UnBias: Unveiling Bias Implications in Deep Learning Models for Healthcare Applications.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Deep learning AI is increasingly used in critical sectors like healthcare, leading to concerns about fairness across demographic groups.
• Existing deep learning models may exhibit biases that negatively impact decision-making processes.
• Ethical implications of AI bias require rigorous investigation to ensure equitable outcomes.

Purpose of the Study:

• To introduce the UnBias approach for assessing and detecting bias in deep neural network architectures.
• To investigate how bias influences the learning process and feature focus in AI models.
• To advance the development of fair and trustworthy AI applications, with a focus on healthcare.

Main Methods:

• Developed and applied the UnBias methodology to evaluate bias in various deep neural network architectures.
• Utilized chest X-ray datasets from public repositories for a case study on COVID-19 detection.
• Tested five gender-based models across four deep learning architectures: ResNet50V2, DenseNet121, InceptionV3, and Xception.

Main Results:

• The UnBias approach successfully detected instances of bias within the deep learning models.
• Bias was observed to shift model focus away from salient features, impacting performance.
• Variations in bias were noted across different deep learning architectures and gender-based models.

Conclusions:

• The UnBias approach provides a framework for identifying and addressing bias in deep learning models.
• Mitigating bias is crucial for deploying equitable and reliable AI systems in sensitive applications like healthcare.
• Further research is needed to refine bias detection and mitigation strategies for diverse AI applications.