The relationship between text message sentiment and self-reported depression.

Area of Science:

• Digital phenotyping
• Computational psychiatry
• Mobile health technology

Background:

• Personal sensing offers potential for identifying behavioral markers of depression.
• Limited research exists on personal sensing for cognitive and affective states.
• Digital language in text messages can serve as a marker for these states.

Purpose of the Study:

• To correlate privacy-preserving sentiment analysis of text messages with self-reported depression severity.
• To develop machine learning models for predicting depression status using text and sensor data.

Main Methods:

• A 16-week longitudinal observational study with 219 U.S. adults.
• Participants used a personal sensing app for self-report depression assessments (PHQ-8), phone sensor data collection, and anonymized text message sentiment analysis.
• Machine learning models were trained to predict depression status using text and sensor features.

Main Results:

• Language categories like depression, emotion, and personal pronoun use strongly correlated with self-reported depression.
• Classification models predicted binary depression status with a leave-one-out AUC of 0.72 using text features alone.
• Combining text features with smartphone sensor data improved prediction AUC to 0.76.

Conclusions:

• Text message sentiment analysis, particularly when integrated with personal sensor data, can effectively predict depression.
• This approach holds potential for comprehensive mental health monitoring and intervention.