Extraction of Explicit and Implicit Cause-Effect Relationships in Patient-Reported Diabetes-Related Tweets From 2017...

Area of Science:

• Computational linguistics
• Social media analytics
• Health informatics

Background:

• Diabetes distress requires understanding patient perspectives.
• Social media offers direct insight into patient experiences and disease understanding.
• Identifying causes of diabetes distress is key for intervention and prevention.

Purpose of the Study:

• To extract explicit and implicit cause-effect relationships from patient-reported diabetes tweets.
• To develop a methodology for understanding patient opinions and feelings from a causality perspective.
• To analyze the diabetes online community's shared observations regarding disease causes.

Main Methods:

• Collected over 30 million diabetes-related tweets (April 2017 - January 2021).
• Applied deep learning and natural language processing, focusing on personal/emotional content.
• Trained BERTweet and conditional random field (CRF) models with Bidirectional Encoder Representations from Transformers (BERT)-based features for cause-effect detection and extraction.

Main Results:

• Detected causal sentences with 68% recall.
• CRF model with BERT features outperformed BERTweet for cause-effect detection.
• Identified 96,676 sentences with cause-effect relationships, with "diabetes," "death," and "insulin" as central clusters.
• Found frequent associations between insulin pricing and death.

Conclusions:

• Developed a novel methodology using BERT-based architectures to detect causal sentences and extract cause-effect relationships from social media.
• Visualized findings in an interactive cause-effect network.
• Patient-reported outcomes from social media provide a valuable complementary data source for diabetes research.