Fine-Grained Emotional Characterization of Dementia Caregivers in Online Support Communities Using Large Language Models

Background: Dementia caregiving carries substantial emotional and psychological consequences, but most evidence comes from structured surveys and interviews that are resource-intensive and may incompletely capture spontaneous, contextual experience. Scalable methods to characterize caregiver experience from real-world narratives are lacking. Objective: To characterize how emotional expression among dementia caregivers varies by caregiving relationship and context, using structured extraction of caregiver narratives from an online support community at scale. Methods: We analyzed 7,198 publicly available posts from 3,350 authors across three forums of an online dementia caregiver community. A large language model (LLM) extracted caregiver role and relationship, caregiving objective, emotional valence (-1 to +1), and emotional themes from each post. Reproducibility of the extracted annotations was assessed through agreement between two independent LLMs, using Cohen's {kappa} for categorical fields and correlation for continuous valence. Results: Caregiver narratives were predominantly negative (89.3% of posts; mean valence -0.46), with interpretable structure. Parent caregivers (-0.51) and adult children caring for fathers (-0.52) and mothers (-0.50) expressed more negative valence than spouse caregivers (-0.47). Emotional valence varied most by post objective - most negative for immediate safety/crisis (-0.73), family conflict (-0.61), and end-of-life (-0.55) contexts, and net-positive only for resource sharing (+0.05). Even net-positive posts often retained concern alongside hope or gratitude rather than expressing uniform positivity. Concern, frustration, and sadness were the most prevalent emotional themes. Greater cumulative posting activity was associated with more positive expression. Inter-model agreement was high for relationship category ({kappa}=0.80) and emotional valence (r=0.95). Conclusions: Caregiver emotional expression is systematically patterned by caregiving relationship and by the objective of a post, with crisis and family conflict situations most negative and resource sharing the only net positive context, in ways that coarse sentiment or topic-modeling approaches have not shown. Emotional valence reflects expressed experience rather than clinical burden. Applied at scale, structured LLM extraction complements survey-based methods and could support distress screening and longitudinal monitoring of caregiver experience and inform the design of caregiver-support programs.