Bridging the conversational gap in epilepsy: Using large language models to reveal insights into patient behavior and concerns from online discussions.

Abstract

Objective: This study was undertaken to explore the experiences and concerns of people living with epilepsy by analyzing discussions in an online epilepsy community, using large language models (LLMs) to identify themes, demographic patterns, and associations with emotional distress, substance use, and suicidal ideation.

Methods: We analyzed 56 970 posts and responses to them from 21 906 users on the epilepsy forum (subreddit) of Reddit and 768 504 posts from the same users in other subreddits, between 2010 and 2023. LLMs, validated against human labeling, were used to identify 23 recurring themes, assess demographic differences, and examine cross-posting to depression- and suicide-related subreddits. Hazard ratios (HRs) were calculated to assess the association between specific themes and activity in mental health forums.

Results: Prominent topics included seizure descriptions, medication management, stigma, drug and alcohol use, and emotional well-being. The posts on topics less likely to be discussed in clinical settings had the highest engagement. Younger users focused on stigma and emotional issues, whereas older users discussed medical treatments. Posts about emotional distress (HR = 1.3), postictal state (HR = 1.4), surgical treatment (HR = .7), and work challenges (HR = 1.6) predicted activity in a subreddit associated with suicidal ideation, whereas emotional distress (HR = 1.5), surgical treatment (HR = .6), and stigma (HR = 1.3) predicted activity in the depression subreddit. Substance use discussions showed a temporal pattern of association with seizure descriptions, implying possible opportunities for intervention.

Significance: LLM analysis of online epilepsy communities provides novel insights into patient concerns often overlooked in clinical settings. These findings may improve patient-provider communication, inform personalized interventions, and support the development of patient-reported outcome measures. Additionally, hazard models can help identify at-risk individuals, offering opportunities for early mental health interventions.