Gender-based data bias and model fairness evaluation in benchmarked open-access disease prediction datasets

Area of Science:

• Artificial Intelligence
• Machine Learning
• Data Science
• Healthcare Informatics

Background:

• Open-access datasets are widely used for machine learning (ML) model validation.
• Concerns exist regarding data bias and model fairness, especially gender bias.
• Investigating gender bias in disease prediction datasets is critical for equitable AI.

Purpose of the Study:

• To systematically investigate gender-based data bias in disease prediction datasets.
• To evaluate the fairness of various ML algorithms trained on these datasets.
• To identify datasets and algorithms that exhibit or mitigate gender bias.

Main Methods:

• Selected 74 datasets from Kaggle and UCI ML Repository with gender and classification labels.
• Quantified data bias using Earth Mover's Distance and assessed statistical significance via bootstrapping.
• Evaluated fairness of 7 ML algorithms using k-fold cross-validation and Equalized Odds/Treatment Equality definitions.

Main Results:

• 35 out of 74 datasets showed significant gender-based data bias, predominantly affecting females.
• Heart disease datasets had the highest bias prevalence; lung cancer and mental health datasets were bias-free.
• Decision Tree algorithms exhibited fewer fairness issues compared to Logistic Regression.

Conclusions:

• Gender-based data bias is prevalent in disease prediction datasets, impacting model fairness.
• Bias-free datasets and specific algorithms (e.g., Decision Tree) contribute to more equitable AI.
• Addressing data bias and algorithm selection is essential for fair and reliable ML applications in healthcare.