Preventing Sensitive Information Leakage via Post-Hoc Orthogonalization With Application to Chest Radiograph Embeddings

MCML Authors

Tobias Weber

* Former Member

→ Group Bernd Bischl
 Statistical Learning and Data Science

Michael Ingrisch

Prof. Dr.

Principal Investigator

Clinical Data Science in Radiology

Bernd Bischl

Prof. Dr.

Director

Statistical Learning and Data Science

David Rügamer

Prof. Dr.

Principal Investigator

Statistics, Data Science and Machine Learning

Abstract

Deep learning has substantially advanced data analysis across various fields. However, research indicates that protected characteristics, such as age, sex, and race, are often implicitly encoded within the deep feature representations, or embeddings, generated by neural networks. This encoding can lead to inherent biases, which in turn may influence decision-making processes. In clinical settings, in particular, such biases risk leading to unfair treatment of certain subgroups, potentially resulting in serious consequences. After analyzing the sources of these biases in the field of radiology, we illustrate how embeddings of chest radiographs (CXRs) can be corrected to remove the influence of protected features. To showcase the harms of such incidents, we study the MIMIC and CheXpert datasets with three prominent pre-trained models: a supervised contrastive, a self-supervised contrastive, and a baseline classifier model. Our experiments reveal a significant influence of protected features on predictions of pathologies in CXRs, demonstrating the potential harm of such practices. We then propose a correction method, removing these harmful effects while maintaining competitive predictive performance.

inproceedings

---

PAKDD 2025

29th Pacific-Asia Conference on Knowledge Discovery and Data Mining. Sydney, Australia, Jun 10-13, 2025.

---

Authors

T. Weber • M. Ingrisch • B. Bischl • D. Rügamer

---

Links

DOI GitHub

---

Research Areas

 A1 | Statistical Foundations & Explainability

 C1 | Medicine