Using natural language processing to track negative emotions in the daily lives of adolescents

Tracking emotion fluctuations in adolescents' daily lives is important for understanding mood dynamics and identifying early markers of affective disorders. This study examines whether natural language processing (NLP) methods, including large language models (LLMs), can approximate adolescents' concurrent self-reported negative affect from ecological momentary assessment narratives. We compared nomothetic (group-level) and idiographic (individualized) models as well as different NLP techniques for capturing within-person emotion fluctuations. Ecological momentary assessment text responses from 97 adolescents (ages 14-18, 77.3% female, 22.7% male, N_EMA = 7,680) were analyzed. Text features were extracted using dictionary-based approaches, topic modeling, and emotion ratings derived from a LLM (GPT-4o). Random Forest and Elastic Net Regression models predicted negative affect from text features, and SHapley Additive exPlanations values were used to assess feature importance. Key findings, interactive visualizations, and model comparisons are available via the website: https://emotracknlp.streamlit.app/. Idiographic models combining text features from different NLP approaches performed comparably to nomothetic models and to Generative Pre-Trained Transformer 4 Omni (GPT-4o) alone in R² (nomothetic: .06-.11, idiographic: .06-.10, GPT-4o: .06-.10) but yielded lower root-mean-squared error (nomothetic: .52-.88, idiographic: .47-.76, GPT-4o: .85-1.36), improving within-person precision. Importantly, there were substantial between-person differences in model performance and predictive linguistic features. When selecting the best-performing model for each participant (e.g., nomothetic or idiographic), significant correlations between predicted and observed emotion scores were found for 90.7%-94.8% of participants. Findings suggest that although nomothetic models and LLMs offer scalability, idiographic models provide greater predictive precision with sufficient within-person data. A personalized approach selecting optimal models for each individual may enhance emotion monitoring and inform targeted interventions through contextual text insights. (PsycInfo Database Record (c) 2026 APA, all rights reserved).