Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Demographic Predictability in 3D CT Foundation Embeddings

Guangyao Zheng, Michael A. Jacobs, Vishwa S. Parekh

TL;DR

This study investigates whether self-supervised 3D CT foundation embeddings leak demographic information. Using NLST data and 1408-feature embeddings, the authors train multiple classifiers to predict age, sex, and race from patient embeddings and evaluate with RMSE/MAE and AUC metrics. They find age and sex are highly encodable (age RMSE $=3.8$ years; sex AUC $=0.998$) whereas race is less encodable (AUC $=0.878$), highlighting potential privacy and fairness implications in radiology AI. The results underscore the need to audit foundation-model representations and develop strategies to mitigate unwanted demographic leakage while preserving clinical utility. This work lays groundwork for responsible deployment of self-supervised medical-imaging models by clarifying what demographic information their embeddings may encode.

Abstract

Self-supervised foundation models have recently been successfully extended to encode three-dimensional (3D) computed tomography (CT) images, with excellent performance across several downstream tasks, such as intracranial hemorrhage detection and lung cancer risk forecasting. However, as self-supervised models learn from complex data distributions, questions arise concerning whether these embeddings capture demographic information, such as age, sex, or race. Using the National Lung Screening Trial (NLST) dataset, which contains 3D CT images and demographic data, we evaluated a range of classifiers: softmax regression, linear regression, linear support vector machine, random forest, and decision tree, to predict sex, race, and age of the patients in the images. Our results indicate that the embeddings effectively encoded age and sex information, with a linear regression model achieving a root mean square error (RMSE) of 3.8 years for age prediction and a softmax regression model attaining an AUC of 0.998 for sex classification. Race prediction was less effective, with an AUC of 0.878. These findings suggest a detailed exploration into the information encoded in self-supervised learning frameworks is needed to help ensure fair, responsible, and patient privacy-protected healthcare AI.

Demographic Predictability in 3D CT Foundation Embeddings

TL;DR

This study investigates whether self-supervised 3D CT foundation embeddings leak demographic information. Using NLST data and 1408-feature embeddings, the authors train multiple classifiers to predict age, sex, and race from patient embeddings and evaluate with RMSE/MAE and AUC metrics. They find age and sex are highly encodable (age RMSE years; sex AUC ) whereas race is less encodable (AUC ), highlighting potential privacy and fairness implications in radiology AI. The results underscore the need to audit foundation-model representations and develop strategies to mitigate unwanted demographic leakage while preserving clinical utility. This work lays groundwork for responsible deployment of self-supervised medical-imaging models by clarifying what demographic information their embeddings may encode.

Abstract

Self-supervised foundation models have recently been successfully extended to encode three-dimensional (3D) computed tomography (CT) images, with excellent performance across several downstream tasks, such as intracranial hemorrhage detection and lung cancer risk forecasting. However, as self-supervised models learn from complex data distributions, questions arise concerning whether these embeddings capture demographic information, such as age, sex, or race. Using the National Lung Screening Trial (NLST) dataset, which contains 3D CT images and demographic data, we evaluated a range of classifiers: softmax regression, linear regression, linear support vector machine, random forest, and decision tree, to predict sex, race, and age of the patients in the images. Our results indicate that the embeddings effectively encoded age and sex information, with a linear regression model achieving a root mean square error (RMSE) of 3.8 years for age prediction and a softmax regression model attaining an AUC of 0.998 for sex classification. Race prediction was less effective, with an AUC of 0.878. These findings suggest a detailed exploration into the information encoded in self-supervised learning frameworks is needed to help ensure fair, responsible, and patient privacy-protected healthcare AI.

Paper Structure

This paper contains 4 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Materials and Methods
  3. Results
  4. Discussion

Figures (1)

  • Figure 1: (a) Logistic regression models predicted age vs. actual age compared. (b) ROC curve of the logistic regression model classification for sex. (c) ROC curve of the logistic regression model classification for race. (d) Visualization of the data colored by sex after T-SNE non-linear dimensionality reduction to two features. (e) Visualization of the data colored by sex after Isomap non-linear dimensionality reduction to two features.