Giovanni Montana

PhD

University of Warwick

Professor

PLCO (Learn more about this study)

PLCOI-1042

Sep 6, 2022

Ethnicity bias in automated chest X-ray classification

We are investigating ethnicity as a source of model bias in automated processing of chest X-rays using deep learning. It has recently been shown to be possible to accurately predict patient ethnicity from a chest X-ray, implying that there exist systematic anatomical feature variations between ethnicities in this modality. This could lead in turn to systematic model errors in pathology prediction for minority classes in a training set if the ethnicity-linked image features intersect with pathology-linked features identified by the model for the majority class. We intend to train an interpretable ethnicity classifier giving insight into anatomical sources of potential systematic errors in classifier systems. We further intend to evaluate systematic ethnicity bias in the results of an existing X-ray abnormality classifier and attempt to correct for this in future work, potentially resolving an ongoing concern in clinical adoption.

Evaluate the accuracy of ethnicity detection from chest X-rays using machine learning using a variety of approaches.

Quantify the presence and degree of bias in a pre-trained multi-label abnormality classifier attributable to patient ethnicity.

Identify the mechanism by which ethnicity-linked features leads to bias in machine learning models for chest X-rays.