The Give And Take: Automated Classification of Types of Support Given and Received over an Autoimmune Hepatitis (AIH) Online Support Group (Preprint)

<sec> <title>BACKGROUND</title> Autoimmune Hepatitis (AIH) is a rare disease involving the body attacking its own liver and affects fewer than 1.2 out of 100,000 people in the United States. Provider support for the disorder can be limited due to the frequency of provider appointments; as is the case with many rare diseases, AIH’s patients turn to peer support from online social media venues to supplement the advice of their providers. </sec> <sec> <title>OBJECTIVE</title> Here, we characterize the types and dynamics of support exchanged over an AIH-related Facebook™ page. </sec> <sec> <title>METHODS</title> Data from a coauthor-administered AIH-related Facebook™ page were downloaded from the browser front end. A brief literature review was performed to elucidate baseline types of support that could be exchanged over health-related social media venues. A portion of the resulting group user communications were qualitatively analyzed to determine the types of support exchanged within them, and key terms that indicated these types of support. Key terms were then used in dictionaries to drive a computational classification algorithm to classify the remaining user communications. The nature of communication (by type of support exchanged) was compared to various characteristics (sole poster; sole commenter; tenure at time of communication) of the user and the communication itself (post vs. comment status). </sec> <sec> <title>RESULTS</title> In the literature review, two directions (inbound/requesting and outbound/offering) and two content types (advice/information and emotional/social) of support were found. In addition to the 3 out of 4 permutations offered by these types and directions, three additional categories (asking the coauthor administrator [CSL] for advice; personal inquiry in providing support; grateful acknowledgement of support) were found directly from annotation of user communications. The search algorithms powered by the created dictionaries yielded reliability (F1) scores ranging from 0.308 (offering advice/information support) to 0.700 (offering emotional/social support). Significant differences existed across communication type (post vs. comment) when compared to support direction; the requesting of support primarily was seen in posts and the offering of support primarily in comments. Users who were deleted were more likely to have requested advice from CSL (OR 6.445, 95% CI 5.454-7.616, P&lt;.001) but significantly less likely to offer emotional/social support (OR 0.155, 95% CI 0.069-0.348, P&lt;.001). A longer tenure in the group correlated qualitatively with more offering of support; such tenure likewise correlated with less requesting of support, with the exception of requesting advice from CSL. Finally, CSL’s own support metrics revealed that he was significantly more likely to offer advice compared (OR 2.824, 95% CI 2.255-3.538, P&lt;.001) to other group members, and less likely to perform personal inquiries in provision of support (OR 0.401, 95% CI 0.178-0.900, P=.022). </sec> <sec> <title>CONCLUSIONS</title> The generated algorithm shows modest but reasonable gold standard comparison (F1)-based reliability in detecting support types by content type and direction. Conclusions drawn on correlations between user characteristics and detected communication types lend further intuition-based reliability to the algorithm. Future research is needed to more accurately detect the offering of advice/information and more importantly, to go beyond simply the types of support and to characterize the reasons of support. </sec>