Maintenance Notice

Due to necessary scheduled maintenance, the JMIR Publications website will be unavailable from Monday, March 11, 2019 at 4:00 PM to 4:30 PM EST. We apologize in advance for any inconvenience this may cause you.

Who will be affected?

Advertisement

Citing this Article

Right click to copy or hit: ctrl+c (cmd+c on mac)

Published on 07.04.17 in Vol 19, No 4 (2017): April

This paper is in the following e-collection/theme issue:

    Original Paper

    Understanding Health Care Social Media Use From Different Stakeholder Perspectives: A Content Analysis of an Online Health Community

    1School of Economics and Management, Beijing University of Chemical Technology, Beijing, China

    2School of Management, Shanghai University, Shanghai, China

    3School of Business, East China University of Science and Technology, Shanghai, China

    4Antai College of Economics and Management, Shanghai Jiao Tong University, Shanghai, China

    Corresponding Author:

    Yingjie Lu, PhD

    School of Economics and Management

    Beijing University of Chemical Technology

    North Third Ring Road 15, Chaoyang District

    Beijing, 100029

    China

    Phone: 86 1064448681

    Fax:86 1064448681

    Email: lvyj@mail.buct.edu.cn


    ABSTRACT

    Background: Health care social media used for health information exchange and emotional communication involves different types of users, including patients, caregivers, and health professionals. However, it is difficult to identify different stakeholders because user identification data are lacking due to privacy protection and proprietary interests. Therefore, identifying the concerns of different stakeholders and how they use health care social media when confronted with huge amounts of health-related messages posted by users is a critical problem.

    Objective: We aimed to develop a new content analysis method using text mining techniques applied in health care social media to (1) identify different health care stakeholders, (2) determine hot topics of concern, and (3) measure sentiment expression by different stakeholders.

    Methods: We collected 138,161 messages posted by 39,606 members in lung cancer, diabetes, and breast cancer forums in the online community MedHelp.org over 10 years (January 2007 to October 2016) as experimental data. We used text mining techniques to process text data to identify different stakeholders and determine health-related hot topics, and then analyzed sentiment expression.

    Results: We identified 3 significantly different stakeholder groups using expectation maximization clustering (3 performance metrics: Rand=0.802, Jaccard=0.393, Fowlkes-Mallows=0.537; P<.001), in which patients (24,429/39,606, 61.68%) and caregivers (12,232/39,606, 30.88%) represented the majority of the population, in contrast to specialists (2945/39,606, 7.43%). We identified 5 significantly different health-related topics: symptom, examination, drug, procedure, and complication (Rand=0.783, Jaccard=0.369, Fowlkes-Mallows=0.495; P<.001). Patients were concerned most about symptom topics related to lung cancer (536/1657, 32.34%), drug topics related to diabetes (1883/5904, 31.89%), and examination topics related to breast cancer (8728/23,934, 36.47%). By comparison, caregivers were more concerned about drug topics related to lung cancer (300/2721, 11.03% vs 109/1657, 6.58%), procedure topics related to breast cancer (3952/13,954, 28.32% vs 5822/23,934, 24.33%), and complication topics (4449/25,701, 17.31% vs 4070/31,495, 12.92%). In addition, patients (9040/36,081, 25.05%) were more likely than caregivers (2659/18,470, 14.39%) and specialists (17,943/83,610, 21.46%) to express their emotions. However, patients’ sentiment intensity score (2.46) was lower than those of caregivers (4.66) and specialists (5.14). In particular, for caregivers, negative sentiment scores were higher than positive scores (2.56 vs 2.18), with the opposite among specialists (2.62 vs 2.46). Overall, the proportion of negative messages was greater than that of positive messages related to symptom, complication, and examination. The pattern was opposite for drug and procedure topics. A trend analysis showed that patients and caregivers gradually changed their emotional state in a positive direction.

    Conclusions: The hot topics of interest and sentiment expression differed significantly among different stakeholders in different disease forums. These findings could help improve social media services to facilitate diverse stakeholder engagement for health information sharing and social interaction more effectively.

    J Med Internet Res 2017;19(4):e109

    doi:10.2196/jmir.7087

    KEYWORDS



    Introduction

    People are increasingly turning to the Internet for health information. Some social media services, such as Wikipedia, Facebook, online forums, and message boards, are increasingly being used to obtain health information and share health care experiences [1,2]. A Pew survey showed that 61% of American adults looked online for health information and that 41% of this group had read about someone else’s medical experiences posted in social media. More than half thought the information online affected a decision about how to treat an illness or condition [3].

    There are some reasons why patients and their caregivers turn to social media for health care information. First, they feel that doctors are too busy to answer their questions [4]; many doctors only tell their patients how to take a treatment but do not take time to fully explain the detailed reasons [5]. Second, health care social media thus allows users to learn more about their health problems through the use of social support and make decisions about treatment options [6], especially for those with chronic diseases [7]. Third, convenience and anonymity [4] are also important reasons why patients and their caregivers turn to social media, where they feel less embarrassed about asking online experts or communicating with other online members about their health conditions [8].

    In addition to patients and their caregivers, other health care stakeholders are involved in health care social media, including doctors, physicians, nurses, aides, social workers, spiritual caregivers, counselors, therapists, and volunteers [9-17]. Some medical experts like to communicate with social media users and offer them professional medical knowledge. In addition, some fellow patients are willing to share their experiences with treatment, which are considered valuable for patients with newly diagnosed conditions. These individuals devote their time to helping other patients get answers in their time of need. These information providers are recognized as health specialists and make great contributions to health care social media.

    Health care social media is not only used to exchange health information [18-20], but also provides a platform for emotional communication [21-23]. Some emotional expressions, such as venting, sending positive energy, and showing compassion or empathy, can help patients with serious diseases to cope. It is important especially for patients with cancer and chronic illnesses to receive emotional support and encouragement from fellow patients [24]. A study indicated that patients with low-survival-rate diseases have more emotional communication than those with high-survival-rate diseases [25]. Another study found that an estimated 75% to 85% of patients engaged in health care social media changed their sentiment in a positive direction through online interactions with other users.

    Thus, different types of members discuss topics they are concerned about and share their feelings in health care social media. Understanding the concerns of different stakeholders and how they use health care social media is very meaningful work. Such information can enable users, especially newcomers, to obtain a sense of what social media is, help them quickly find the issues they are concerned about, increase their involvement more easily, and obtain the resources they need for health self-education, self-caring, and self-management. For health care social media sites, a better understanding of user online behaviors could help webmasters provide humanized supporting functions to facilitate diverse stakeholder engagement. For researchers in the field of health care, such studies could clearly summarize the behavior characteristics of different stakeholders and provide advice and guidance for health care social media.

    However, given the tremendous amounts of health-related messages posted by social media users, it is difficult and time consuming to employ traditional statistical approaches to find valuable information. Moreover, social media users are usually reluctant to offer personal information due to privacy concerns, and user identification data are usually proprietary and hence not easily accessible. To address these problems, text mining techniques have recently been applied to health social media in some studies [26] to shed light on the demographics of health-related social media users by extracting keywords and other key attributes. However, clearly identifying different health care stakeholders is still a problem that needs to be solved. We therefore proposed a novel framework using text mining techniques to comprehensively analyze the content of an online health community from the perspectives of diverse stakeholders. We used a computational social science approach to process the large amount of text data for stakeholder analysis, topic analysis, and sentiment analysis. This approach allowed us to determine the characteristics of different stakeholders involved in health care social media to facilitate an understanding of health care social media use.


    Methods

    Data Collection

    We chose the online health community MedHelp.org (MedHelp International, San Francisco, CA, USA) as our data source. This site is one of the most popular health care social media platforms. It consists of more than 170 discussion boards about different diseases. Since the site opened in 1994, millions of threads have been posted in the community. It also attracts over one million visitors every month.

    We selected representative disease forums. Lung cancer and breast cancer are common cancers with high mortality, and some studies have shown that both cancers are among the most common cancers that Internet users seek information about [27]. Diabetes, as a chronic disease, is also among the most frequently discussed diseases in health care social media. Therefore, we chose lung cancer, breast cancer, and diabetes as our research subjects. We obtained all the webpages from these 3 disease forums by using webcrawler software Offline Explorer 6.8 (MetaProducts Systems), and then we parsed the pages to extract available messages and stored the messages in a database. Next, we filtered out noisy and unreliable data by text preprocessing. We finally collected 138,161 messages posted by 39,606 members in these 3 disease forums in the 10 years from January 2007 to October 2016 as our experimental data. Table 1 shows the statistical results.

    Table 1. Data collection statistics (January 2007 to October 2016) from 3 disease forums on MedHelp.org.
    View this table

    We only used messages that were open to the public. We never used any user identification data. Personal information such as name, age, and other demographics was not used or reported as part of the results of the study. Therefore, this study did not raise any ethical or legal concerns.

    Experiment Design

    In this paper, we propose a novel framework using text mining techniques to comprehensively analyze the content of the messages posted in 3 disease forums. The framework consists of 3 parts: stakeholder identification, topic identification, and sentiment analysis, as Figure 1 shows.

    First, to distinguish different stakeholders engaged in the online health community, we based stakeholder analysis on text mining techniques to identify different user groups. Users were automatically partitioned into different groups based on the similarity of their posts. An unsupervised approach such as a clustering technique was applied to stakeholder identification. Then, to better understand the concerns of different stakeholders, we based topic analysis on text mining techniques to identify health-related hot topics. Next, we analyzed sentiment to assess the valence and intensity of the messages posted by different stakeholders, as the emotional support and encouragement offered by community members are also an important component of online communication.

    Figure 1. Research framework.
    View this figure
    Stakeholder Identification

    Stakeholder identification included the following 4 steps: (1) message collection, (2) feature extraction, (3) probabilistic clustering, and (4) keyword extraction and stakeholder identification.

    For message collection, we first collected all messages posted in the 3 disease forums, and then merged all the messages posted by the same user into one.

    For feature extraction, we used a comprehensive textual feature representation encompassing writing style-based features and content-specific features for authorship analysis [28,29]. Writing style-based features were the number of messages, average number of sentences per message, and frequency of words per message. Content-specific features were word n-grams, medical domain-specific terminologies, and kinship terminologies. The n-grams referred to word unigrams, bigrams, and trigrams and represented important keywords and phrases in the messages. Medical domain-specific terminologies were helpful in distinguishing diverse stakeholders, as some studies have shown that patients and their caregivers used more lay concepts and fewer professional terminologies than health professionals do [30-33]. We selected Medical Subject Headings (MeSH), a controlled thesaurus, to extract health-related terminologies. Family members of patients serving as primary caregivers usually used kinship terms to refer to the patients they cared for in their posts, and therefore we selected kinship terminologies as part of the content-specific features to distinguish caregivers from other stakeholders. Kinship terminologies in this study were mainly identified by one Unified Medical Language System (UMLS) semantic type, family group, which has been used in previous studies to identify family member concepts [34].

    For probabilistic clustering, without a priori knowledge of the number of user groups and their specific characteristics, we chose a probabilistic clustering that leveraged the expectation maximization algorithm among various clustering techniques for stakeholder identification [35]. We found the optimal number of clusters by cross-validating different numbers of clusters. We then used 3 metrics to evaluate the clustering results: the Rand index, Jaccard similarity coefficient, and the Fowlkes-Mallows (FM) index, which we used in our previous study [36]. The clustering result was a probabilistic distribution of instances of belonging to each stakeholder cluster. We labeled each user with a group number according to the cluster with the highest probability assignment.

    For keyword extraction, the keywords were terms that best distinguished the corresponding user group from other groups. We extracted these keywords from the messages posted by each user group according to their scores. Supposing all resultant clusters were C1,C2...CN, then for each n-gram (unigram, bigram, and trigram) term w in a cluster Ci, its score was calculated as f(w, Ci) × log (N/|{Cj|f(w, Cj) ≥ f(w, Ci), j=1,2...N}|) (equation 1), where f(w, Ci) is the frequency of w in cluster Cj, and |{Cj|f(w, Cj) ≥ f(w, Ci), j=1,2...N}| is the total number of clusters with a frequency of term w greater than or equal to the term w frequency of the cluster evaluated. We ranked keywords with high scores, from which we could infer the stakeholder identification.

    Topic Identification

    We identified topics in a manner similar to identifying stakeholders, by the following steps: message collection, feature extraction, probabilistic clustering, keyword extraction, and topic identification.

    For message collection, we only chose informational messages involving health conditions, online reviews of particular drugs or medical treatments, and so on. We filtered out messages containing no informational support but only offering emotional support and other spam messages.

    For feature extraction, the features used in topic identification included 2 parts: word n-grams and domain-specific terminologies. Domain-specific terminologies were from the UMLS Metathesaurus. The UMLS Metathesaurus as the world’s largest repository of biomedical concepts, consisting of 1.7 million biomedical concepts, where each concept is assigned to at least one of the 134 semantic types. We only chose the terminologies mapped to health-related semantic types that we used in our previous study [36]. To obtain these medical terminologies automatically, we used MetaMap 2014 (US National Library of Medicine), a highly configurable program that maps biomedical text to concepts in the UMLS Metathesaurus. Using a Java application programming interface offered by MetaMap, we could parse the messages to get the health-related terminologies.

    We performed clustering and keyword extraction for topic identification in the same manner as stakeholder identification.

    Sentiment Analysis

    We analyzed sentiment to assess the valence and intensity of the messages by community members. Lexicon-based approaches are widely used in sentiment analysis, and some well-known sentiment lexicons, such as SentiWordNet [37], have been successfully applied in sentiment analysis [38,39]. The SentiWordNet lexicon provides positive- and negative-intensity scores for each sentiment term. In this study, we used the SentiWordNet lexicon to extract sentiment terms from the messages and calculate their sentiment scores. We used the following 3 sentiment measures to evaluate sentiment expression: PositiveScores, NegativeScores, and SubjectiveScores [40,41], where PositiveScores is positive-polarity scores divided by the number of messages; NegativeScores is negative-polarity scores divided by the number of messages; and SubjectiveScores is subjective-polarity scores divided by the number of messages.


    Results

    Stakeholder Analysis

    We used the expectation maximization clustering algorithm for stakeholder identification. Expectation maximization clustering can evaluate and determine the optimal number of clusters by cross-validating different numbers of clusters. We identified 3 stakeholder groups in 3 disease forums. We then performed the 2-sample t test to evaluate whether there was a significant difference in the 3 stakeholder groups. The null hypothesis was that the difference between 2 groups was 0. The conclusion was to reject the null hypothesis (P<.001) and that the 3 stakeholder groups were significantly different. Then, we extracted some keywords from each group according to equation (1) after filtering out meaningless or invalid phrases. Table 2 shows the results.

    Table 2. Stakeholder analysis in the 3 disease forums.
    View this table

    Then, we could infer stakeholder identification according to the extracted keywords. Taking clustering results in the lung cancer forums as an example, in the cluster 1 group, members talked about their own conditions (eg, my chest, my lungs, my left lung, my right lung, of my chest, in my chest, on my chest) and procedures they underwent (eg, I was diagnosed, my xray result, my ct scan). They hoped to get help (eg, please help me, my question, i was wondering) and then gave thanks (eg, be greatly appreciated). We therefore assigned cluster 1 as the patient group. In the cluster 2 group, members talked more about their family members (eg, my husband, my mother, my father, my dad, my mom, my sister, father in law) and expressed their concern about their family members (eg, i am worried). They also raised some questions of concern (eg, my question is) and gave thanks (eg, thank you, thanks in advance, thanks so much). We assigned cluster 2 as the caregiver group. In the cluster 3 group, members were more likely to give other members advice and suggestions (eg, you should, you need to, you could consider) and offer help (eg, hope this helps) or advise them to undergo further consultation and procedures with their doctors (eg, with your doctor, see your doctor, with your physician). Members in the cluster 3 group often expressed their compassion and encouragement to other members (eg, stay positive, good luck, all the best, sorry to hear, best of luck). We thus assigned cluster 3 as the health specialist group.

    Then we used the 3 evaluation metrics to test whether the textual features proposed in this study could significantly distinguish different stakeholders. By incorporating writing style-based features (F1), word n-grams (F2), medical domain-specific terminologies (F3), and kinship terminologies (F4) into the feature set in turn, we evaluated the results of clustering based on different feature sets, as Table 3 shows. The results showed that the 3 measure values increased as more features were incorporated into the feature set, and reached a maximum (eg, Rand=0.802, Jaccard=0.393, FM=0.537 for breast cancer) when all features were incorporated, indicating that these textual features could improve the performance of stakeholder identification significantly.

    Table 3. Performance measures for distinguishing stakeholders using different textual feature setsa.
    View this table

    In terms of the distributions of different stakeholders in the 3 disease forums, as Figure 2 shows, patients (24,429/39,606, 61.68%) and caregivers (12,232/39,606, 30.88%) were the majority of the population, in contrast to health specialists (2945/39,606, 7.43%). The proportions of patients (1202/2416, 49.75% vs 5738/11,571, 49.59%) and caregivers (1053/2416, 43.58% vs 4836/11,571, 41.79%) were similar in the lung cancer and diabetes forums. By contrast, in the breast cancer forum, the proportion of patients (17,489/25,619, 68.27%) was significantly greater.

    The reason for the different distributions may be that women are the prominent group with breast cancer and are also more likely than men to seek online information, as indicated by a previous study [42].

    Table 4 shows descriptive statistics of postings by different stakeholders. Both patients and caregivers published fewer than 2 messages (1.48 and 1.51) on average. By contrast, health specialists, although constituting a minority of the population in the online community compared with patients and caregivers, published many more postings (28.4) on average than the other 2 groups.

    We speculated about the reason for different distributions of posting. Generally, most patients and caregivers aimed to seek health information they were concerned about. They would leave the site after receiving satisfactory answers. Another possibility was that they simply became frustrated with the inability to find answers and left the site soon. Most patients and caregivers thus were short-term participants.

    Figure 2. Distributions of the 3 stakeholder groups.
    View this figure
    Table 4. Clustering results in the stakeholder analysis.
    View this table
    Figure 3. Changes in the proportions of stakeholder groups in (A) lung cancer, (B) diabetes, and (C) breast cancer forums, January 2007 to October 2016.
    View this figure

    In terms of changes in the proportions of the 3 stakeholder groups (see Figure 3), the proportion of the specialist group tended to decrease and then stabilize. By contrast, the proportion of caregivers decreased gradually, and the proportion of patients increased.

    One possible explanation is that most health specialists gradually formed a stable relationship with the online community and only a few left after short-term engagement. By contrast, as health care social media becomes more widespread and widely accepted, a higher proportion of patients than of caregivers prefer to seek online health information. After all, patients themselves are most concerned about their own health conditions.

    Topic Analysis

    Clustering and keyword extraction for topic identification were performed in the same manner as stakeholder identification. We identified 5 significantly different health hot topics (P<.001), as Table 5 shows. They were named based on the extracted keywords with similar UMLS semantic types: symptom, examination, procedure, drug, and complication.

    Table 5. Topic analysis in the 3 disease forums.
    View this table

    We then used the 3 evaluation metrics to test whether the textual features proposed in this study could be used to distinguish different hot topics significantly. By incorporating word n-grams (F1) and medical domain-specific terminologies (F2) into the feature set, we evaluated the results of clustering based on a 2-feature set, as s Table 6 shows. The results showed that feature set F1+F2 outperformed F1 significantly (eg, Rand=0.783, Jaccard=0.369, FM=0.495 for breast cancer), indicating that the 2 types of textual features improved the performance of topic identification significantly.

    Table 6. Performance measures for distinguishing hot topics using different textual feature setsa.
    View this table

    To further explore the concerns of different stakeholders, we examined the distributions of the 3 stakeholder groups in the 5 hot topics (Figure 4). Among the different stakeholders in the different types of diseases, their concerns about various health topics were fairly different.

    A significantly greater proportion of patients with lung cancer (536/1657, 32.34%) were involved in the symptom topics. Patients with diabetes were more interested in drug topics (1883/5904, 31.89%). Patients with breast cancer were more likely to mention their examination or tests online (8728/23,934, 36.47%).

    The reasons may be that some early symptoms of lung cancer, such as coughing or wheezing, are very similar to ailments such as fevers and bronchitis. Patients with such ailments were not sure whether their conditions were signs of lung cancer and thus hoped to obtain valuable information by seeking the experiences of others or consulting about their conditions on health care social media. By contrast, patients who found an abnormality of their breast tissues usually underwent further examination. Because of a lack of other obvious symptoms, they mainly discussed their examination or tests for the diagnosis of breast cancer. Diabetes is a common chronic disease that is difficult to cure. Patients with diabetes use regular long-term medication and thus were naturally concerned about topics related to antidiabetic drugs to learn about their medical effects and possible side effects.

    The distribution of hot topics among caregivers was similar to that of patients. Both groups appeared to be mainly concerned with the hottest health issues related to a specific disease, as patients and caregivers were both information requesters. However, the proportional distribution of topics was distinguishable between patients and caregivers. For example, a significantly greater proportion of caregivers than of patients were involved in drug topics (300/2721, 11.03% for caregivers vs 109/1657, 6.58% for patients) in lung cancer forums and in procedure topics (3952/13,954, 28.32% for caregivers vs 5822/23,934, 24.33% for patients) in breast cancer forums.

    These results indicate that caregivers were more concerned about disease treatment topics than disease diagnosis because drug and procedure topics are both related to disease treatment. Complication was another health topic of interest to caregivers, possibly because it may greatly affect their caregiving responsibilities. As Figure 4 shows, the proportion of complication topics was significantly greater among caregivers than among patients in the 3 disease forums (4449/25,701, 17.31% for caregivers vs 4070/31,495, 12.92% for patients).

    As Figure 4 shows, health specialists were equally distributed among the 5 health topics, in contrast to patients and caregivers. This result indicates that specialists, as information providers, were more likely to focus on various health topics and to share their experiences and knowledge. Broad knowledge is indispensable for health specialists to offer help to patients in their time of need.

    Figure 4. Distributions of (A) patients, (B) caregivers, and (C) specialists in the 5 hot topics.
    View this figure

    Sentiment Analysis

    Emotional messages could be distinguished from informative messages by setting a threshold of sentiment measures in SubjectiveScores. As Figure 5 shows, informative messages (108,519/138,161, 78.54%) were more prevalent than emotional messages (29,642/138,161, 21.45%).

    This result suggests that most users make use of health care social media more to exchange health information than to provide a platform for emotional communication. This result is also in agreement with the above argument that most patients and caregivers were short-term participants and that many only sought health information of interest.

    The distribution of informative and emotional messages among different stakeholders indicated that patients were more likely to express their emotions than caregivers and specialists. The proportion of emotional messages was 25.05% (9040/36,081) for patients, 14.39% (2659/18,470) for caregivers, and 21.46% (17,943/83,610) for specialists. Particularly among patients in the diabetes forums, patients usually encouraged each other to overcome their diseases by pursuing long-term treatment. By contrast, caregivers preferred to share informative messages and were the least likely to exchange emotions among the 3 stakeholders.

    Figure 5. Distributions of informative messages and emotional messages by (A) disease and by stakeholder group in (B) lung cancer, (C) diabetes, and (D) breast cancer forums.
    View this figure
    Figure 6. Sentiment measures of the 3 stakeholder groups.
    View this figure

    One possibility is that the caregiving responsibilities of caregivers do not leave sufficient time for online communication. Especially for the caregivers involved in the 2 cancer forums, most only sought diagnosis and treatment information of interest online to help their patients cure their high-survival-rate diseases.

    We calculated the 3 sentiment measures of PositiveScores, NegativeScores, and SubjectiveScores to evaluate sentiment expression, as Figure 6 shows. The subjective sentiment scores of patients (2.46) were lower than those of caregivers (4.66) and specialists (5.14), indicating that, although patients were more likely to express their emotions, as Figure 5 shows, the emotional intensity of patients was lower than that of specialists and caregivers. The negative sentiment scores of caregivers (2.56) were higher than their positive scores (2.18). Among health specialists, positive sentiment scores (2.62) were higher than their negative sentiment scores (2.46).

    We speculated about the reasons for the differences in sentiment expressions. Patients appear to prefer to regard health care social media as an emotion-exchange platform rather than a place to vent their negative feelings. Conversely, caregivers undertaking long-term caretaking activities usually expressed strong emotions about their patients. They preferred to use negative words with high emotional intensity to vent their emotional stress and burden [43,44], as Figure 6 shows that the negative sentiment scores of caregivers were higher than their positive scores. Many caregivers appeared to regard health care social media as a good platform to vent and relieve stress. For health specialists, higher positive sentiment scores implied that this group provided emotional support by encouraging patients and their caregivers to confront their diseases rather than expressing sympathy and pity.

    Further analysis identified differences in sentiment expression by different stakeholders in different hot topics, as Figure 7 shows. Overall, for patients and caregivers, the proportions of negative emotional messages were greater than those of positive emotional messages among the topics related to symptom, complication, and examination; for example, for patients, the proportion was 30.83% (562/1823) positive versus 69.17% (1261/1823) negative in symptom topics, 36.43% (1055/2896) versus 63.57% (1841/2896) in examination topics, and 37.85% (450/1189) versus 62.15% (739/1189) in complication topics.

    This result indicates that patients and caregivers who discussed complications and examination topics feared illness, depression, hopelessness, anxiety, and other negative emotions and thus were more likely to express negative emotions. In particular, among caregivers, negative emotional intensity was significantly higher for the complication topic than for the other topics, indicating that complications were a considerable burden for caregivers performing long-term caretaking activities, and this group expressed strong negative feelings of disappointment and even despair. By contrast, drug and procedure topics are usually more closely related to the treatment of disease, and thus all stakeholders who discussed these 2 topics were more likely to have a positive attitude; for example, for patients, the proportion was 47.94% (580/1210) positive versus 52.06% (630/1210) negative in drug topics and 59.11% (1136/1922) versus 40.89% (786/1922) in procedure topics. In their discussion of the 2 topics, patients were happy to be getting better, and the caregivers expressed their satisfaction with the treatment; in addition, specialists conveyed their congratulations and best wishes to the patients.

    Finally, we analyzed trends in emotional changes among the different stakeholders to illustrate whether members who engaged in health care social media changed their sentiment expression based on online interactions with other users. We incorporated the messages posted by different members every half month and calculated the changes in the 3 sentiment measures of PositiveScores, NegativeScores, and SubjectiveScores over time in the first year of their involvement. Figure 8 shows the results.

    As Figure 8 shows, the subjective sentiment scores of the messages from patients tended to increase in the early weeks and then stabilized. Patients preferred to express concern about their health topics of interest during their early involvement and then gradually began to express their feelings for social support. From the perspective of sentiment polarity, we found that positive sentiment scores increased gradually while negative sentiment scores decreased, implying that the patients engaged in health care social media changed their emotional state in a positive direction through online interactions with other users. Some possible reasons for this change are that the patients eliminated fears and anxiety after their problems were resolved or that professionals and fellow patients encouraged them to confront their diseases.

    The subjective sentiment scores of the messages from caregivers remained high. Caregivers apparently preferred to use subjective words with high emotional intensity in their posts compared with patients. From the perspective of sentiment polarity, positive sentiment scores increased gradually while negative sentiment scores decreased, similar to the trend changes in patients, possibly for the same reasons discussed above. However, in contrast to the sentiment expression of patients, for caregivers the negative sentiment scores remained higher than the positive sentiment scores, implying that caregivers appeared to regard health care social media as a platform to share their negative feelings. Caregivers may complain about their long-term, heavy caretaking work online to relieve their stress at any time.

    All 3 sentiment scores of the messages from specialists remained stable over time. This result indicates that specialists were involved long term in the online community to provide emotional support and to help patients and caregivers in various ways. They expressed their sympathy and pity for the unfortunate users by using some negative words while encouraging them to fight their illness using some positive words.

    Figure 7. Distributions of positive and negative messages posted by the 3 stakeholder groups in the 5 hot topics.
    View this figure
    Figure 8. Changes in the sentiment expression of (A) patients, (B) caregivers, and (C) specialists over the first year of each user’s involvement.
    View this figure

    Discussion

    Understanding the concerns of different stakeholders and how they use social media is very meaningful work. In this paper, we proposed a novel framework using text mining techniques to perform a comprehensive content analysis of an online health community from the perspectives of diverse stakeholders. We used a computational social science approach to process the large amount of text data for stakeholder analysis, topic analysis, and sentiment analysis. We identified significant differences in hot topics of interest and sentiment expression among different stakeholders involved in different types of disease forums. These valuable conclusions provide a better understanding of health care social media use by different stakeholders that may aid improvements in social media services to facilitate diverse stakeholder engagement for more effective health information sharing and social interaction.

    This study also has some limitations that must be considered further. Further research should examine how to describe and measure the impacts of health care social media use on health self-management. Users involved in social media maintained good communication and developed online social networks, and thus future studies should include social network analysis of different stakeholder groups to determine the impacts of these relationships on their engagement. In addition, some deeper stakeholder analysis by using text mining techniques, such as how to distinguish knowledgeable patients and professional doctors in specialist groups, would be considered in our further study.

    Acknowledgments

    This work was supported by the National Natural Science Foundation of China (grant no. 71401005 and no. 71371005) and the Major Research Plan of the National Natural Science Foundation of China (grant no. 91646205).

    Conflicts of Interest

    None declared.

    References

    1. Lasker JN, Sogolow ED, Sharim RR. The role of an online community for people with a rare disease: content analysis of messages posted on a primary biliary cirrhosis mailinglist. J Med Internet Res 2005 Mar 31;7(1):e10 [FREE Full text] [CrossRef] [Medline]
    2. Adams S, de Bont A, Berg M. Looking for answers, constructing reliability: an exploration into how Dutch patients check web-based medical information. Int J Med Inform 2006 Jan;75(1):66-72. [CrossRef] [Medline]
    3. Fox S, Jones S. The social life of health information. Washington, DC: Pew Internet & American Life Project; 2009.   URL: http://www.pewinternet.org/files/old-media//Files/Reports/2009/PIP_Health_2009.pdf [accessed 2017-03-29] [WebCite Cache]
    4. Umefjord G, Petersson G, Hamberg K. Reasons for consulting a doctor on the Internet: Web survey of users of an Ask the Doctor service. J Med Internet Res 2003 Oct 22;5(4):e26 [FREE Full text] [CrossRef] [Medline]
    5. Dickerson SS, Boehmke M, Ogle C, Brown JK. Seeking and managing hope: patients' experiences using the Internet for cancer care. Oncol Nurs Forum 2006 Jan 01;33(1):E8-17. [CrossRef] [Medline]
    6. Gerber BS, Eiser AR. The patient physician relationship in the Internet age: future prospects and the research agenda. J Med Internet Res 2001;3(2):E15 [FREE Full text] [CrossRef] [Medline]
    7. Bansil P, Keenan NL, Zlot AI, Gilliland JC. Health-related information on the Web: results from the HealthStyles Survey, 2002-2003. Prev Chronic Dis 2006 Apr;3(2):A36 [FREE Full text] [Medline]
    8. Anderson JG, Rainey MR, Eysenbach G. The impact of CyberHealthcare on the physician-patient relationship. J Med Syst 2003 Feb;27(1):67-84. [Medline]
    9. Culver JD, Gerr F, Frumkin H. Medical information on the Internet: a study of an electronic bulletin board. J Gen Intern Med 1997 Aug;12(8):466-470 [FREE Full text] [Medline]
    10. Varlamis I, Apostolakis I. Medical informatics in the Web 2.0 era. In: Tsihrintzis GA, Virvou M, Howlett RJ, Jain LC, editors. New Directions in Intelligent Interactive Multimedia. Berlin, Germany: Springer; 2008:513-522.
    11. Kernisan LP, Sudore RL, Knight SJ. Information-seeking at a caregiving website: a qualitative analysis. J Med Internet Res 2010;12(3):e31 [FREE Full text] [CrossRef] [Medline]
    12. Ginossar T. Online participation: a content analysis of differences in utilization of two online cancer communities by men and women, patients and family members. Health Commun 2008;23(1):1-12. [CrossRef] [Medline]
    13. Colineau N, Paris C. Talking about your health to strangers: understanding the use of online social networks by patients. New Rev Hypermedia Multimedia 2010 Apr;16(1-2):141-160. [CrossRef]
    14. Hughes B, Joshi I, Lemonde H, Wareham J. Junior physician's use of Web 2.0 for information seeking and medical education: a qualitative study. Int J Med Inform 2009 Oct;78(10):645-655. [CrossRef] [Medline]
    15. Domingo MC. Managing healthcare through social networks. Computer 2010 Jul;43(7):20-25. [CrossRef]
    16. Durant KT, McCray AT, Safran C. Identifying temporal changes and topics that promote growth within online communities: a prospective study of six online cancer forums. Int J Comput Models Algorithms Med 2011 Apr;2(2):1-22 [FREE Full text] [CrossRef] [Medline]
    17. Cho J, Noh H, Ha MH, Kang SN, Choi J, Chang YJ. What kind of cancer information do Internet users need? Support Care Cancer 2011 Sep;19(9):1465-1469. [CrossRef] [Medline]
    18. Macias W, Lewis LS, Smith TL. Health-related message boards/chat rooms on the Web: discussion content and implications for pharmaceutical sponsorships. J Health Commun 2005 Apr;10(3):209-223. [CrossRef] [Medline]
    19. Schultz PN, Stava C, Beck ML, Vassilopoulou-Sellin R. Internet message board use by patients with cancer and their families. Clin J Oncol Nurs 2003;7(6):663-667. [CrossRef] [Medline]
    20. Attard A, Coulson NS. A thematic analysis of patient communication in Parkinson's disease online support group discussion forums. Comput Hum Behav 2012 Mar;28(2):500-506. [CrossRef]
    21. Greaves F, Ramirez-Cano D, Millett C, Darzi A, Donaldson L. Use of sentiment analysis for capturing patient experience from free-text comments posted online. J Med Internet Res 2013;15(11):e239 [FREE Full text] [CrossRef] [Medline]
    22. Mohd Roffeei SH, Abdullah N, Basar SKR. Seeking social support on Facebook for children with autism spectrum disorders (ASDs). Int J Med Inform 2015 May;84(5):375-385. [CrossRef] [Medline]
    23. Jin J, Yan X, Li Y, Li Y. How users adopt healthcare information: an empirical study of an online Q&A community. Int J Med Inform 2016 Feb;86:91-103. [CrossRef] [Medline]
    24. Klemm P, Wheeler E. Cancer caregivers online: hope, emotional roller coaster, and physical/emotional/psychological responses. Comput Inform Nurs 2005;23(1):38-45. [Medline]
    25. Buis LR, Whitten P. Comparison of social support content within online communities for high- and low-survival-rate cancers. Comput Inform Nurs 2011 Aug;29(8):461-467. [CrossRef] [Medline]
    26. Sadah SA, Shahbazi M, Wiley MT, Hristidis V. A study of the demographics of Web-based health-related social media users. J Med Internet Res 2015;17(8):e194 [FREE Full text] [CrossRef] [Medline]
    27. Castleton K, Fong T, Wang-Gillam A, Waqar MA, Jeffe DB, Kehlenbrink L, et al. A survey of Internet utilization among patients with cancer. Support Care Cancer 2011 Aug;19(8):1183-1190. [CrossRef] [Medline]
    28. Zheng R, Li J, Chen H, Huang Z. A framework for authorship identification of online messages: writing-style features and classification techniques. J Am Soc Inf Sci 2006 Feb 01;57(3):378-393. [CrossRef]
    29. Abbasi A, Chen H. Writeprints: a stylometric approach to identity-level identification and similarity detection in cyberspace. ACM Trans Inf Syst 2008 Mar 01;26(2):1-29. [CrossRef]
    30. Zeng QT, Tse T. Exploring and developing consumer health vocabularies. J Am Med Inform Assoc 2006;13(1):24-29 [FREE Full text] [CrossRef] [Medline]
    31. Zeng QT, Tse T, Divita G, Keselman A, Crowell J, Browne AC, et al. Term identification methods for consumer health vocabulary development. J Med Internet Res 2007 Feb 28;9(1):e4 [FREE Full text] [CrossRef] [Medline]
    32. Keselman A, Smith CA, Divita G, Kim H, Browne AC, Leroy G, et al. Consumer health concepts that do not map to the UMLS: where do they fit? J Am Med Inform Assoc 2008;15(4):496-505 [FREE Full text] [CrossRef] [Medline]
    33. Doing-Harris KM, Zeng-Treitler Q. Computer-assisted update of a consumer health vocabulary through mining of social network data. J Med Internet Res 2011 May 17;13(2):e37 [FREE Full text] [CrossRef] [Medline]
    34. Goryachev S, Kim H, Zeng-Treitler Q. Identification and extraction of family history information from clinical reports. AMIA Annu Symp Proc 2008 Nov 06:247-251 [FREE Full text] [Medline]
    35. Witten I, Frank E. Data Mining: Practical Machine Learning Tools and Techniques. San Francisco, CA: Morgan Kaufmann Publishers; 2011.
    36. Lu Y, Zhang P, Deng S. Exploring health-related topics in online health community using cluster analysis. 2013 Presented at: 46th Hawaii International Conference on System Sciences; January 7-10, 2013; Wailea, HI, USA p. 802-811. [CrossRef]
    37. Esuli A, Sebastiani F. SENTIWORDNET: a publicly available lexical resource for opinion mining. 2006 Presented at: 5th Conference on Language Resources and Evaluation; May 24-26, 2006; Genoa, Italy p. 417-422.
    38. Tetlock PC. Giving content to investor sentiment: the role of media in the stock market. J Finance 2007;62(3):1139-1168. [CrossRef]
    39. Tetlock PC, Saar-Tsechansky M, Macskassy S. More than words: quantifying language to measure firms' fundamentals. J Finance 2008 May 09;63(3):1437-1467. [CrossRef]
    40. Qiu B, Zhao K, Mitra P, Wu D, Caragea C, Yen J, et al. Get online support, feel better: sentiment analysisdynamics in an online cancer survivor community. 2011 Presented at: 3rd IEEE International Conference on Social Computing; October 9-11, 2011; Boston, MA, USA.
    41. Denecke K, Taytsarau M, Palpanas T, Brosowski M. Topic-related sentiment analysis for discovering contradicting opinions in weblogs. Technical report #DISI-09-037. Trento, Italy: Dipartimento di Ingegneria e Scienze Dell'Informazione; 2009.   URL: http://eprints.biblio.unitn.it/1644/1/037.pdf [accessed 2017-03-29] [WebCite Cache]
    42. Blank TO, Adams-Blodnieks M. The who and the what of usage of two cancer online communities. Comput Hum Behav 2007 May;23(3):1249-1257. [CrossRef]
    43. Grbich C, Parker D, Maddocks I. The emotions and coping strategies of caregivers of family members with a terminal cancer. J Palliat Care 2001;17(1):30-36. [Medline]
    44. Flaskerud JH, Carter PA, Lee P. Distressing emotions in female caregivers of people with AIDS, age-related dementias, and advanced-stage cancers. Perspect Psychiatr Care 2000;36(4):121-130. [Medline]


    Abbreviations

    FM: Fowlkes-Mallows.
    MeSH: Medical Subject Headings
    UMLS: Unified Medical Language System


    Edited by G Eysenbach; submitted 02.12.16; peer-reviewed by T Vaughan, B Bie; comments to author 10.02.17; revised version received 27.02.17; accepted 05.03.17; published 07.04.17

    ©Yingjie Lu, Yang Wu, Jingfang Liu, Jia Li, Pengzhu Zhang. Originally published in the Journal of Medical Internet Research (http://www.jmir.org), 07.04.2017.

    This is an open-access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on http://www.jmir.org/, as well as this copyright and license information must be included.