Interdisciplinary expertise to advance human-centered explainable AI.

Area of Science:

• Health Informatics
• Artificial Intelligence
• Bias Detection

Background:

• Artificial intelligence (AI) significantly impacts healthcare, yet persistent bias and subgroup performance issues hinder its full potential.
• Existing research emphasizes data representativeness and model performance for reducing bias and advancing population health.
• Social epidemiology, behavioral science, and health services research offer valuable frameworks for interpreting AI findings, especially with explainable AI (XAI).

Purpose of the Study:

• To introduce a novel human-in-the-loop model integrating generative XAI with an interdisciplinary expert panel.
• To critically assess AI model explanations from multiple perspectives to identify bias and guide future research.
• To leverage multidisciplinary expertise for improving AI rigor and fostering team science.

Main Methods:

• Coupling generative explainable AI (XAI) with a wraparound interdisciplinary expert panel review process.
• Human-centered approach for critically assessing AI model explanations.
• Incorporating historical and contextual information through expert interpretation.

Main Results:

• The interdisciplinary expert panel effectively identified areas of bias in AI models.
• Expert interpretations provided historically and contextually informed insights.
• Potential confounders were identified, and scientific evidence from previous studies was integrated to reduce bias.

Conclusions:

• The human-centered framework enhances methodological rigor for AI model improvement and implementation in healthcare.
• This approach proactively fosters team science and collaborative research opportunities.
• The model offers a promising pathway to mitigate bias and improve AI performance in health applications.