional methods, such as qualitative coding, may be too time-consuming and costly. Therefore, modern sociological researchers frequently use computational methods, such as sentiment analysis and topic modeling, to analyze the data. Originating from the field of natural language processing (NLP), sentiment analysis is optimized to deal with the detection and classification of sentiments in (a large number of) texts. By using sentiment analysis, we can infer whether a given text has a positive, negative, or more fine-grained emotional orientation in a given context []. In studying social media, researchers analyze the data on social media to obtain public perceptions on a specific topic in contribution to the study and advancement of society []. Some researchers have also applied sentiment analysis to measure customers’ needs from their social media posts, thereby obtaining unique insight to improve a brand’s products or services []. Researchers have also applied sentiment analysis on social media to predict mental health issues, for example, Wang et al [] used sentiment analysis to detect users with depression on social networking services.

Regarding breast cancer, sentiment analysis may play a more important role in exploring the patients’ psychological state, such as their perceptions, cognitions, and emotions []. Through analyses of tweet sentiments, previous research has confirmed that patients with breast cancer have different polarities (valence) of emotional expression for topics related to breast cancer []. For example, support seeking and treatments are associated with positive sentiment, but health care and insurance are associated with negative sentiment. Moreover, posters may not necessarily be patients themselves posting about their experiences or concerns but could be posting about a loved one, a relative, or an acquaintance with breast cancer. Accordingly, posters’ emotional expressions on social media may not only display differences in sentiment, depending on their specified content or aspects (eg, treatment stage or success), but also show differences, depending on their relationship with the patient [] or if the posters themselves are the patients. In this paper, we define this as the “poster-patient relationship.” Therefore, in studying the usage of social media for emotional expression in the context of breast cancer, we propose the necessity to distinguish the poster-patient relationships for each post—whether posts originate from patients themselves or from their friends and relatives or other people.

The Research

Before examining emotional expressions and sentiment, we intended to discern the relationships between poster and patient through the post. Due to the large volume of data, we turned to machine learning for this task. “Machine learning” is the term used to describe both the academic discipline and the collection of techniques that allow computers to undertake complex tasks, and recent advances in machine learning have driven advances in the development of NLP and artificial intelligence (AI) []. In NLP, the past 5 years have seen rapid advances in the transformer-based framework, resulting in cutting-edge pretrained language models, such as Bidirectional Encoder Representations from Transformers (BERT) [], Robustly Optimized BERT Pretraining Approach (RoBERTa) [], and Generative Pretrained Transformer (GPT)-3 [], which have greatly improved the effectiveness of downstream tasks (eg, text classification), opening up new avenues for researchers to study society and language [].

Our aim was to study how users on the Chinese social media platform Weibo post about breast cancer–related topics on social media. Although we took a hypothesis-blind, exploratory approach to data analysis, we focused our discussion on topics surrounding the issue of emotional expression by examining differences in emotional expression, depending on poster-patient relationships. In step 1, we collected data from Weibo and determined poster-patient relationships through 2 stages of classification: first, we identified whether a post references a patient with breast cancer (as opposed to posts that mention breast cancer without naming a specific patient), followed by the poster-patient relationship classification that determined the relationship between the mentioned patient and the author of the post (poster). Ultimately, these 2 stages in step 1 constituted a single classification pipeline to identify poster-patient relationships: whether the post authors are themselves the patients or (1) a family member (family_members); (2) a friend or relative (friends_relatives); (3) an acquaintance (acquaintances); (4) from a parasocial relationship, such as a celebrity or public figure (heard_relation); or (5) no patient mentioned (no_patient). In step 2, we used the LIWC-based dictionary to count the word frequency for each post, with 5 emotional categories (sadness, anger, anxiety, positive, and negative), thereby expanding our target beyond just positive and negative sentiments. Despite the lack of discreet positive emotion categories in the LIWC dictionary, we chose it because it is 1 of the most widely used and accessible sentiment dictionaries in psycholinguistic research. Next, we used topic modeling to further examine the main topics discussed between each class and how these topics differ across classes. This will allow us to see how social media narratives for patients and posters differ, while shedding light on possible implications for emotional expression via social media.

MethodsEthical Considerations

As all data used in this study are publicly available and no personal identifiers were obtained, our study was exempt from institutional ethics review. Where applicable, all posts included in this analysis have been paraphrased so that they cannot be traced back to the user. No identifying information (eg, usernames, IDs, or pictures) are included in the main manuscript or in the supplementary material.

Step 1: Poster-Patient Relationship ClassificationData Collection

Since Sina Weibo does not maintain a public application programming interface (API), we used a previously constructed web crawler to request publicly available Weibo posts. Our web crawler simulates a user visiting Weibo’s official website and searches for relevant posts (see the next paragraph for the search procedure). Through this approach, each web search request can obtain up to 50 posts before reinitiating a new search request to retrieve a new set of posts. In our crawler, we were able to set adjustable parameters to specify keywords, the publishing date, location, and interval times between 2 search requests.

We conducted 2 searches with different queries: “breast cancer (‘乳腺癌’)” and “sadness (‘悲伤’)”, as well as “breast cancer (‘乳腺癌’)” and “record (‘记录’)” in Chinese, from January 1, 2018, to December 31, 2021. For both queries, the interval time was set to 15 seconds and the location was unspecified, meaning that we searched for posts from across China. Finally, for the 2 searches with different queries, we obtained 160,182, and 144,125 posts, respectively. For each post, we additionally obtained the user id, username, user type, publish time, post text, location, number of comments, likes, and reposts, which were removed before commencement of analyses.

Next, for the data-cleaning phase, we combined the search results of the 2 queries into a single data set. Duplicate posts were removed through string matching, and obvious advertisements and irrelevant posts were removed by manually checking the data set. This was to ensure the posts were related to narrative accounts pertaining to breast cancer. Finally, this resulted in a cleaned data set containing relevant breast cancer–related narratives from individual users, for a total of 10,322 posts.

Poster-Patient Relationship Classification Criteria

First, we set up 6 categories based on the relationship of the mentioned patient and the author of the post: “post_user,” where the authors are themselves the patients (coded as 0); “family_members,” where the authors mention a family member (eg, parent) as the patient (coded as 1); “friends_relatives,” where a friend or nonimmediate relative (eg, cousins, aunt) is the patient (coded as 2); “acquaintances,” where a colleague or neighbor (social relationships) is the patient (coded as 3); “heard_relation,” where the author may be posting about a celebrity or a famous patient with cancer (coded as 4); and “no_patient,” where breast cancer is mentioned generally without being associated with a specific person (coded as 5).

Data Annotation

We randomly portioned 3000 (29.1%) of the 10,322 posts for manual annotation based on the classification criteria, with each data point (post) assigned a label from the 6 aforementioned categories. In the process of labeling, first we determined whether there was a patient in the post (binary classification task), and then we determined whether the poster-patient relationship could be inferred and labeled according to the prespecified classification criteria (multiclass classification task). All data labeling was performed by 1 of the authors who is a native Chinese speaker. See for the annotation proportions, and Table S1 in for examples of annotated posts.

To verify that our annotations were objectively labeled and free of subjective bias, we randomly selected 600 (20%) of the 3000 annotated posts, and these were reannotated in the same procedure by another native Chinese annotator who was not part of the research team. Across the 6 categories, the interannotator agreement was good (Cohen κ=0.67) [], and the original annotations were used to train the classification model.

Table 1. Distribution of annotated posts.
Posts, nNo_patient1089Heard_relation509Family_members443Acquaintances356Post_user338Friends_relatives265Data Preprocessing

In our study, we chose the pretrained Chinese-RoBERTa-wwm-ext (Chinese RoBERTa) [] model as our classification model. The Chinese RoBERTa is a large language transformer model based on the RoBERTa architecture [], trained on a large corpus of the in house–collected extended data containing an encyclopedia, news articles, and web forums, which has 5.4 billion words and is over 10 times bigger than the Chinese Wikipedia [], and is frequently used for Chinese NLP tasks. To improve the accuracy of the multiclass text classification, we decomposed the classification task over 2 stages (see Ref. []): a binary classification task to determine whether a patient was mentioned, followed by a multiclass classifier on posts where a patient was mentioned in order to identify the poster-patient relationship.

The pretrained language model (Chinese RoBERTa) has a limited input character length of 512, and 522 posts in our data set were longer than this character length limit. As such, we used automated text summarization to condense the text length to within 512 characters for these 522 posts using SnowNLP, a Python library that can perform Chinese word segmentation, part-of-speech tagging, sentiment analysis, text categorization, pinyin conversion, traditional simplification, text keyword extraction, text summarization, sentence segmenting, and text similarity estimation []. The SnowNLP tool segments posts by sentence and using the TextRank algorithm [] calculates the weight of each sentence in the post according to the extent to which the content of the sentence represents the content of the text. Finally, all the small units are sorted in reverse order according to their weight scores. When implementing this tool, by setting a number parameter, the corresponding number of sentences is output accordingly, resulting in summarized texts. In , we included some examples of automatic summarization.

Classifier Training

Following annotation and data preprocessing, 2 classifiers were constructed for this study in a 2-stage process. In the first stage, a binary classification model was trained to identify whether a patient is mentioned. This was followed by training a multiclass classification model to identify the poster-patient relationship for each post where a patient was mentioned in 1 of 5 classes: post_user, family_members, friends_relatives, acquaintances, and heard_relation. This resulted in a total of 6 classes corresponding to the annotations, with the inclusion of the “no_patient” class from the earlier binary classification model. In constructing the 2 classifiers, we specified the task of the RoBERTa model as classification. We monitored the training performance for each epoch through cross-entropy loss. Fine-tuning was implemented under the Pytorch framework, where we used the Amda Optimizer to optimize and update model parameters for training purposes. For testing, sklearn metrics were used to evaluate the binary classification and multiclass classification. In addition, 2400 (80%) of the 3000 annotated posts were used to train the model, and the main parameters for the model training were as follows: batch size=16, learning rate=1.0 × 10–5, and training epochs=5. We used 600 (20%) posts to test the fine-tuned model.

In the second stage, we removed the “no_patient” class from the annotated data. In total, 1515 (50.5%) posts were used to fine-tune the Chinese RoBERTa model. The main parameters were similar to the binary classifier, with batch size=16, learning rate=1.0 × 10–5, and training epochs=5. For validation, we used 396 (13.2%) posts to test the trained model.

Step 2: Examining Differences in Emotional ExpressionAnalysis 1: Sentiment Analysis Based on the LIWC

The LIWC program is a text analysis program that calculates the degree of use for different categories of words across a wide array of texts []. This tool was originally developed in English, but researchers have since produced a Chinese version of the LIWC dictionary based on the same criteria []. We used an open source Python package to access the Chinese LIWC dictionary. The LIWC dictionary has proved extremely useful in a number of different disciplines and has had a large impact on our understanding of how lexical elements related to cognition, affect, and personal concerns can be used to better understand human behavior [].

In our study, we focused on the emotion categories to implement the sentiment analysis in our corpus of Weibo posts. We used the LIWC program and its Chinese dictionary to examine 5 emotion categories available in the Chinese LIWC dictionary: positive emotions, negative emotions, sadness, anger, and anxiety. The LIWC dictionary operates by counting the number of terms in each post that corresponds to its internal dictionary for each emotion category, and outputs a score representing the ratio of relevant terms to all identified terms in the post. We then conducted Kruskal-Wallis tests to determine whether positive emotion terms, negative emotion terms, anxiety terms, sadness terms, and anger terms significantly differed between each poster-patient relationship class. If there was a significant effect of the emotion category, we conducted post hoc Dwass-Steel-Critchlow-Fligner (DSCF) pairwise comparisons to compare differences between specific categories.

In this paper, our data are in Chinese, so we had to tokenize our data. We used Jieba for tokenization, which is 1 of the most popular Chinese tokenization tools in NLP []. To clean out the noise, we excluded more than 2000 stop words, which were collected from an open source Chinese dictionary of stop words.

Analysis 2: Topic Modeling

Making sense of a large unstructured corpus through qualitative means is difficult. Therefore, we used topic modeling to better assist us in interpreting data. Topic modeling is a widely used approach to extract common, recurring themes from large amounts of text data through identification and clustering of repeated patterns in words and sentences. In this paper, we adopted the open source BERTopic algorithm [] to achieve this. BERTopic leverages transformers and class-based term frequency–inverse document frequency (c-TF-IDF) to create dense clusters of words, allowing for easily interpretable topics, while keeping important words in the topic descriptions []. Past research [] has also found that BERTopic-based topic modeling generally yields more theoretically interpretable results than other forms of topic modeling (eg, latent Dirichlet allocation or Top2Vec). As the BERTopic algorithm only assigns 1 topic to every document (post), we were able to compute topics per class, which allowed uniform comparison of topic distribution for every class (poster-patient relationships), enabling us to observe general trends: which topics are more frequently observed in which class of poster-patient relationship. As long texts are more suitable for modeling and there is no limit to the length of input sentences, during the topic modeling, we replaced the summarized sentences with the original ones. For identified topics, we deliberated on the schema associated with as many words in the topic as possible. Note that this process is largely subjective, so we encourage readers to additionally reference the words contained in each topic, rather than relying solely on the authors’ labels.

In this paper, our data are in Chinese and because the BERTopic model is based on the clustering of individual words to implement topic modeling; therefore, in the process of topic modeling, similar to the sentiment analysis, we needed to tokenize our Chinese data. We again used Jieba for tokenization []. To obtain meaningful entities from the topic models, we excluded more than 2000 stop words, which were collected from an open source Chinese dictionary of stop words.

ResultsStep 1: Poster-Patient Relationship ClassificationBinary Classifier

This model was trained to distinguish each post as either mentioning (“patient” class) or not mentioning (“no_patient” class) a patient. We merged the “post_user,” “family_members,” “friends_relatives,” “acquaintances,” and “heard_relation” classes into a superordinate “patient” class. The model achieved a high F1-score (see ).

Table 2. Binary classifier’s metrics report.ClassPrecisionRecallF1-scoreSupportno_patient0.900.900.90204patient0.950.950.95396Macro average0.920.920.92600Multiclass Classifier

Next, we constructed a multiclass classifier to focus on patient classification: “post_user,” “family_members,” “friends_relatives,” “acquaintances,” and “heard_relation.” Results are reported in .

Table 3. Multiclass classifier’s metric report.ClassPrecisionRecallF1-scoreSupportacquaintances0.760.670.7175heard_relation0.830.830.83102famliy_members0.930.900.9186post_user0.860.910.8982friends_relatives0.740.840.7951Macro average0.820.830.83396Post Classification

After excluding the annotated data, we were left with 7322 (70.9%) of the 10,322 data points (posts). These posts then underwent the 2-stage classification process. The first stage included a binary classifier to determine whether patient information was identifiable from the post (patient and no_patient), and if a patient was detected, the post then passed to the second stage. This included a multiclass classifier to classify the relationship between the patient and the Weibo poster. In the first stage, 4494 (61.4%) posts were classified as having a patient and 2828 (38.6%) posts as having no patient. Of the former, the relation classifications were as follows (): the patient was identified as a friend or relative (friends_relatives; n=667, 14.8%), as the poster (post_user; n=705, 15.7%), as an acquaintance (acquaintances; n=781, 17.4%), as a family member (family_members; n=961, 21.4%), and as someone they had only heard about (heard_relation; n=1380, 30.7%).

As and show, the rankings of categories by the number of relevant posts were similar regardless of whether the data were manually labeled or predicted by our classifier. The ranking list was no_patient > heard_relation > family_members > acquaintances > post_user > friends_relatives. We noted that the “no_patient” class that did not mention a specific patient was the majority class, which accounted for one-third of the total number of posts (n=2828, 38.6%). We think that posters use the target words (“breast cancer”) to share some personal thoughts, not necessarily about specific instances of breast cancer or for a targeted patient. Alternatively, they may feel no need to talk about the patient due to the content and style of the post. Except for this class, the distribution of the other poster-patient relationship classes was relatively balanced in the data set.

Table 4. Distribution of predicted posts.
Posts, nNo_patient2828Heard_relation1380Family_members961Acquaintances781Post_user705Friends_relatives667Step 2: Examining Differences in Emotional ExpressionsSentiment Analysis

For subsequent analyses, our aim was to maximize the information we could extract from the data, so manual annotations were combined with the machine-learned predictions for a total of 10,322 posts. We applied the LIWC and the matched Chinese dictionary to count the emotion-related words for each tokenized post. We mainly focused on positive emotion, negative emotion, sadness, anger, and anxiety categories. We calculated the ratio of each emotion category in each post (number of emotion words/number of all tokens). To visualize broad emotional differences among the classified poster-patient relationship classes, we plotted the mean scores for 6 identity categories in each of the 5 emotion categories.

For positive emotions, the “friends_relatives” class had a relatively higher value than the other 5 classes (). For negative emotions, the “no_patient” class had a relatively higher value than the other 5 classes. For angry terms, the “no_patient” class had a significantly higher value than the other 5 classes, which had almost the same values. For anxiety terms, the “family_members,” “no_patient,” and “post_user” classes had a higher value than the other 3 classes; the “heard_relation” class had the lowest value. For sadness terms, the “family_members,” “no_patient,” and “post_user” classes had a relatively higher value than the other 3 classes.

Table 5. Emotion distribution for each class in the 5 emotion categories (positive emotions, negative emotions, anger, anxiety, and sadness).
Mean ratio of each emotion category in each postaPositive emotions
no_patient0.05567
heard_relation0.05785
family_members0.05469
acquaintances0.06581
post_user0.05382
friends_relatives0.07490Negative emotions
no_patient0.11920
heard_relation0.09202
family_members0.09933
acquaintances0.09118
post_user0.09759
friends_relatives0.09386Anger
no_patient0.01020
heard_relation0.00490
family_members0.00467
acquaintances0.00479
post_user0.00469
friends_relatives0.00489Anxiety
no_patient0.00699
heard_relation0.00389
family_members0.00674
acquaintances0.00465
post_user0.00595
friends_relatives0.00430Sadness
no_patient0.01094
heard_relation0.00894
family_members0.01107
acquaintances0.00845
post_user0.01110
friends_relatives0.00928

aNumber of emotion words/number of all tokens.

Next, we statistically examined differences in emotions across poster-patient relationships. Kruskal-Wallis tests showed significant effects for positive emotions (posemo: χ25=185.9, P<.001), negative emotions (negemo; χ25=156.8, P<.001), anxiety (anx; χ25=50.6, P<.001), anger (anger; χ25=38.2, P<.001), and sadness (sad; χ25=56.8, P<.001). This suggests that for all emotion categories, significant effects were detected across the 6 poster-patient relationship classes. reports the post hoc DSCF pairwise comparisons.

Although there were a number of significant effects, here we comment primarily on consistent patterns of results that may be indicative of broader trends in Weibo users with respect to the emotional language used when posting about breast cancer. We noticed that the “friends_relatives” class had significantly higher positive emotions than all other poster-patient relationship classes, and this was followed closely by the “acquaintances” class, which had higher positive emotions than the other remaining poster-patient relationship classes.

Table 6. Pairwise comparisons for the 5 emotion categories.Class 1Class 2Positive emotionsNegative emotionsAnxietyAngerSadness

WaP valueWP valueWP valueWP valueWP valueacquaintancesfamily_members–7.87<.001b3.94.067.55<.001b1.05.986.42<.001bacquaintancesfriends_relative5.67<.001b1.87.771.29.941.75.823.02.27acquaintancesheard_relation–6.75<.001b0.640.990.91.981.81.792.47.50acquaintancesno_patient–10.73<.001b12.13<.001b2.65.426.38<.001b–0.46.99acquaintancespost_user–8.49<.001b3.13.235.15.004b1.74.825.16.004bfamily_membersfriends_relatives13.42<.001b–1.83.79–5.90<.001b0.86.99–3.01.27family_membersheard_relation1.78.81–3.96.06–7.94<.001b0.81.99–4.81.01bfamily_membersno_patient–2.62.438.63<.001b–6.92<.001b5.71<.001b–8.88<.001bfamily_memberspost_user–1.08.97–0.61.99–2.13.660.86.99–0.87.99friends_relativeheard_relation–12.65<.001b–1.54.89–0.57.99–0.21.99–1.04.98friends_relativeno_patient–16.21<.001b9.41<.001b0.94.984.01.05–4.23.03bfriends_relativepost_user–13.68<.001b1.19.963.69.09–0.07.992.02.71heard_relationno_patient–5.15.004b14.14<.001b2.01.725.55.001b–4.14.04bheard_relationpost_user–2.76.372.98.285.02.005b0.12.993.47.14no_patientpost_user1.29.94–8.31<.001b3.76.08–4.23.03b6.98<.001b

aStandardized Wilcoxon statistic from Dwass-Steel-Critchlow-Fligner (DSCF) pairwise comparisons.

bSignificant P values.

In addition, we found that “no_patient” posts had consistently higher negative emotions than the posts in all other poster-patient relationship classes, but no strong and consistent pattern of difference was observed between other poster-patient relationship classes. This pattern was mirrored strongly in the anger emotion category, suggesting that “no_patient” posts were higher on anger compared to posts in the other poster-patient relationship classes. As “negative emotions” is a broad emotion category containing many other negative emotion words in its dictionary, we think that strong differences observed in anger could be driving the significant difference found in the negative emotions category.

Furthermore, we noticed that with the exception of the “post_user” class, the “family_members” class was generally significantly higher in anxiety than the “acquaintances,” “friends_relatives,” “no_patient,” and “heard_relation” poster-patient relationship classes and higher in sadness than the “acquaintances,” “no_patient,” and “heard_relation” poster-patient relationship classes.

Clustered Topics

To gain an overview of why some poster-patient relationship classes were consistently higher in some emotions than other classes, we turned to topic modeling. Using the topics per class function of the BERTopic model, we aimed to compare topical relationships that mirrored some of the identified effects from the sentiment analysis.

We initially found that 139 topics were automatically generated from BERTopic, but this included several topics of low significance, where post counts numbered less than 50. As we wanted to focus on topics of greater relevance, we narrowed our analysis to include only the top 30 (21.6%) topics by topic prevalence across the entire data set, which was sufficient to cover more than 6000 (58.1%) posts. In and in Table S2 in , we list the top 30 topics with top 30 representative terms and provide a summarized theme for each topic. These are represented by an ID, which represents the ranked prevalence of each topic, while the topic number represents the topic labels assigned for the initial generation. We also visualized the distribution of (poster-patient relationship) classes per topic, which was used to identify topics that were more prevalent in a particular class for the analysis. These visualizations are available in our GitHub repository [].

Table 7. Top 30 terms of top 30 topics from topic modeling.IDTopic numberLabelTop 30 representative words (Chinese)Top 30 representative words (translated into English)00Anger生气,脾气,气死我了,情绪,真的angry, temper, I’m angry, emotions, really11Laments去世,家里,回来,生活,记得passed away, at home, come back, life, remember23Symptoms乳腺,乳房,肿块,增生,结节,breast, breast, lump, hyperplasia, node34Hospital stays医生,病人,主任,医院,手术doctor, patient, director, hospital, surgery47Hope and prayers希望,幸福,生活,人生,幸运hope, happiness, life, life, lucky56Hospitalization手术,医院,化疗,住院,医生surgery, hospital, chemotherapy, hospitalization, doctor68Lamenting hospitalization病房,医院,病人,恐惧,患者ward, hospital, patient, fear, patient72Dreams and nightmares梦里,梦见,梦到,昨晚,做梦dream, dreaming, dreaming, last night, dreaming810Diagnosis一年,手术,去年,确诊,希望a year, surgery, last year, diagnosed, hope95Chinese dramas刘静,女主,男主,欢喜,英子Liu Jing, heroine, hero, cheerful, Yingzi1013School老师,学生,家长,班主任,上课teacher, student, parent, classroom, lesson1120Friends朋友,闺蜜,离婚,聊天,命理friend, bestie, divorce, chat, numerology1218Sleep-wake cycles熬夜,睡觉,晚上,睡不着,睡着stay up, sleep, night, sleepless, sleep1312Passing去世,消息,难过,死者,刚刚passed away, news, sad, deceased, just1426Treatment processes放疗,化疗,结束,治疗,转移radiotherapy, chemotherapy, end, treatment, metastasis1533Treatment effects治愈,治疗,方案,效果,患者cure, treatment, protocol, effect, patient16113Appeal to emotion开心,心情,事情,几率,难过happy, mood, things, odds, sad1742Initiative面对,压力,生活,健康,人生face, pressure, life, health, life1811A Little Red Flower (a popular Chinese movie released in 2020)小花,一朵,千惠,小红花,病魔little flower, a, Chie, little red flower1945Suspicion of breast cancer怀疑,焦虑症,返祖,胸痛,检查suspicion, anxiety, revert, chest pain, examination2048Other cancers肺癌,肝癌,胃癌,肠癌,吸烟lung cancer, liver cancer, stomach cancer, bowel cancer2164Anxiety焦虑,担心,烦躁,考研,心情anxiety, worry, irritable, exam, mood2217Metastasis of cancer cells转移,癌症,癌细胞,患者,闫宏微transfer, cancer, cancer cells, patient, Yan Hongwei2322Weibo follows关注,微博,抗癌,荔枝,记录concern, microblogging, anti-cancer, lychee2423Weibo usage微博,媽媽,努力做到,更新,不想microblogging, mom, trying to do, update, don\'t want2585Treatment side effects头发,假发,化疗,光头,掉头发hair, wig, chemotherapy, bald, lose hair2627Check-up姐夫,电话,昨天,医生,回去brother-in-law, phone, yesterday, doctor, go back2763Female physiology没事,预防,增生,例假,一去Nothing, prevention, hyperplasia, period, a go289Public figures陈晓旭,李明,伤官,林黛玉,李婷Chen Xiaoxu, Li Ming, hurt official, Lin Daiyu, Li Ting2958Treatment stages化疗,第二次,第三次,结束,白细胞chemotherapy, second, third, end, white blood cellsNotable TopicsNegative Emotions and Anger

The sentiment analysis suggested that the “no_patient” class had consistently higher negative emotions and anger than all other poster-patient relationship classes. Next, we examined the top 30 topics to identify topics with a similar pattern, which were topics 0, 2, 3, 18, 13, 23, 42, 45, 48, 64, and 113. These spanned a number of overlapping themes. Topic 0, for example, contained terms that directly expressed anger and also appeared to carry the speculation that anger is a cause of breast cancer. Similarly, topics 42, 64, and 113 comprised emotive posts about being positive or hopeful in the face of breast cancer, as well as the anxiety and stress it causes. Posts on topics 3, 48, and 63 contained physiological and medical terms, particularly cancer-related terms, their comorbidities, and their antecedents, and posts on topic 45 appeared to express anxiety at the poster facing a possible cancer diagnosis. Finally, topics 2 and 18 contained posts about the user having a nightmare about breast cancer while sleeping, and topics 13 and 20 were about cancer in everyday life. A guiding theme for these topics is that they seem to relate to the posters’ fears and anger toward cancer in general.

Sadness and Anxiety

Topics 26 and 58 resembled the patterns of relationship classes for sadness and anxiety, in that with the exception of the “post_user” class, the “family_members” class was more prevalent than the other poster-patient relationship classes. These topics shared a common theme, in that they discussed treatment options for breast cancer (eg, chemotherapy, immunotherapy). One explanation could be that immediate family members, as caregivers, were more concerned about breast cancer treatment.

Error Analysis for Machine Learning Classification

Although our classifiers predicted posts well to some extent, we noticed that some cases were mistakenly classified into other categories, according to the metrics from and . To explore the possible reasons behind this misclassification, we implemented error analysis.

We found that 1 common reason for these errors was when the patient in a post was unclear and what they said needed to be inferred through semantic understanding. In Table S3 in , for example, in post I, the breast cancer patient in the post was the post author (we inferred that the patient should be the poster from reading the post), so according to our classification definition, the true label would be “post_user,” but the predicted label from our classifiers was “acquaintances.” We think that this could be attributed to a mention of a colleague at the beginning of the post and was mistakenly classified into the “acquaintances” class instead. We observed another reason for errors was when the patient was clearly mentioned but there were multiple other actors mentioned in the post as well. Such appearances can greatly affect the classifiers’ prediction. In post II, based on our understanding, the patient appeared to be the poster, but there were many other family members present (eg, father, baby, son, daughter-in-law, granddaughter, grandma). Therefore, post II was mistakenly classified into the “family_members” class instead of the “post_user” class.

DiscussionPrincipal FindingsStep 1: Poster-Patient Relationship Classification

We fine-tuned the pretrained language model Chinese RoBERTa on our annotations on poster-patient relationships to construct a classification model capable of identifying patients’ relationships with the posters of Weibo posts concerning breast cancer. We subsequently used those classifiers to implement a 2-stage classification process. Both classifiers performed well, and we were generally able to classify poster-patient relationships with moderate-to-high accuracy. This comprised step 1, the poster-patient relationship classificati