Fairness in multimodal machine learning applications in clinical decision support: a systematic review

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Health Equity

Background:

• Multimodal AI models are increasingly integrated into clinical decision support (CDS) systems.
• Assessing fairness in these AI systems is crucial for equitable healthcare delivery.
• However, the extent to which fairness is evaluated in multimodal AI-based CDS is not well understood.

Purpose of the Study:

• To systematically review studies evaluating fairness in multimodal AI-based CDS systems.
• To determine the prevalence of fairness evaluations in multimodal CDS research, specifically for chest x-rays and sepsis.

Main Methods:

• A systematic literature review was conducted using two targeted searches.
• Studies were screened for fairness evaluations in multimodal AI-based CDS, excluding rule-based or knowledge-based systems.
• A specific focus was placed on studies related to chest x-rays and sepsis.

Main Results:

• Out of 3059 search results, 160 articles evaluated fairness, with only 11% (18/160) using multimodal data and employing 29 different fairness metrics.
• For chest x-ray and sepsis CDS, 8% (7/88) of studies using multimodal data included fairness evaluations.
• Despite the availability of sensitive attribute data in 83% (74/88) of these studies, fairness was infrequently assessed.

Conclusions:

• Fairness evaluations are notably uncommon in research on multimodal AI-based CDS systems.
• There is a significant need for improved guidance and tools to support fairness assessments in this domain.
• Enhancing fairness evaluations is essential for the responsible development and deployment of multimodal AI in clinical settings.