In medicine, how do we machine learn anything real?

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Health Equity

Background:

• Traditional machine learning assumes objective data, but health data often reflects historical discrimination.
• Embodied data from human bodies can lead to AI systems that do not function as intended.
• Naive application of machine learning in healthcare is problematic due to data complexity and historical biases.

Purpose of the Study:

• To demonstrate the pervasive nature and historical depth of biases within medical data.
• To advocate for machine learning models that actively recognize, reduce, or remove biases.
• To encourage systemic changes in data generation practices to mitigate health disparities.

Main Methods:

• Enumeration of diverse examples illustrating biases in medical data.
• Analysis of how these biases manifest in machine learning applications.
• Discussion of the link between biased data and discriminatory healthcare outcomes.

Main Results:

• Identification of widespread and historically embedded biases in healthcare data.
• Demonstration of how these biases are automated and amplified by machine learning algorithms.
• Evidence that current machine learning practices can exacerbate existing health disparities.

Conclusions:

• Machine learning models must be designed to counteract, not perpetuate, societal biases.
• Addressing bias at the data generation stage is essential for improving healthcare.
• Mitigating algorithmic bias is a critical step toward achieving health equity.