Potential for Algorithmic Bias in Clinical Decision Instrument Development

Area of Science:

• Health Informatics
• Medical Ethics
• Health Equity

Background:

• Clinical decision instruments (CDIs) aim to standardize care and reduce disparities.
• However, standardization may inadvertently perpetuate bias and inequality in healthcare.
• Potential sources of bias in CDI development require systematic investigation.

Purpose of the Study:

• To quantitatively characterize potential sources of bias in the development of Clinical Decision Instruments (CDIs).
• To systematically review 690 CDIs for evidence of bias in their development process.

Main Methods:

• Quantitative systematic review of 690 Clinical Decision Instruments (CDIs).
• Analysis focused on four potential sources of bias: participant demographics, investigator team geography, predictor variable selection, and outcome definitions.

Main Results:

• Evidence of potential algorithmic bias was found in CDI development.
• Participant demographics were skewed (e.g., 73% White, 55% male).
• Investigator teams showed geographic skew (52% North America, 31% Europe).
• CDIs utilized potentially biased predictor variables (e.g., 1.9% used Race and Ethnicity).
• Outcome definitions, particularly those involving follow-up (26%), may introduce socioeconomic bias.

Conclusions:

• CDIs, while intended to improve care, may contain inherent biases.
• Factors such as skewed demographics, geographic representation, and variable selection contribute to potential bias.
• Recommendations include considering these factors during CDI development and transparently communicating them to clinicians.