Monitoring Depression Trends on Twitter During the COVID-19 Pandemic: Observational Study.

Area of Science:

• Computational linguistics and mental health informatics.
• Application of artificial intelligence in public health surveillance.
• Analysis of social media data for psychological insights.

Background:

• The COVID-19 pandemic significantly impacted global mental health, with anecdotal evidence of increased depression.
• A lack of systematic studies on depression detection and monitoring during the pandemic.
• Need for scalable methods to track mental health trends during public health crises.

Purpose of the Study:

• To develop an automated, scalable method for creating large-scale depression user datasets.
• To validate transformer-based deep learning models for identifying depression from everyday language.
• To assess the importance of psychological text features in depression classification and monitor trends.

Main Methods:

• Creation of the largest English Twitter depression dataset (2575 users) using a regular expression-based search.
• Training and evaluation of three transformer-based deep learning models for depression classification.
• Development of a fusion classifier combining deep learning, psychological text features, and demographic data.

Main Results:

• The fusion model achieved 78.9% accuracy in identifying depression.
• Key features included conscientiousness, neuroticism, first-person pronouns, and expressions of sadness.
• Monitoring revealed depressive users responded later to the pandemic; Florida showed lower depression levels than NY and CA.

Conclusions:

• An efficient method for analyzing depression levels on Twitter was proposed.
• The study highlights COVID-19's mental health impact and the utility of noninvasive monitoring systems.
• The developed system is adaptable for monitoring mental health during future public health events.