Researchers balance privacy and accuracy in federated learning for cardiac risk diagnosis

Federated learning—a technique that trains machine learning models on data spread across multiple hospitals without moving sensitive patient information to a central server—offers real promise for collaborative medical research. When paired with differential privacy (a formal mathematical guarantee of individual privacy), it creates systems with strong security properties. Yet researchers face a tough engineering challenge: adding privacy protections often reduces the accuracy and reliability of clinical predictions.

A team led by Rodrigo Tertulino has tackled this head-on by designing a complete workflow for cardiovascular risk prediction that addresses two stubborn real-world obstacles: the strict privacy requirements of healthcare data and the severe class imbalance endemic to medical datasets (where dangerous conditions are rare compared to normal cases).

The core problem#

The researchers began with a standard federated learning setup for cardiovascular risk prediction. Initial results were discouraging: conventional methods achieved zero recall on the minority (high-risk) class—essentially missing all the patients at genuine risk. This is clinically unacceptable.

The solution: combining two techniques#

The team integrated two complementary approaches:

• SMOTETomek (Synthetic Minority Over-sampling Technique with Tomek Links), applied at the individual hospital level, generates synthetic examples of the rare condition and removes misleading borderline cases, effectively rebalancing the training data
• FedProx, a federated learning algorithm specifically designed to handle non-identical data distributions across hospitals, which outperformed the standard FedAvg approach

Mapping the privacy–utility trade-off#

The final results revealed a clear, non-linear relationship: as privacy budgets tightened (smaller epsilon values in differential privacy terms), recall declined. However, the researchers identified an operational "sweet spot": at epsilon = 9.0, the system maintained strong privacy guarantees while achieving recall above 77%—clinically acceptable performance.

Why it matters#

This work provides a practical blueprint for real-world healthcare systems, showing that it is feasible to build diagnostic tools that satisfy genuine privacy requirements without sacrificing the ability to identify patients at risk. The methodological approach is designed to work with the messy, heterogeneous data that actual hospital networks contend with every day.