FRAMM: Fair ranking with missing modalities for clinical trial site selection

Area of Science:

• Biomedical Informatics
• Clinical Trial Management
• Artificial Intelligence in Healthcare

Background:

• Underrepresentation of diverse populations in clinical trials impacts treatment efficacy and subgroup analysis.
• Existing clinical trial site selection methods struggle with incomplete data and balancing enrollment with diversity.

Purpose of the Study:

• To introduce FRAMM, a deep reinforcement learning framework for fair clinical trial site selection.
• To address challenges of incomplete data modalities and simultaneous optimization of enrollment and diversity.

Main Methods:

• FRAMM utilizes a modality encoder with masked cross-attention to handle missing data.
• A deep reinforcement learning approach with a custom reward function optimizes for both enrollment and fairness.
• Evaluation performed using real-world historical clinical trial data.

Main Results:

• FRAMM outperforms leading baselines in enrollment-only scenarios.
• The framework significantly improves participant diversity in clinical trials.
• Demonstrates effective trade-offs between enrollment numbers and demographic representation.

Conclusions:

• FRAMM offers a novel solution for fair and efficient clinical trial site selection.
• The framework effectively addresses data limitations and optimizes for multiple objectives.
• FRAMM has the potential to improve the generalizability and precision of clinical trial findings across diverse populations.