Automatic Detection of Twitter Users Who Express Chronic Stress Experiences via Supervised Machine Learning and...

Area of Science:

• Computational Social Science
• Digital Health
• Natural Language Processing

Background:

• Americans experience significant chronic stress, exacerbated by events like the COVID-19 pandemic.
• Traditional stress measurement methods like surveys have limitations.
• Social media offers a valuable, underutilized resource for understanding stress experiences.

Purpose of the Study:

• To develop and evaluate an automated system for identifying Twitter users who self-report chronic stress.
• To leverage social media data for a novel approach to stress detection.

Main Methods:

• Collected tweets using the Twitter API with stress-related keywords.
• Filtered tweets using predefined text patterns and manually annotated them for self-reported chronic stress.
• Trained and evaluated multiple machine learning classifiers, including Bidirectional Encoder Representation from Transformers (BERT) and Support Vector Machines (SVM).

Main Results:

• Achieved high inter-annotator agreement (0.83 Cohen's kappa) on 4195 annotated tweets.
• The BERT-based classifier demonstrated superior performance with 83.6% accuracy, outperforming SVM (76.4%).
• Analysis of positive tweets revealed insights into stress sources and health impacts.

Conclusions:

• Automated identification of self-reported chronic stress on Twitter is feasible and effective.
• This approach holds significant potential for large-scale public health surveillance and intervention strategies.