(Predictable) performance bias in unsupervised anomaly detection.

Area of Science:

• Medical Imaging AI
• Algorithmic Fairness in Healthcare
• Machine Learning for Disease Detection

Background:

• The increasing volume of medical imaging data necessitates AI tools for clinicians.
• Unsupervised anomaly detection (UAD) models are crucial for early disease identification.
• Fairness in UAD models remains an underexplored area, unlike supervised models.

Purpose of the Study:

• To investigate how dataset composition affects UAD model performance across demographic subgroups.
• To quantify fairness disparities in UAD models using chest X-ray datasets.
• To introduce a metric for measuring fairness in machine learning models.

Main Methods:

• Evaluated UAD model performance on three large-scale public chest X-ray datasets.
• Assessed performance variations across protected variables based on dataset subgroup representation.
• Utilized two state-of-the-art UAD models and introduced subgroup-AUROC (sAUROC) for fairness quantification.

Main Results:

• Discovered empirical 'fairness laws' demonstrating linear relationships between subgroup representation and anomaly detection performance.
• Observed performance disparities even with balanced training data.
• Identified compound effects that worsen performance for individuals in multiple underrepresented groups.

Conclusions:

• Quantified disparate performance of UAD models across demographic subgroups.
• Demonstrated that balanced representation alone does not mitigate unfairness; some subgroups are harder for models to learn.
• The discovered 'fairness laws' enable estimation of disparate performance and guide optimal dataset composition for equitable AI.