Skip to Content
SBMI Horizontal Logo

Context-aware Content-sensitive Modeling of Online Peer Interactions in Social Media: Towards Personalized Digital Experiences for Behavior Change and Health Promotion

Author: Tavleen Kaur Ranjit Singh, MS (2023)

Primary advisor: Sahiti Myneni, PhD

Committee members: Kirk Roberts, PhD; Kayo Fujimoto, PhD

PhD thesis, McWilliams School of Biomedical Informatics at UTHealth Houston.


Online health communities (OHCs) have emerged as prominent platforms for behavior modification. The digitization of peer interactions from these OHCs has afforded researchers rich avenues to examine and model multilevel mechanisms that drive behavior change among individuals. However, siloed approaches towards content (semantics) and context (pragmatics) modeling combined with limited incorporation of behavioral science theory in analyzing online peer conversations is an important methodological limitation. In this research, I present a novel methodological framework called Pragmatics to Reveal Intent in Social Media (PRISM), to facilitate granular characterization of peer interactions by combining multidimensional facets of human communication. Using this framework, I analyzed peer interactions from two different OHCs - QuitNet for tobacco cessation (n = ~2.39 million peer interactions from 2000- 2015) and American Diabetes Association (ADA) support community for diabetes self- management (n = ~73k peer interactions from 2012-2021). Firstly, a labeled set of peer interactions was generated (n = 2,005 for QuitNet; n = 1,501 for ADA) through manual annotation along the three attributes: communication themes (CTs), behavior change techniques (BCTs), and speech acts (SAs). Secondly, deep learning models were used to scale the qualitative codes to the entire datasets. Thirdly, the validated model was then applied to perform a retrospective analysis of these two datasets. Finally, using social network analysis (SNA), specifically affiliation exposure models and stochastic actor- oriented models, I portrayed large-scale patterns and relationships among the aforementioned communication attributes embedded in peer interactions in both communities.

Qualitative analysis showed that social support was the most prevalent CT in both communities; feedback and monitoring was the most commonly employed BCT among the users of both communities. The users most commonly expressed their intentions using emotion SA in the QuitNet community and assertion SA in the ADA community. I also evaluated the association between the communication attributes and across observed tiers of user engagement and self-reported behavior profiles. With additional in-domain pre-training, bidirectional encoder representations from Transformers (BERT) outperformed other deep learning models on the classification tasks from both datasets. Content-specific stochastic actor-oriented models (SAOMs) revealed that there is an inherent tendency for the communication to be reciprocated (p < 0.05). The results did not show an influence of network dynamics on the abstinence status of QuitNet users. However, engaging in dense triads via expressing comparison of behavior BCT with interactive turn-taking communication style SAs or feedback and monitoring BCT with push-in communication style SAs significantly affected the abstinence status of individuals in a way that triads influence the beginning stages of the quit (p < 0.05). SNA analysis from the ADA dataset showed that the community users exposed to other users by expressing specific content using push-in or interactive communication style SAs tended to engage more in the community. These network patterns integrated with theoretical underpinnings led to the generation of a knowledge graph that informs the personalization of digital health tools for behavior modification; for example, delivering motivating content using specific SAs (e.g., assertion or emotion) can positively influence users who are indirectly peers with abstinent users to start their quit. Future evaluation and trial designs for implementing these novel digital health infrastructures will be discussed