Predicting social anxiety disorder based on communication logs and social network data from a massively multiplayer online game: Using a graph neural network.

Abstract

Aim: Social anxiety disorder (SAD) is a mental disorder that requires early detection and treatment. However, some individuals with SAD avoid face-to-face evaluations, which leads to delayed detection. We aim to predict individuals with SAD based on their communication logs and social network data from a massively multiplayer online game (MMOG).

Method: The study included 819 users of Pigg Party, a popular MMOG in Japan. Participants completed the Japanese version of the Liebowitz Social Anxiety Scale (LSAS-J) and a social withdrawal scale (hikikomori) questionnaire. Participants scoring ≥60 on the LSAS-J were classified as having SAD, while those scoring <60 were classified as not having SAD (non-SAD). A total of 142,147 users' communication logs and 613,618 social edges from Pigg Party were used as input to predict whether participants had SAD or non-SAD. Graph sample and aggregated embeddings (Graph SAGE) was utilized as a graph neural network model.

Results: Individuals with SAD were more likely to be socially withdrawn in the physical community (hikikomori), had fewer friends, spent less time in other users' virtual houses, and showed lower entropy in their visitation times in MMOG. Based on their social network data, the Graph SAGE model predicted SAD, with an F1 score of 0.717.

Conclusion: The communication logs and social network data in an MMOG include indicators of interpersonal avoidance behaviors, which is typical of individuals with SAD; this suggests their potential use as digital biomarkers for the early detection of SAD.