Language sentiment predicts changes in depressive symptoms

Area of Science:

• Psychiatry
• Computational Linguistics
• Natural Language Processing

Background:

• Depression is a significant public health issue requiring predictive tools.
• Previous studies indicate depression affects language use, but predictive power is unclear.

Purpose of the Study:

• To determine if language sentiment in brief written responses predicts changes in depression.
• To compare the predictive accuracy of human raters, LLMs, and LIWC for depression symptom changes.

Main Methods:

• Two studies involving 467 participants who provided written responses and completed depression assessments (PHQ-9).
• Sentiment analysis of written responses by human raters (N=470), ChatGPT 3.5/4.0, and LIWC.
• Mood dynamics quantified via a risky decision-making task and momentary happiness measurements.

Main Results:

• Language sentiment, assessed by human raters and LLMs, predicted a three-week increase in depressive symptoms.
• LIWC sentiment analysis did not predict symptom changes.
• Language sentiment correlated with current mood but independently predicted future symptom changes.

Conclusions:

• Sentiment analysis of brief written responses using AI tools is a scalable method for predicting future depression.
• AI-powered sentiment analysis matches human performance in predicting psychiatric symptom changes.
• This approach offers a novel tool for early identification and intervention in depression.