Behavioral Characteristics of Medical Students on Social Media Platforms: Topic Modeling and Social Network Analysis.

Background: Medical students experience sustained academic, clinical, and psychosocial pressures. Online forums provide an informal space where students seek information, express concerns, and exchange peer support. However, few studies have analyzed these interactions longitudinally based on real-world data from online social platforms. Objective: This study aimed to integrate topic modeling, temporal topic analysis, and social network analysis (SNA) to investigate evolving concerns, network positions, and influence structures within the bulletin board system of a medical university over a 7-year period. Methods: We collected all posts and comments from the "Medical Center" board of the Peking University bulletin board system forum between 2017 and 2023. After excluding irrelevant posts, nontextual elements, blank content, and noise terms, the textual corpus was segmented and standardized for analysis. Latent Dirichlet allocation (LDA) was used for topic modeling to identify primary discussion topics. The number of topics was selected through 5-fold cross-validation by evaluating candidate models with 5 to 20 topics using perplexity. Annual topic prevalence was calculated to examine temporal changes in student concerns. SNA was also used to construct interaction networks. Users were defined as nodes, and directed edges were established when one user replied to another user's post or comment. Degree centrality, eigenvector centrality, and betweenness centrality were calculated to identify active users, influential hubs, and bridging users. A topic-network coupling analysis further linked topic categories with the centrality measures of users. Results: The corpus included 4639 posts and 58,942 comments. The annual post numbers fluctuated over the years, reaching their lowest point in 2019 and peaking in 2022. Posting activity increased after 8 AM, remained relatively stable from noon to 10 PM, and declined after 10 PM. Seven major topics of concern were identified, with feedback on living facilities (1025/4639, 22.1%) being the most prevalent, followed by information inquiry (1004/4639, 21.6%) and medical education (883/4639, 19.0%). Temporal analysis showed that concerns related to living facilities remained a recurring topic across years, and emotional support remained a low-frequency but persistent topic. The overall network was sparse, with a small subset of users occupying active, influential, or bridging positions. Topic-network coupling showed that high-volume topics were not necessarily network-central: feedback on living facilities had the largest discussion volume but the lowest median centrality values, whereas current affairs had the highest median degree and eigenvector centrality, and psychological and emotional support had the highest median betweenness centrality. Conclusions: Online forums can reveal not only what students discuss but also how concerns persist, change, and circulate through peer interaction networks. Combining topic modeling with network analysis may help educational administrators identify persistent concerns related to the learning environment, time-sensitive issues, and structurally important topics requiring targeted communication or support strategies.