Using Natural Language Processing to Track Negative Emotions in the Daily Lives of Adolescents

Area of Science:

• Computational Psychiatry
• Affective Computing
• Adolescent Psychology

Background:

• Tracking adolescent emotion fluctuations is crucial for understanding mood dynamics and early detection of affective disorders.
• Text-based analysis offers a scalable method for monitoring daily emotional states.

Purpose of the Study:

• To compare nomothetic (group-level) and idiographic (individualized) models for predicting adolescent negative affect (NA) from text features.
• To evaluate various Natural Language Processing (NLP) techniques for capturing within-person emotion fluctuations.

Main Methods:

• Analyzed 7,680 ecological momentary assessment (EMA) text responses from 97 adolescents (ages 14-18).
• Extracted text features using dictionary-based methods, topic modeling, and GPT-derived emotion ratings.
• Employed Random Forest and Elastic Net Regression for NA prediction, comparing nomothetic and idiographic approaches.

Main Results:

• Idiographic models integrating diverse NLP features showed comparable R-squared to nomothetic models but lower prediction error (RMSE), enhancing within-person precision.
• Significant individual differences observed in model performance and predictive linguistic features.
• Personalized model selection achieved high correlations between predicted and observed emotion scores (90.7-94.8% of participants).

Conclusions:

• Idiographic models offer superior predictive precision for adolescent NA compared to nomothetic models when sufficient individual data is available.
• A flexible, personalized approach to emotion monitoring can leverage text data for contextual insights and inform interventions.
• Text-based emotion tracking, particularly with individualized models, holds promise for enhancing mental health monitoring in adolescents.