In this work, we used datasets from 2 different social media sources: cancerdusein.org and Facebook groups. Table 1 summarizes statistics from these 2 datasets.

The first dataset contained the forum posts from cancerdusein.org, a French health forum with more than 16,000 posts. These posts cover a large number of topics related to health issues. This forum is recommended to patients in a brochure of the Institut National du Cancer (INCA), which is the French reference organization in oncology. The forum is recommended for patients to exchange information and find comfort and potential solutions to their problems. It serves as an online cancer support community, where cancer patients, cancer survivors, and their families share information about cancer and their conditions. The second dataset contains posts from groups on Facebook, one of the most well-known social networks. We extracted 70,092 posts from 4 different public groups or communities on Facebook: Cancer du sein, Octobre rose 2014, Cancer du sein - breast cancer, and brustkrebs. We collected data from groups focusing on the adult population (the targeted users) and in which users were very active.

On both social media platforms, patients freely exchange information without the need for moderators to supervise discussions. New messages can either be added to an existing thread or be posted to open a new thread. In cancerdusein.org, a thread appears in exactly 1 of the 13 predefined subforums, for example, Discussion générale [general discussion], Vivre mon cancer au quotidien [daily life with my cancer], Les bonnes nouvelles [good news], or Récidives et combats au long cours [relapses and long-term battles]. In Facebook groups, there are no predefined topics to index the threads. Structuring topics according to the subforum structure is possible in cancerdusein.org, but this structure underlines the relatively uninformative and widely spread topics, covering a strongly unbalanced number of messages. Such indexing is not possible in Facebook groups. Interestingly, we propose to accomplish a finer analysis of topics in the next section, which further enables the presence of several topics within 1 message.

Texts on social media are often strongly heterogeneous and noisy, with many deviations from standards of spelling, syntax, and abbreviations, which impede efficient natural language processing. The French language has a rich spelling and grammar, characterized by special characters such as ç, various kinds of accented vowels (eg, é, è, ê, ë, â, and à), and many flexional variants. Additional rules exist for linking subsequent terms in certain situations (eg, the contraction du formed from de+le and the contraction des formed from de+les). As a consequence, automatic correction of text not obeying those rules is relatively difficult in practice. Furthermore, semantic analysis of texts is complicated by a large number of homonymy relationships: for example, pas can either mean step (noun) or can be the negation adverb not. As Balahur [42] and Farzindar and Inkpen [43] have pointed out, these linguistic peculiarities may affect classification performance. For this reason, we developed the following preprocessing steps.

Today, detection of latent semantic structures and topics has become a very active field of research in the text mining community. We focused on the LDA model [27], which has become a standard model for unsupervised topic detection from a text corpus. It is a probabilistic model with a hierarchical definition of its components. With the LDA model, we generated new documents from a given model. Based on the relatively simple and robust bag-of-words representation of text documents, it leaves the order of occurrence of terms and sentence structure out for consideration. For a given corpus of D documents, we first defined the relevant vocabulary V, a preprocessed collection of terms occurring in the corpus. Typical preprocessing steps include spelling correction, lemmatization, and the removal of noisy or irrelevant terms. To define a topic t, we associated a nonnegative weight ω_ti with each of the vocabulary’s terms, w_i, so that weights summed up to 1 (∑^V_i=1 ω_ti=1). In practice, each topic typically consisted of a relatively small number of terms with nonnegligible weight. An LDA model uses a fixed number K>1 of topics. For each document d, weights ω_dt≥0 indicate the occurrence probability of terms from topic t, where the sum of ω_dt over all topics t yields 1 (∑^K_t=1 ω_dt=1). If document d contains l_d terms (or “positions”), we associated a topic t_dj with each of the positions j=1,..., l_d, where the probability of associating topic t is α_dt. Finally, each position was filled with a term, w_dj, from the vocabulary, where the probability of using term w_i is ω_tdj.

The corpus-generation model is proposed by the algorithm shown in Figure 1.

The principal information that we can learn from using such a model on a corpus of text data is the structure of represented topics and the distribution of topics over the documents contained in the corpus. The high number of unknown parameters in this model makes inference challenging, yet Bayesian techniques such as Gibbs sampling [46] have proven reliable. Based on prior assumptions about the distribution of the weights of terms in topics and of topics in documents on a range from very uniform to very spiked, these inference techniques are applied to the data to estimate the posterior distributions of the model. Most importantly, the most likely topic structure and the occurrence probabilities for topics in each document are proposed. In this work, we considered a message as a document.

Besides K, 2 parameters often denoted as α and β strongly influence the distribution of topic probabilities for each of the messages. They are concentration parameters for the prior distributions of topics over a message (α) and of words over a topic (β). When α or β is smaller than 1 and decreases, prior mass concentrates closer and closer to the border of the simplex with spikes at each of its vertices. Then, 1 or fewer components (topics for α, words for β) carry strong probability in the mixture distribution. In the limit 0, a single component is selected with a probability of 1. On the contrary, when α or β is larger than 1 and increases, mass concentrates more and more in the barycenter of the simplex, leading to a mixture of the distribution, which is more and more balanced over all components. In the limit ∞, each component is selected with a probability of 1 over the number of components.

Now we will explain our choice of α based on the influence of α on the distribution of topic probabilities for messages and of term distributions for topics. When α=1, the prior distribution for the vector of topic probabilities corresponds to a uniform distribution on the simplex with K vertices. As α increases, the distribution concentrates more and more strongly toward the center of the simplex, such that most of the probabilities are closer to 1/ K. As α decreases, it concentrates more and more strongly toward the vertices, leading to some probabilities being further away from 1/ K. For fixed α, probabilities concentrate more and more around 1/ K as K increases. In Griffiths and Steyvers [47], values α=α₀/ K with the constant α₀=50 are encouraged, where dividing through K constantly keeps a certain complexity measure of the model. Exploratory analysis showed that α₀=50 led to very flat probability vectors in our case, which made it difficult to attribute a small number of topics for indexation to each message. On the other hand, smaller values of α₀ led to topics becoming more difficult to interpret due to flatter distribution of term probabilities within topics and similar dominating terms in multiple topics. After careful analysis of topics and posterior distributions for a range of values of α₀, we decided to fix α₀=10. Whereas higher values of α₀ yielded a better fit of the model in terms of its likelihood, it led to very flat posterior probabilities for the topic distribution of messages. As in Griffiths and Steyvers [47], we decided to fix the value of parameter β to 0.1 for our experiments.

There is evidence [48] that automatic choice of parameters through a model selection criterion may result in an unsatisfactory topic collection, whose interpretation is more challenging than topics associated with suboptimal values of the criterion. Often, the calculation of held-out likelihood is used, allowing for approaches such as likelihood cross-validation. However, the likelihood calculation is not trivial, and some standard methods produce inaccurate results (see [49]).

To avoid noisy topics that are difficult to interpret, it is useful to focus on terms with potential medical relevance. Here, we defined terms as sequences of words, and often there was only a single word. To begin, we used terms indexed in the French version of the Medical Subject Headings (MeSH) [50]. Then we added terms figuring in a list of breast cancer drugs (extracted from the online resource) or appearing in a list of nonconventional treatments (extracted from the French Wikipedia entry). We denoted this term set as MED. We retained 481,111 occurrences of 18,672 terms in 16,868 messages on cancerdusein.org, and 626,043 occurrences of 18,741 terms in 70,092 messages on Facebook. The resulting topics, often strongly dominated by a single term, appeared to be rather difficult to interpret by clinical experts, possibly due to the relatively small dimension of the term-document space. We categorized terms figuring in the representative terms according to their grammatical role: nouns/proper names (NN), verbs (V), and adjectives (A). Then, we extracted topics by applying LDA to the original MED term set, extended by terms according to scenarios MED+NN+V+A. Based on the exploratory inspection of topics extracted by LDA in the approaches presented in the following, we further removed a small number of strongly represented terms leading to strong noise (femme [woman], temps [time or weather]), and medically meaningless topics.

We succeeded in interpreting the 20 topics obtained from the output of our model on the cancerdusein.org corpus. Table 2 presents the 10 first topics and the top 10 words obtained by our model that were interpreted by an expert. Some relationships were established. In the QLQ-C30, we found matches for all of the topics except for global health status and QoL. In the QLQ-BR23 form, we matched all of the topics.

We succeeded in interpreting the 20 topics obtained from the output of our model on the Facebook corpus. Table 3 presents the 10 first topics and the top 10 words obtained by our model that were interpreted by an expert. Some relationships were established. In the QLQ-C30, we found matches for all of the topics except for role functioning, cognitive functioning, and global health status and QoL. In the QLQ-BR23 form, we matched all of the topics.

The first limitation of this study is the type of users, which produced the patient-authored text exploited in our process. Indeed, unless a group has formal gatekeeping of members, it is difficult to know for sure whether people posting to a forum or in a Facebook group are patients, survivors, health care professionals, care providers, family, or friends of patients. Consequently, topics extracted with our method may have been generated by users who do not have breast cancer. In particular, it has been known for decades that health information is sought principally by friends or family members, and then after that by patients [56]. In this work, we assumed that the relatives’ topics of interest were similar to patients’ topics of interest. However, in a previous work [57], we proposed a method to automatically deduce the role of the forum user. This method can be used at the beginning of our chain to exclude the posts of individuals who are not actual patients.

The second limitation is that we harvested data from only 1 forum and different Facebook groups. However, this forum is frequently recommended by French physicians to patients. It is also recommended by INCA, which is the French reference organization in oncology. We deliberately selected this forum and these Facebook groups to examine similarities and differences within and between these 2 particular communities. Of course, there are certainly many other online communities related to breast cancer, and the users in these 2 online communities were not necessarily representative of users of all breast cancer social media.

It is also important to note that our method can be easily applied to other diseases. For example, we can (1) use brain cancer forum data to align topics discussed by patients with items of the EORTC QLQ-C30 and the brain cancer module (QLQ-BN20) [58] questionnaires, and (2) use lung cancer forum data to align topics discussed by patients with items of the QLQ-C30 and the lung cancer module (QLQ-LC13) [59] We have already also applied a similar approach to study other social media data such as Twitter [60]. The main adaptation is relative to the acquisition of the patient terms, which are specific to the disease and the social media as mentioned in the Data Preprocessing section above.

A third limitation was the choice of LDA. LDA requires much manual tuning of its parameters, which vary from task to task. We spent a lot of time finding the best parameters so that the results could be interpreted meaningfully. Such analysis makes itself a sort of “overfitting” to the task at hand, making it very hard to generalize the method to other datasets and other tasks. However, we efficiently defined parameters of 2 types of text (forum and Facebook posts), which can be reused for other studies on comparable corpora.

Topics covered on social media focused on a specific domain, breast cancer. It was difficult to adjust the number of topics because topics were closed: all of the users were discussing breast cancer. When we adjusted the model and sought the optimal K with methods such as those used in other studies (eg, [47,61,62]), we obtained more than 50 topics. An interesting perspective was using the heuristic approach defined by Zhao et al [63] to determine an appropriate number of topics. This method is based on the rate of perplexity change [62,64]. This measure is commonly used in information theory to evaluate how well a statistical model describes a dataset, with lower perplexity denoting a better probabilistic model [63]. Finally, as in Arnold et al [65], we observed that an expert is not able to interpret so many topics. In this study, we manually fixed K=20. We interpreted all the topics with minimal redundancies.