Privacy-Aware Early Detection of COVID-19 Through Adversarial Training

Omid Rohanian, Samaneh Kouchaki, Andrew A. S. Soltan, Jenny Yang, Morteza Rohanian, David A. Clifton

In the rush to deploy machine learning models for pandemic management, privacy often takes a backseat to performance. However, a team of researchers from the University of Oxford and the University of Surrey is proving that we don’t have to choose between clinical accuracy and patient confidentiality. Their recent paper, Privacy-aware Early Detection of COVID-19 through Adversarial Training, introduces a framework that predicts COVID-19 status from routine blood tests while actively "scrubbing" sensitive demographic information from the model’s internal memory.

TL;DR

Researchers developed deep learning models that detect COVID-19 using routine clinical data (blood tests and vitals) while employing adversarial training to prevent the "leakage" of sensitive patient attributes like age, gender, and ethnicity. The result is a robust diagnostic tool that maintains high accuracy while successfully thwarting adversarial attacks designed to recover private patient data.

What Problem Does This Paper Address?

Machine learning models trained on Electronic Health Records (EHR) are surprisingly good at "reading between the lines." Even if you remove a patient's "Gender" or "Ethnicity" column from a dataset, a neural network can often reconstruct that information because clinical markers—like hemoglobin levels or white blood cell counts—frequently correlate with demographic groups.

This creates a major privacy risk: a "property inference attack." If an adversary gains access to a trained model, they might extract sensitive demographic profiles of the training population. In a healthcare setting, this isn't just a technical glitch; it’s a potential legal and ethical disaster. The challenge is building a model that is smart enough to diagnose a disease but "blind" to the sensitive traits it doesn't need to know.

Methods

The authors explored two primary adversarial architectures to bake privacy directly into the learning process:

  1. Adversarial Training with Gradient Reversal (ADV): This is the core innovation. The model consists of a main classifier (predicting COVID-19) and several "discriminators" (the adversaries). These discriminators try to guess the patient's age, gender, or ethnicity from the model's internal representations. During training, a Gradient Reversal Layer (GRL) is used. It acts like a double agent: in the forward pass, it lets data through normally, but during backpropagation, it flips the sign of the gradients. This forces the model to move away from representations that help the discriminators, effectively "un-learning" demographic associations while keeping the diagnostic features intact.
  2. Fast Gradient Sign Method (FGSM) Regularization: Originally designed to make models robust against "noise," the authors used this to perturb the input data during training. The goal here is general regularization—making the model's internal "hidden states" less specific and therefore harder for an attacker to exploit.

Datasets

The study utilized a massive amount of real-world data from the UK National Health Service (NHS):

  • OUH (Oxford University Hospitals): The primary dataset used for training and initial testing, containing over 115,000 instances.
  • External Validation Sets: To prove the model works in the "wild," it was tested on three additional trusts:
  • University Hospitals Birmingham (UHB)
    • Bedfordshire Hospitals (BH)
    • Portsmouth Hospitals University (PUH)

The researchers focused on "routine" data—blood gases, vital signs, and standard laboratory blood tests—which are typically available within an hour of a patient arriving at a hospital.

Key Insights and Ideas

  • Privacy isn't just about deleting columns: The paper highlights that clinical data is deeply "entangled." For example, certain ethnic groups have different reference ranges for neutrophil counts. A model that learns to use these counts for diagnosis might accidentally learn to identify the patient's ethnicity.
  • The "Privacy-Performance Trade-off" is surprisingly small: One might assume that forcing a model to ignore demographic data would make it less accurate. However, the authors found that the adversarial models performed nearly as well as the "unprotected" baseline, suggesting that demographic bias isn't actually necessary for accurate COVID-19 screening.
  • Adversarial training as a "De-biasing" tool: While the primary focus was privacy, this method naturally helps prevent the model from becoming biased toward specific demographic groups, a common pitfall in medical AI.

Key Results

  • Diagnostic Accuracy: The adversarial models achieved an AUC (Area Under the Curve) of ~0.85 to 0.86 for COVID-19 detection, matching the performance of standard, non-private models.
  • Privacy Protection: The "ADV" model was remarkably effective at hiding data. When an attacker tried to guess a patient's gender from the model's internal representation, their success rate (measured by AUC) plummeted from 0.91 (in a standard model) to 0.45 (essentially a random guess).
  • Robustness: The models were "cross-tested" (e.g., trained on men, tested on women) and showed no significant drop in performance, proving they had learned generalizable clinical features rather than demographic shortcuts.

Limitations and Future Work

  • Binarization: For this study, the authors simplified attributes (e.g., age was split into "over 64" and "under 64," ethnicity into "white" and "non-white"). Real-world demographics are far more nuanced, and future work will need to address multi-class or continuous variables.
  • Tabular Focus: The current work is limited to tabular EHR data. Expanding these adversarial privacy techniques to medical imaging (like X-rays or CT scans) is a logical and necessary next step.
  • Regression Tasks: The authors suggest exploring how these models handle continuous data protection, where an attacker tries to guess an exact age rather than just an age bracket.

Summary updated

Abstract

Early detection of COVID-19 is an ongoing area of research that can help with triage, monitoring and general health assessment of potential patients and may reduce operational strain on hospitals that cope with the coronavirus pandemic. Different machine learning techniques have been used in the literature to detect potential cases of coronavirus using routine clinical data (blood tests, and vital signs measurements). Data breaches and information leakage when using these models can bring reputational damage and cause legal issues for hospitals. In spite of this, protecting healthcare models against leakage of potentially sensitive information is an understudied research area. In this study, two machine learning techniques that aim to predict a patient's COVID-19 status are examined. Using adversarial training, robust deep learning architectures are explored with the aim to protect attributes related to demographic information about the patients. The two models examined in this work are intended to preserve sensitive information against adversarial attacks and information leakage. In a series of experiments using datasets from the Oxford University Hospitals (OUH), Bedfordshire Hospitals NHS Foundation Trust (BH), University Hospitals Birmingham NHS Foundation Trust (UHB), and Portsmouth Hospitals University NHS Trust (PUH), two neural networks are trained and evaluated. These networks predict PCR test results using information from basic laboratory blood tests, and vital signs collected from a patient upon arrival to the hospital. The level of privacy each one of the models can provide is assessed and the efficacy and robustness of the proposed architectures are compared with a relevant baseline. One of the main contributions in this work is the particular focus on the development of effective COVID-19 detection models with built-in mechanisms in order to selectively protect sensitive attributes against adversarial attacks. The results on hold-out test set and external validation confirmed that there was no impact on the generalisibility of the model using adversarial learning.
BibTeX
@article{rohanian2023privacy,
  author = {Omid Rohanian and Samaneh Kouchaki and Andrew A. S. Soltan and Jenny Yang and Morteza Rohanian and Yang Yang and David A. Clifton},
  title = {{Privacy-Aware Early Detection of COVID-19 Through Adversarial Training}},
  journal = {IEEE J. Biomed. Health Informatics},
  year = {2023},
  url = {https://openreview.net/forum?id=q2V4oZusGe},
}

Loading related content...

Figure