Impact of Gender Bias in Training Data for Machine Learning Models predicting Myocardial Infarction

Abstract

The use of biomarkers reference ranges derived from clinical trials to detect and diagnose the presence of cardiovascular disease (CVD) and the occurrence of myocardial infarction (MI) is well established.  However, the predominance of older male participants in these trials has been shown to contribute to increased rate of misdiagnoses among women.  The application of Machine Learning (ML) to medical diagnosis promises the potential to improve accuracy.  However, ML also has the potential to perpetuate this problem of gender bias in trial data leading to worse outcomes for female patients.  This research found that models trained using data containing only male patient data were less accurate at predicting MI among female patients than those trained using data sets with greater representation of females.  For the model trained only with male patient data, this increased false negative rate corresponds to 2% of female patients with MI not being correctly diagnosed.   Addressing this issue will require (a) the collection of more female patient data to support the construction of training data sets which accurately reflect the patient population (b) consistent reporting of gender mix and other demographic information such as age when ML model performance is reported.