Detecting dataset bias in medical AI using a generalized and modality agnostic auditing approach

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Machine Learning

Background:

• Artificial Intelligence (AI) shows promise in healthcare but faces challenges with bias amplification from non-representative datasets.
• Latent bias in AI models can be introduced during training and remain undetected during testing, posing risks in clinical deployment.

Purpose of the Study:

• To develop and present a novel method for detecting AI bias in medical datasets.
• To introduce Generalized Attribute Utility and Detectability-Induced bias Testing (G-AUDIT) as a tool to quantify shortcut learning risks.

Main Methods:

• G-AUDIT is a data modality-agnostic approach for auditing datasets.
• It examines the relationship between task-level annotations, sensor-level measurements, and patient, environmental, and acquisition characteristics.
• The method automatically quantifies risks associated with shortcut learning in AI models.

Main Results:

• The applicability of G-AUDIT was demonstrated across diverse medical datasets, including images, text, and tabular data.
• The method successfully identified potential shortcuts in machine learning tasks that traditional qualitative assessments often miss.
• G-AUDIT provides a quantitative measure of bias, enhancing the reliability of AI in healthcare.

Conclusions:

• G-AUDIT offers a robust solution for identifying and mitigating AI bias in medical applications.
• This automated auditing approach can improve the safety and trustworthiness of AI tools deployed in healthcare.
• The method's modality-agnostic nature makes it broadly applicable for ensuring AI fairness and performance.