Physical Activity, Sedentary Behavior, and Sleep on Twitter: Multicountry and Fully Labeled Public Data Set for...

Area of Science:

• Digital epidemiology
• Public health surveillance
• Machine learning applications in health

Background:

• Social media data offers potential for public health surveillance (PHS) but lacks trust and robust datasets.
• Traditional PHS data sources are often outdated, costly, and limited in scope.
• Developing reliable datasets is crucial for advancing supervised machine learning (ML) in PHS.

Purpose of the Study:

• To present a large-scale, multicountry, longitudinal, and fully labeled dataset for digital PASS surveillance research.
• To supplement the dataset with PHS-related metadata to support high-quality surveillance.
• To facilitate the development and evaluation of digital PHS systems.

Main Methods:

• Collected 122,135 PASS-related tweets from Twitter (Nov 2018 - Jun 2020) using the livestream API.
• Iteratively refined tweet selection using regex, NLP, ontologies, and linguistic analysis.
• Utilized Amazon Mechanical Turk for crowd-labeling with a quality control pipeline and ML validation.

Main Results:

• LPHEADA comprises 366,405 crowd-generated labels for 122,135 tweets from AU, CA, UK, US.
• Includes demographic and location metadata for each tweet, crucial for PHS.
• Validated dataset components using ML, latent semantic analysis, and linguistic analysis.

Conclusions:

• LPHEADA is a novel and comprehensive resource for digital PASS surveillance.
• Addresses the limitations of isolated, small-subset datasets currently available.
• Will serve as an invaluable resource for public health researchers and practitioners.