The proposed research was conducted in the context of two different OHCs: one for tobacco cessation and another for DSM.

QuitNet (QN) is one of the largest OHCs promoting tobacco cessation amongst its users [44]. The users of this community are usually smokers willing to quit or ex-smokers willing to stay abstinent. Initial studies have shown a strong correlation of an individual’s participation in QN with abstinence compared to individuals who do not participate in such communities [45]. The data set used in this study was drawn from a previously studied quality improvement database spanning from 2000 to 2015, consisting of nearly 2.46 million peer interactions organized into 289,456 unique threads exchanged by 64,884 unique users. Using the inbuilt sample() function of the random module in Python, we randomly selected a subset of 2005 messages from this data set, organized into 132 unique threads posted by 884 unique community users.

The American Diabetes Association (ADA) support community is an online support group for individuals who have diabetes (type 1, type 2, or prediabetes) to engage with their peers as well as their caregivers [46]. The community users interact with one another on a wide variety of topics ranging from medication use, diet, and physical activity to daily monitoring of blood glucose levels. Even though the outcomes among type 1 diabetes mellitus (T1DM), T2DM, or prediabetes are impacted by behaviors (eg, lifestyle, medication use, self-monitoring of blood glucose) that can be heavily influenced by an individual’s social infrastructure, for this research, we focused on interactions related to T2DM. The data set spanned from 2015 to 2018, consisting of nearly 58,965 peer interactions specific to T2DM organized into 5829 unique topics posted by 1909 users. Using the inbuilt sample() function of the random module in Python, we randomly selected a subset of 1006 messages from this data set, organized into 99 unique topics posted by 141 unique community users.

Based on our prior research that focused on the qualitative analysis of digital communication from these OHCs [22,39], we expanded our data sets for this study in order to attain thematic saturation and a justifiable data set size amenable to manual analysis. Further, we selected the given size ranges for the two data sets in order to create a gold-standard annotated data set that can be efficiently used for downstream automated text mining techniques.

Further, the associations among user behavior and engagement profiles with speech acts and communication themes were evaluated by statistical analysis approaches, such as the chi-square test or the Fisher test, depending on the sample size. We also used the Cramer V to assess their correlation strengths. All statistical analyses were performed using the R programming language (using the stats package), and the significance level was P<.05.

The qualitative analysis of the 3011 messages exchanged by QN and ADA community users provided insights into the nature of the thematic interest of the community users and the variance in the expression of intentions among them (Figures 2 and 3).

The associations between all categories of speech acts and the five most prevalent categories of communication themes within the QN and ADA communities are shown in Tables 2 and 3, where P and V refer to the P value and the Cramer V value, respectively. In concrete terms, it means that in the QN community, the emotion speech act, for example, was often used in the context of the user’s readiness to quit smoking or was often expressed when a user participated in community-specific traditions (eg, bonfires, pledges). Similarly, stance and statement speech acts, for example, were often prevalent in the discussions outlining motivations to quit smoking (teachable moments). Declarative and directive speech acts, for example, were often used in the context of a user’s progress with smoking cessation. The strength of association was highest for the desire speech act within the social support theme (Cramer V=.26), followed by the directive speech act within the social support theme (Cramer V=.20) and the declarative speech act within the traditions theme (Cramer V=.15); see Table 2.

In the ADA community, for example, the assertion speech act was mostly used in the context of a user’s readiness to manage lifestyle behaviors that influence diabetes. The commissive speech act was often prevalent within the social support, readiness, and teachable moments communication themes. Messages conveying the user’s experiences with hurdles (teachable moments) often contained expressive and stance speech acts. The social support theme often had higher prevalence of the emotion, expressive, and stance speech acts (see Table 3).

Figure 5 shows the distribution of speech acts across the three user engagement levels for both communities, where the x axis represents the speech act categories and the y axis represents the percentage of messages in which a given speech act was present. In the QN community, low-engagement users had a high prevalence of the desire (2%) and expressive (18%) speech acts compared to high- or medium-engagement users, which indicates that low-engagement users are willing to incorporate behavior change but somehow need more motivation to engage with other users of the community who have successfully quit. Emotion (37%) and declarative (28%) speech acts were more frequently expressed by high-engagement users compared to other engagement levels, showing how active users in the QN community play an essential role in providing emotional support to other users in this community. Statistical analysis showed that there is a significant association (P<.05) between the overall expression of speech acts and user engagement levels based on the counts of messages expressing different speech acts exchanged by individual user groups, while the overall strength of association is weak (Cramer V=.07). Post hoc analysis revealed that there is a significant association (P<.05) between the assertion, declarative, question, and statement speech acts and the user engagement levels in the QN community (see Table 4).

Within the ADA community, the low-engagement users expressed the desire (6%) and declarative (17%) speech acts to a greater extent compared to other engagement levels. The high-engagement users of the ADA community provided directives (29%) to other users compared to other engagement levels, which indicates that these users might be able to appropriately guide other community users per users’ concerns. High-engagement users within the ADA community were also quite opinionated as they expressed high levels of stance (22%). Statistical analysis showed that there is a significant association (P<.05) between the overall expression of speech acts and user engagement levels, while the overall strength of association is weak (Cramer V=.13). Post hoc analysis revealed that there is a significant association (P<.05) between the commissive, declarative, directive, and stance speech acts and the user engagement levels in the ADA community (see Table 5).

The distribution of different speech acts by the percentage of messages exchanged per the QN community user’s abstinence status is shown in Figure 6, where the x axis represents the speech act categories and the y axis represents the percentage of messages in which a given speech act was present. Abstinent users had different informational needs than non-abstinent users. Their role and context in making contributions differed from those new to the community or in a different stage of behavior change. They had a high prevalence of emotion (28%) and expressive (19%) speech acts. The other commonly occurring speech acts among abstinent users were assertion (18%), directive (17%), and statement (17%). The most highly prevalent speech act among non-abstinent users was directive (21%), and they also expressed themselves with the commissive (17%) and declarative (13%) speech acts through promises to themselves or the community. The other commonly occurring speech acts among non-abstinent users were emotion (13%) and statement (12%). The stance speech act was only expressed by abstinent users and had low prevalence (2%). The prevalence of the desire speech act was comparable across QN community users per their abstinence status—4% for abstinent users and 5% for non-abstinent users.

We also examined the distribution of different speech acts by the percentage of messages exchanged per the ADA community users’ medication status, as shown in Figure 7, where the x axis represents the speech act categories and the y axis represents the percentage of messages in which a given speech act was present. The most highly prevalent speech acts for both medication and nonmedication users were assertion (50% and 54%, respectively) and statement (both 38%). The other commonly occurring speech acts among medication and nonmedication users were directive (27% and 28%, respectively), stance (20% and 24%, respectively), and question (18% and 17%, respectively). Regarding the distribution of different speech acts by the percentage of messages exchanged per ADA community users’ diagnosis status (Figure 7), the most highly prevalent speech acts for newly diagnosed T2DM and advanced T2DM users were assertion (46% and 52%, respectively) and statement (38% and 39%, respectively). The newly diagnosed T2DM users had a higher prevalence of the stance (25%) speech act, while advanced T2DM users had a higher prevalence of the directive (29%) speech act. The users who incorporated lifestyle changes provided directives (29%) to others and also shared their health practices through the use of statements (40%). The users who incorporated lifestyle changes had a high prevalence of the assertion (54%), question (21%), and stance (26%) speech acts in their peer interactions compared with users who had not incorporated lifestyle changes (Figure 7).

Table 6 shows the results of the statistical analysis across user behavior profiles within the two OHCs, where P and V refer to the P value and the Cramer V value, respectively. There was a statistically significant (P=.002) association between the expression of speech acts and abstinence status in the QN community. Similarly, there was a statistically significant (P=.03) association between the prevalence of various communication themes and abstinence status in the QN community. There was a statistically significant (P=.007) association between the expression of speech acts and lifestyle change status in the ADA community. This significance was also noted in the prevalence of communication themes and lifestyle change status within the ADA community (P=.001). The relationship strength between speech acts and abstinent status (Cramer V=.22) and between speech acts and lifestyle change status (Cramer V=.12) was weak. There was no significant relationship between the expression of speech acts and the ADA community users’ medication or diagnosis status. Further analysis is required to identify the specific categories of speech acts that might contribute to these significant relationships.

Our study focused on analyzing the speech acts embedded within the user communications in OHCs using the mixed-methods approach in order to gain an understanding of the factors contributing toward behavior change in the context of two health care domains, tobacco cessation and diabetes self-management (DSM). The study of communication intent to understand and model health-related user communications is not without precedent. For example, earlier research used a similar pragmatics-based discourse analytic framework to investigate the expression of empathy within a T2DM Facebook support group and found that empathy may be expressed and perceived differently depending on the user’s perspectives [51]. Similarly, in our study, although management of T2DM may include promises of behavior change through diet, exercise, and increased adherence to medication regimens, commissive speech acts formed only a small fraction of the discussion. Although we used the 10 categories of speech acts from Searle's taxonomy of speech acts, other studies have focused on only stance-taking intentions of complaining or disagreeing [52] or relationship-building through use of the self-praise speech act to understand how this speech act instills positive values among the smoking cessation community users [53]. However, it is important to look at all the categories of speech acts in order to characterize what constitutes persuasive communications so that we can better enable individuals to improve their health-related behaviors through high-impact, just-in-time adaptive support.

Another study [54] used a smoking cessation public forum called SmokingisBad to understand how information is shared between initiators/help seekers and respondents/advice givers using the analytical lens of persuasion and interpersonal pragmatics. This study reported how advice givers focus more on the interpersonal side of interaction to provide motivation and support to initiators to become successful quitters [54]. In our study, too, we noticed statistically significant differences in the expression of communication themes as well as speech acts based on the abstinence status of QN community members. The statistically significant association between the speech acts and communication themes discussed among different user behavior subgroups (abstinent vs non-abstinent users, users with lifestyle changes vs no lifestyle changes) corroborates how the intent of peer interactions and social dialogue are unique to an individual’s behavior profile and also differ for a shared collection of communication themes. Abstinent users engaging with a smoking cessation community are likely to have different needs than those seeking support to stop smoking. Non-abstinent users making a promise to themselves or peers through the use of a commissive statement (eg, “I will stop smoking tomorrow.”) logically follows. Existing tobacco cessation interventions oftentimes target an individual’s readiness based on their stage of behavior change (as defined by the transtheoretical model of change [55]) and have proven to be effective in helping individuals reduce their nicotine intake [56,57]. In addition to the behavioral stage, our study enables the identification of the cognitive and behavioral state of individuals, as manifested in their peer interactions at a granularity that was not previously explored through incorporation of attributes that describe communication content and intent. The uniqueness of our study lies in understanding the intent of the community users, not just based on one’s conversations with their peers, but also based on their self-reported behavior profiles as well as observed engagement levels, thus offering intrapersonal and interpersonal contexts of psychosocial and behavioral domains crucial to self-management of risky behaviors and chronic health conditions.

Significant associations between speech acts, communication themes, and user engagement within the two OHCs also suggest variances in preferred methods of expression and peer discussions, depending on the engagement levels of users. It is also interesting to note that ADA community users with low engagement had a higher prevalence of the desire speech act in their communication with peers, reflecting on their needs to fulfill their self-management goals. However, most ADA community users with high engagement had a high prevalence of the directive speech act, which may reflect their ability to provide guidance to their peers, thus emphasizing the role of peer mentors and patient expertise [58]. This provides an additional layer of social complexity that can be harnessed in the design of various digital health tools through the inclusion of advanced interactions in the form of recommendation engines that facilitate meaningful peer and content connections, conversational agents for guided training, etc. Such technological implementations can take into consideration the specific intentions associated with the engagement status of users within an OHC in order to help low-engagement users who need motivation and social support by exposing or recommending them to high-engagement users who have been able to maintain constant engagement using such platforms or to relevant topics based on their thematic interests. The age group of the community users also affects their perceptions and engagement with technology and is another important factor to consider in order to maximize the potential of digital health tools to improve consumer engagement with such platforms to facilitate behavior change [59], which is outside the scope of our analysis due to data constraints associated with our data sets.

In terms of the associations between speech acts and communication themes, all of the speech acts (except question) were significantly associated with the social support theme within the QN community. Only a few speech acts (commissive, emotion, expressive, and stance) were significantly associated with the social support theme in the ADA community. As social support is delivered within the full range of speech acts, this indicates robust use of many different subtle forms of transmission, emphasizing how the messages are formed matters. Areas of differentiation in how these ideas are shared (and why it differs in areas such as progress, readiness for change, and traditions) indicate areas where coupling pragmatic action with message content has the potential for impact. The differences observed among online communities to discuss communication themes such as traditions, obstacles, and patient-related outcomes also highlight the distinctive approaches necessary to address independent communities. Previous research has shown that smokers who receive comprehensive cessation counseling over the telephone using techniques that focus on cognitive, emotional, and coping processes tend to have increased acceptance of cravings to smoke [60]. Similarly, in our study, we found that in the QN community, the directive speech act was highly prevalent in the cravings theme; thus, digital behavior change interventions can benefit from incorporating such additional insights into cravings to improve message crafting and content tailoring capabilities per the user’s needs.

Our study is not without limitations. Although QN and ADA messages were selected at random, the relatively small and uneven sample sizes may have resulted in inaccurate representations of the overall prevalence of speech acts and thematic emphasis. However, our sample of 3011 messages using qualitative research methods is appropriate for the study objectives. Even though we used empirically sound and reliable methods for coding the peer conversations, our analysis might be limited by subjective bias in the annotation process that is inherent in human coding in qualitative research [61]. Although we extracted the behavior profiles of the users who had self-reported their abstinence status within the QN forum messages and those who had created self-reported signatures within the ADA community, such extraction may be affected by incomplete or inaccurate self-report accounts and as such may not be representative of the general population. Moreover, our analysis did not consider sociodemographic and cultural factors, which can also result in differences in expressions. Our analysis was primarily focused on text-based user communications and did not consider other aspects of online user behavior within an OHC (eg, likes, dislikes, shares), which can also provide additional insights. OHCs can also be used for spreading health-related misinformation [62], which can have a negative influence on such behaviors. However, we did not consider the manifestation of misinformation as a separate theme within our data sets, which should be formalized through content-flagging mechanisms in future studies. The timeline of the data sets used in the study is not recent (2000-2015 for QN and 2015-2018 for ADA). We will attempt to obtain current data sets in future studies; however, we believe the findings still hold since the basic tenet of the social interactions remains the same. Future work should consider coding more messages in order to ensure category saturation for speech acts and communication themes so as to curate a balanced annotated data set. Our efforts to develop semiautomated methods to scale up the application of communication themes and speech act labels to large-scale data sets using deep learning models (eg, convolutional neural networks [CNNs]) [63], bidirectional encoder representations from transformers (BERT) [64]), and social influence models [65-67] are underway.

Digital health platforms, specifically OHCs, provide unprecedented opportunities for researchers to transform into empathetic listeners of people’s endeavors, from small-scale, 1-day-at-a-time to large-scale life-consuming behavior modifications. Such listening capabilities will help us glean the needs of individuals to initiate and sustain positive health changes. This paper proposes an analytical framework to enable deep social listening capabilities to uncover the unmet needs of individuals, thus laying the foundation for next-generation technology innovation efforts. Our efforts to overlay the communication intent implicit in peer interactions alongside content-specific theory-linked characterizations of social media discourse have provided insights into topic diversity and latent interactions of users within these topics. Such content-specific and intent-sensitive methodological framework can inform the development of cognitively enabled big data analytics and machine learning models that better harvest the digital footprint of social media users.