Evaluating subgroup disparity using epistemic uncertainty in mammography

Charles Lu, Andréanne Lemay, Katharina Hoebel, Jayashree Kalpathy-CramerarXiv 2021 · OpenReview

Abstract

As machine learning (ML) continue to be integrated into healthcare systems that affect clinical decision making, new strategies will need to be incorporated in order to effectively detect and evaluate subgroup disparities to ensure accountability and generalizability in clinical workflows. In this paper, we explore how epistemic uncertainty can be used to evaluate disparity in patient demographics (race) and data acquisition (scanner) subgroups for breast density assessment on a dataset of 108,190 mammograms collected from 33 clinical sites. Our results show that even if aggregate performance is comparable, the choice of uncertainty quantification metric can significantly impact the subgroup level. We hope this analysis can promote further work on how uncertainty can be leveraged to increase transparency of machine learning applications for clinical deployment.

Introduction

Machine learning (ML) models in healthcare often suffer from "hidden" biases. A model might boast high overall accuracy yet perform poorly on specific demographics, such as minority racial groups, or data from specific hardware, such as older mammography scanners. This disparity is a major hurdle for clinical deployment, where fairness and generalizability are safety requirements.

This study addresses the challenge of subgroup disparity in breast density assessment—a critical task because high breast density is both a risk factor for cancer and a factor that makes tumors harder to detect on a mammogram. The goal is to move beyond simple accuracy metrics and use the model’s internal epistemic uncertainty (the model's own "doubt") to detect and mitigate these disparities.

Methodology and Data

Uncertainty Quantification

The researchers employed Monte Carlo (MC) Dropout to estimate epistemic uncertainty. By running the same image through the model multiple times with different neurons randomly deactivated, the model's consistency can be measured. High fluctuation in predictions indicates high uncertainty. Four metrics were compared:

  1. Maximum Softmax Probability (Naive): The confidence of the single highest prediction.
  2. Predictive Variance: The fluctuation of predictions across multiple passes.
  3. Predictive Entropy: The "disorder" or spread of the prediction distribution.
  4. Bhattacharyya Coefficient (BC): The similarity between the distributions of the top two predicted classes.

Dataset and Metrics

The study utilized the Digital Mammographic Imaging Screening Trial (DMIST) dataset, comprising 108,230 mammograms from 21,729 patients across 33 clinical sites. The analysis focused on five racial categories (White, Black, Hispanic, Asian, and Other) and four scanner types (Senograph, Senscan, ADS, and Other).

The authors proposed a new subgroup disparity metric ($\delta$), which calculates the average difference in performance between all unique pairs of subgroups across various "rejection thresholds." A rejection threshold allows a clinical system to triage highly uncertain cases to a human expert.

Results: The Impact of Uncertainty Metrics on Disparity

Aggregate vs. Subgroup Performance

The model achieved an overall Kappa score of 0.646 ± 0.016. However, aggregate metrics proved deceptive. Performance was consistently higher for the most represented groups (e.g., White patients and Senograph scanners) and lower for minority groups and less common hardware.

The "Triage" Trap

A common assumption is that rejecting the most uncertain cases will improve performance across the board. However, the study found that for some metrics, this actually increases disparity, making the model less fair for minority groups as more data is excluded.

Subgroup disparity across racial categories. Predictive variance (purple) maintains the lowest disparity as uncertain cases are excluded, whereas the Naive metric (blue) shows increasing disparity.

Predictive Variance as a Robust Metric

Predictive variance emerged as the most effective metric for reducing bias. At a 10% rejection threshold, the race disparity ($\delta$) for the Naive metric was 0.051, while Predictive Variance was significantly lower at 0.034.

Furthermore, when excluding uncertain cases based on predictive variance, the disparity between scanner types actually decreased at higher rejection thresholds (25%), whereas other metrics saw disparity stay flat or increase. This suggests that predictive variance is particularly effective at identifying cases where the model struggles due to unfamiliar hardware.

Subgroup disparity across scanner types. Unlike other metrics, predictive variance (purple) uniquely reduces disparity at higher rejection thresholds, improving fairness across hardware types.

Discussion and Limitations

There was a strong negative correlation between class frequency and uncertainty; the model was most confident on the most common breast density classes (scattered and heterogeneous). This highlights the impact of data scarcity on model reliability.

While promising, this study is limited to breast density assessment. Future work should evaluate if these trends hold for cancer detection. Additionally, the practical integration of uncertainty scores into clinical workflows—specifically how radiologists interact with these scores—remains to be tested.

Conclusion

To build truly fair medical AI, developers must look beyond what a model says and evaluate how much the model "knows" it doesn't know. By leveraging predictive variance as an uncertainty metric, clinical systems can more effectively triage ambiguous cases without inadvertently widening the performance gap between patient subgroups.

Summary updated

Abstract

As machine learning (ML) continue to be integrated into healthcare systems that affect clinical decision making, new strategies will need to be incorporated in order to effectively detect and evaluate subgroup disparities to ensure accountability and generalizability in clinical workflows. In this paper, we explore how epistemic uncertainty can be used to evaluate disparity in patient demographics (race) and data acquisition (scanner) subgroups for breast density assessment on a dataset of 108,190 mammograms collected from 33 clinical sites. Our results show that even if aggregate performance is comparable, the choice of uncertainty quantification metric can significantly the subgroup level. We hope this analysis can promote further work on how uncertainty can be leveraged to increase transparency of machine learning applications for clinical deployment.
BibTeX
@misc{lu2021evaluating,
  author = {Charles Lu and Andréanne Lemay and Katharina Hoebel and Jayashree Kalpathy-Cramer},
  title = {{Evaluating subgroup disparity using epistemic uncertainty in mammography}},
  year = {2021},
  eprint = {2107.02716},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2107.02716},
}
Figure