Mitigating Bias in Radiology Machine Learning: 1. Data Handling.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Medicine

Background:

• Minimizing bias is essential for the clinical adoption of machine learning (ML) models.
• Systematic mathematical biases in ML systems lead to performance discrepancies and suboptimal clinical outcomes.
• Bias can originate from data handling, model development, and performance evaluation phases.

Purpose of the Study:

• To identify and explain 12 suboptimal practices in the data handling phase of ML studies.
• To illustrate how these practices introduce bias into ML models.
• To provide mitigation strategies for preventing bias during data handling.

Main Methods:

• A framework categorizing ML data handling into four steps: data collection, data investigation, data splitting, and feature engineering.
• Review of existing research literature to identify suboptimal practices and their impact.
• Development of code examples in a Google Colaboratory Jupyter notebook to demonstrate bias and prevention.

Main Results:

• Identification of 12 specific suboptimal data handling practices that contribute to ML bias.
• Explanation of the mechanisms through which these practices create bias.
• Demonstration of bias-inducing scenarios and their corresponding mitigation techniques using code.

Conclusions:

• Addressing suboptimal data handling practices is critical for reducing bias in clinical ML.
• Implementing recommended mitigation strategies can improve the reliability and performance of ML models in healthcare.
• This work provides practical guidance and tools for researchers and developers to build less biased ML systems.