Sex-Based Performance Disparities in Machine Learning Algorithms for Cardiac Disease Prediction: Exploratory Study.

Area of Science:

• Artificial Intelligence in Healthcare
• Cardiology Machine Learning
• Algorithmic Bias and Health Equity

Background:

• Artificial intelligence (AI) bias is increasingly recognized across sectors, with potential to exacerbate health disparities in healthcare.
• Algorithmic inequities in healthcare can widen existing health inequalities among demographic groups.

Purpose of the Study:

• To identify and characterize bias in machine learning (ML) algorithms used for heart failure management.
• To specifically investigate sex-based disparities in the performance of cardiology algorithms.

Main Methods:

• Literature search of PubMed and Web of Science for cardiac ML algorithms focusing on demographic bias.
• Reproduced existing ML algorithms on two open-source datasets (UCI Heart Failure and UCI CAD).
• Assessed algorithms for sex bias, focusing on false negative rates (FNR) and evaluating remediation techniques.

Main Results:

• Only 3 out of 127 reviewed papers addressed sex differences; female patients were underrepresented in datasets.
• Reproducing algorithms showed mean accuracies around 85%, but significant sex disparities were found.
• Female patients experienced higher FNR (underdiagnosis), while male patients had higher false positive rates (overdiagnosis).

Conclusions:

• Cardiac ML research has overlooked significant underperformance of algorithms for female patients.
• Sex disparities in algorithmic error rates and feature importance were quantified.
• Current remediation techniques were insufficient to eliminate identified inequities in cardiology algorithms.