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.

What problem does this paper attempt to address?

The core problem that this paper attempts to solve is: **Do self - supervised 3D CT embeddings encode demographic information (such as age, gender, and race)?** Specifically, researchers hope to explore whether these embeddings contain sensitive information such as the patient's age, gender, and race by analyzing the 3D CT image embeddings generated by the self - supervised learning model. ### Background and Motivation Self - supervised learning has made remarkable progress in the field of medical imaging and can efficiently extract general features from large - scale unlabeled datasets. Self - supervised base models have been successfully applied to encode three - dimensional (3D) CT data and have performed well in multiple downstream tasks, such as intracranial hemorrhage detection and lung cancer risk prediction. However, as these models learn from complex data distributions, a key question arises: Do these embeddings capture demographic information (such as age, gender, or race)? The presence of such information may bring advantages (such as personalized clinical decision - support systems), but it may also raise concerns about fairness and privacy. ### Research Objectives The main objective of this study is to preliminarily investigate whether self - supervised 3D CT embeddings encode demographic information, especially: - **Age** - **Gender** - **Race** Through this research, the authors hope to understand whether these embeddings contain information that can be used to predict demographic characteristics and evaluate their potential impacts. ### Methods and Results To achieve this goal, researchers used 3D CT images from the publicly available National Lung Screening Trial (NLST) dataset and their corresponding demographic information (age, gender, and race). They trained multiple machine - learning models (such as softmax regression, linear regression, support vector machines, random forests, and decision trees) to predict these demographic characteristics and evaluated the performance of the models. The results show that models based on CT Foundation embeddings can accurately predict age and gender information, but the prediction effect on race information is poor. Specifically: - **Age prediction**: The linear regression model performed best, with a root - mean - square error (RMSE) of 3.8 years. - **Gender prediction**: The softmax regression model performed best, with an area under the curve (AUC) of 0.998 and an accuracy of 0.993. - **Race prediction**: The softmax regression model had an AUC of 0.878 and an accuracy of 0.471, indicating that the model's ability to predict race was weak. ### Discussion and Conclusions The research results show that self - supervised CT embeddings do encode certain demographic characteristics (such as age and gender), which may have potential application value for personalized clinical decision - support systems. However, this also raises the risk of the model potentially spreading demographic biases and security vulnerabilities. In addition, the poor prediction effect of race information may be due to the unbalanced distribution of races in the dataset (mainly white patients). In conclusion, this study emphasizes the importance of understanding the information encoded in base - model embeddings to ensure the safety and optimized development of downstream clinical applications. Future research should continue to explore the information contained in these embeddings to promote the responsible development of medical - imaging AI, protect patient privacy, and enhance fairness.

Demographic Predictability in 3D CT Foundation Embeddings

Deep Learning Discovery of Demographic Biomarkers in Echocardiography

Algorithmic encoding of protected characteristics in image-based models for disease detection

Demographic Bias of Expert-Level Vision-Language Foundation Models in Medical Imaging

Discovering mechanisms underlying medical AI prediction of protected attributes

Drop the shortcuts: image augmentation improves fairness and decreases AI detection of race and other demographics from medical images

Can Race-sensitive Biomedical Embeddings Improve Healthcare Predictive Models?

Risk of Training Diagnostic Algorithms on Data with Demographic Bias

Demographic bias in misdiagnosis by computational pathology models

Analysing race and sex bias in brain age prediction

Deep Learning Algorithms with Demographic Information Help to Detect Tuberculosis in Chest Radiographs in Annual Workers’ Health Examination Data

The Limits of Fair Medical Imaging AI In The Wild

Learning and diSentangling patient static information from time-series Electronic hEalth Records (STEER)

Improving lung cancer health equity by applying deep learning to low dose CT screening of minority and disadvantaged patients.

Prediction of Gender from Longitudinal MRI data via Deep Learning on Adolescent Data Reveals Unique Patterns Associated with Brain Structure and Change over a Two-year Period

Self-supervised Learning for Chest CT - Training Strategies and Effect on Downstream Applications

Learning and DiSentangling Patient Static Information from Time-series Electronic HEalth Record (STEER)

Time-to-Event Pretraining for 3D Medical Imaging

Sex Estimation of Medial Aspect of the Ischiopubic Ramus in Adults Based on Deep Learning

Atlas-Based Interpretable Age Prediction In Whole-Body MR Images

Machine Learning and Real-World Data to Predict Lung Cancer Risk in Routine Care