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

Area of Science:

• Medical Imaging
• Machine Learning
• Artificial Intelligence

Background:

• Machine learning models in medical imaging often exhibit performance disparities across different population subsets.
• This issue, termed hidden stratification, arises from incomplete data variation description and can lead to models missing critical subgroups, such as rare cancer subtypes.

Purpose of the Study:

• To assess techniques for measuring hidden stratification effects in machine learning models.
• To characterize hidden stratification in synthetic and real-world medical imaging datasets.

Main Methods:

• Utilized synthetic experiments on the CIFAR-100 dataset.
• Analyzed multiple real-world medical imaging datasets.
• Employed several measurement techniques to quantify hidden stratification.

Main Results:

• Hidden stratification was identified in subsets with low prevalence, poor label quality, subtle features, or spurious correlations.
• Performance differences exceeding 20% were observed on clinically significant subsets.
• Evidence suggests hidden stratification can significantly reduce model efficacy on important subgroups.

Conclusions:

• Hidden stratification poses a significant challenge to the clinical utility of medical imaging AI.
• Evaluation of hidden stratification should be integral to deploying machine learning in medical imaging.
• Developing methods to detect and mitigate hidden stratification is essential for equitable AI performance.