Contextual insights from large language models in ambulatory assessment: Differential reactivity of affect and self-esteem to daily events in borderline personality disorder

Ambulatory assessment (AA) research often relies on closed-ended context checklists, whereas open-ended event descriptions preserve richer idiographic context but are rarely analyzed at scale. Borderline personality disorder (BPD) is characterized by marked instability, providing a clinically informative use case for testing scalable approaches to contextual dynamics in daily life. We investigated the association of event context and sentiment (i.e., emotional connotation) derived using a large language model (LLM) with self-reported momentary calmness, affective valence, and self-esteem in a sample of 265 women (131 individuals with BPD and 134 healthy controls) undergoing a 4-day high-frequency AA with 12 hourly assessments per day. We evaluated an automated LLM pipeline for classifying 10,844 free-text entries and for rating event-related sentiment, enabling quantitative modeling of context and appraisal in AA. The free-text entries were classified into five categories: daily routines and household, leisure and recreation, social and personal relationships, work and professional engagements, health and well-being, plus an indeterminate category. ... mehrResults showed the BPD group had lower values across outcomes and reacted with significantly lower calmness and self-esteem during social/personal events, as well as significantly lower self-esteem during health- and well-being-related activities, compared with controls. When sentiment was modeled alongside event categories, more positive-than-usual sentiments were associated with higher calmness, valence, and self-esteem, whereas more negative-than-usual sentiments were associated with lower states, with larger sentiment reactivity in BPD for calmness, valence and self-esteem. By employing LLMs for automated text categorization and sentiment analysis of free-text entries within AA research, we have effectively incorporated contextual information into our analysis.