A Racially Unbiased, Machine Learning Approach to Prediction of Mortality: Algorithm Development Study.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Health Equity Research

Background:

• Racial disparities in healthcare are a significant issue in the U.S.
• Machine learning (ML) in healthcare must be carefully evaluated to prevent exacerbating these disparities.
• Ensuring ML algorithms do not introduce bias is crucial for equitable patient care.

Purpose of the Study:

• To assess a novel ML algorithm designed to minimize racial bias in predicting in-hospital mortality.
• To compare the bias and accuracy of this ML algorithm against established clinical scoring systems.

Main Methods:

• Retrospective analysis of electronic health record data from ICU patients (2001-2012).
• Inclusion criteria: minimum 10 hours of measurements, all prediction variables present, and recorded race/ethnicity.
• Bias assessed using the equal opportunity difference; performance compared to MEWS, SAPS II, and APACHE.

Main Results:

• The ML algorithm demonstrated superior accuracy (sensitivity, specificity, AUC) compared to all comparators.
• The ML algorithm was found to be unbiased (equal opportunity difference = 0.016, P=.20).
• APACHE was also unbiased (0.019, P=.11), while SAPS II (0.038, P=.006) and MEWS (0.074, P<.001) showed significant bias.

Conclusions:

• Commonly used clinical scoring systems may contain significant racial bias.
• ML algorithms have the potential to reduce racial bias in healthcare predictions.
• This study suggests ML can improve both accuracy and fairness in mortality prediction.