Uncertainty-Aware Health Diagnostics via Class-Balanced Evidential Deep Learning

Area of Science:

• Artificial Intelligence
• Medical Diagnostics
• Machine Learning

Background:

• Uncertainty quantification is vital for safe deep learning in healthcare.
• Class imbalance in medical data poses a significant challenge to existing methods.
• Current approaches often fail to provide reliable uncertainty estimates for imbalanced datasets.

Purpose of the Study:

• To propose a novel class-balanced evidential deep learning framework.
• To enhance the fairness and reliability of uncertainty estimates in health diagnostic models.
• To address the limitations of existing methods when dealing with class-imbalanced medical data.

Main Methods:

• Developed a class-balanced evidential deep learning framework.
• Introduced a pooling loss to mitigate class bias in evidence learning.
• Incorporated a learnable prior to regularize posterior distributions and improve uncertainty quality.

Main Results:

• Demonstrated the effectiveness of the proposed framework on benchmark and real-world imbalanced health data.
• Showcased superior performance compared to existing uncertainty quantification methods.
• Validated the ability to provide fair and reliable uncertainty estimates even with significant class imbalance.

Conclusions:

• The proposed framework significantly improves uncertainty quantification for class-imbalanced health data.
• This advancement contributes to the development of more trustworthy and practical deep learning diagnostic systems.
• The research bridges the gap between theoretical uncertainty quantification and real-world healthcare applications.