Health-Related Content in Transformer-Based Language Models: Exploring Bias in Domain General vs. Domain Specific...

Area of Science:

• Health Communication
• Natural Language Processing
• Artificial Intelligence

Background:

• Machine learning models are increasingly used in healthcare.
• Training data bias can lead to disparities in healthcare.
• General domain datasets may contain biases affecting health-related applications.

Purpose of the Study:

• To investigate if domain-specific health communication corpora reduce bias compared to general domain training sets.
• To analyze the presence and impact of bias in health-related datasets.

Main Methods:

• Comparative analysis of bias metrics.
• Evaluation of machine learning models trained on general vs. domain-specific health corpora.
• Examination of health-related text data.

Main Results:

• Bias observed in general health-related training sets persists.
• Domain-specific health communication corpora do not show improvement in mitigating this bias.
• Existing biases are replicated in specialized datasets.

Conclusions:

• Specialized health communication corpora do not inherently solve bias issues.
• Further research is needed to develop effective bias mitigation strategies for health AI.
• Addressing bias in health data requires more than just domain specialization.