Hidden Stratification Causes Clinically Meaningful Failures in Machine Learning for Medical Imaging

TL;DR

Hidden stratification arises when training labels fail to capture meaningful subgroups in medical imaging, risking poor performance on clinically important but underrepresented cases. The authors outline three measurement approaches—schema completion, error auditing, and algorithmic detection via unsupervised clustering—to uncover and quantify these strata. Across hip fracture, MURA, and CXR14 datasets, they demonstrate substantial subgroup performance gaps (up to ~20 percentage points) and show that spurious correlates can exacerbate risk. They advocate incorporating hidden stratification assessment into ML deployment and regulation for safer, more reliable medical imaging systems.

Abstract

Machine learning models for medical image analysis often suffer from poor performance on important subsets of a population that are not identified during training or testing. For example, overall performance of a cancer detection model may be high, but the model still consistently misses a rare but aggressive cancer subtype. We refer to this problem as hidden stratification, and observe that it results from incompletely describing the meaningful variation in a dataset. While hidden stratification can substantially reduce the clinical efficacy of machine learning models, its effects remain difficult to measure. In this work, we assess the utility of several possible techniques for measuring and describing hidden stratification effects, and characterize these effects on multiple medical imaging datasets. We find evidence that hidden stratification can occur in unidentified imaging subsets with low prevalence, low label quality, subtle distinguishing features, or spurious correlates, and that it can result in relative performance differences of over 20% on clinically important subsets. Finally, we explore the clinical implications of our findings, and suggest that evaluation of hidden stratification should be a critical component of any machine learning deployment in medical imaging.

Figures (1)

• Figure 1: ROC curves for subclasses of the (a) abnormal MURA superclass and (b) pneumothorax CXR14 superclass. All subclass AUCs are significantly different than the overall task (p $<$ 0.05).