🤖 machine learning

Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction

This study demonstrates that a federated deep learning approach successfully improves cardiovascular disease risk prediction across two heterogeneous cohorts (Lifelines and the Rotterdam Study) while preserving patient privacy by enabling collaborative model training without sharing sensitive individual-level data.

Original authors: Hyunho Mo, Djura Smits, Mahlet A. Birhanu, Maarten J. G. Leening, Daniel Bos, Pim van der Harst, Esther E. Bron

Published 2026-07-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Hyunho Mo, Djura Smits, Mahlet A. Birhanu, Maarten J. G. Leening, Daniel Bos, Pim van der Harst, Esther E. Bron

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to teach a super-smart robot how to predict who might get a heart problem in the next ten years. Usually, to teach this robot, you'd need to gather every single patient's medical file into one giant, central library. But here's the catch: those files are like secret diaries. Laws (like the GDPR) say you can't just mail those diaries to a central library because that would break people's privacy.

So, how do you teach the robot without seeing the diaries? The authors of this paper tried a clever trick called Federated Learning.

Think of it like a group of students working on a group project, but they aren't allowed to leave their own houses.

Student A lives in a huge neighborhood called Lifelines (with 148,230 people). They have a lot of data, but it's a bit fuzzy because the students just filled out questionnaires about their health. They know if someone had a heart event, but they aren't 100% sure of the exact date.
Student B lives in a smaller, very precise neighborhood called the Rotterdam Study (with 10,155 people). Their data is super sharp because it's linked directly to hospital records. They know the exact dates of heart events.

Instead of sending their diaries to a teacher, each student trains their own mini-robot on their local data. Then, they only send the lessons learned (the math weights of the robot) to a central server. The server mixes these lessons together to make one "Super Robot" and sends the new instructions back. The actual patient data never leaves their homes.

What did they find?
The paper suggests that this "group project" approach actually worked better than if each student had worked alone.

When the Rotterdam Study robot trained by itself, it got a score of 0.728 (out of 1.0, where higher is better).
After joining the group project, its score went up to 0.739.
The Lifelines robot also improved, going from 0.783 to 0.787.

It's like the big, fuzzy neighborhood (Lifelines) shared enough general wisdom to help the small, precise neighborhood (Rotterdam) get even sharper, even though they never saw each other's secret diaries.

What didn't work or wasn't proven?
The paper is careful to say this isn't a magic cure-all.

They did not prove that this works for every hospital or every type of disease. They only tested it on these two specific groups.
They did not say the problem is "solved." In fact, they noticed something interesting: the big, fuzzy neighborhood (Lifelines) was so huge that its lessons sometimes started to pull the small, precise robot's performance down a tiny bit after a while (around the 12th round of lessons). The paper suggests that in the future, we might need a smarter way to mix the lessons so the big group doesn't drown out the small, precise group.
They did not use the robot to actually treat patients yet. This was a test to see if the math works.

How sure are they?
The authors are confident that the method suggests an improvement. They ran the experiment 10 times to be sure the results weren't just luck, and the scores consistently went up. However, because they only tested two specific groups with different sizes and different ways of recording data, they can't claim this is a perfect solution for the whole world just yet. They argue that while the "fuzzy" data helped, the "precise" data was still the star of the show for the final evaluation.

In short, the paper shows that you can build a better heart-risk predictor by letting different groups teach a robot together without ever sharing their private secrets, but we still need to figure out the best way to balance the lessons when one group is much bigger than the other.

Technical Summary: Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction

Problem Statement
Cardiovascular disease (CVD) risk prediction models traditionally rely on data pooled from single institutions or centrally aggregated datasets. However, extending these models across multiple institutions is hindered by privacy regulations (e.g., GDPR) that restrict the sharing of individual-level patient data. While Federated Learning (FL) offers a solution by enabling collaborative model training without data transfer, its application in healthcare faces challenges due to heterogeneity among datasets. These differences include variations in cohort size, population characteristics, outcome definitions (e.g., self-reported vs. clinically verified), and follow-up completeness. Previous studies have focused on validating FL infrastructure using pseudo-splits of single cohorts, leaving a gap in understanding the predictive benefits of true multi-site FL across genuinely heterogeneous populations.

Methodology
The study implements a horizontal federated learning framework to train a Deep Survival Neural Network (DeepSurv) for CVD risk prediction across two distinct population-based cohorts: Lifelines (N=148,230) and the Rotterdam Study (N=10,155).

Data Harmonization: To ensure consistent inputs across sites, a three-stage pipeline converted raw cohort data into an HL7 FHIR-compliant format. This involved transforming CSV data into patient-centered Common Data Format (CDF), applying site-specific mapping rules to generate FHIR resources, and flattening these into tabular structures for the learning algorithm. No raw data were transferred between sites.
Predictors and Outcomes: The model utilized 12 predictors (e.g., age, sex, blood pressure, lipids, smoking status, diabetes, hypertension) selected based on established CVD risk tools. The outcome was a composite CVD endpoint (stroke, myocardial infarction, or heart failure) over a 10-year follow-up. Notably, Lifelines outcomes were self-reported with interval-based event timing, while Rotterdam Study outcomes were digitally linked to clinical records with exact event dates.
Federated Setup: The system utilized the Vantage6 platform with a client-server architecture. Two client nodes (Lifelines and Rotterdam Study) held local, FHIR-harmonized datasets. A central aggregator node performed Federated Averaging (FedAvg).
- Model Architecture: DeepSurv, a feed-forward neural network with an input layer of 12 neurons, two hidden layers of 16 neurons, and a linear output. It replaced the linear predictor of the Cox proportional hazards model with a neural network to capture non-linear associations.
- Training Protocol: Local models were trained for 100 epochs per round using the Adam optimizer. The process ran for 20 aggregation iterations. Model weights were aggregated based on the number of training samples at each node ( $n_k$ ).
- Evaluation: Performance was assessed using the C-statistic (concordance statistic) across 10 independent runs with stratified resampling. Confidence intervals were computed using a corrected resampled t-test.

Key Results
The study compared the performance of local models (trained independently on each cohort without federation) against the federated model (trained via FedAvg).

Rotterdam Study: The C-statistic improved from 0.728 (95% CI: 0.717–0.739) in the local model to 0.739 (95% CI: 0.728–0.749) in the federated model. The learning trajectory showed a rapid initial increase, peaking around iteration 12, followed by a slight decline but remaining above the local baseline.
Lifelines: The C-statistic improved from 0.783 (95% CI: 0.775–0.791) to 0.787 (95% CI: 0.780–0.792). The improvement was more gradual and sustained across all 20 iterations without the decline observed in the Rotterdam Study.
Heterogeneity Handling: Despite the significant differences in cohort size (Lifelines being ~14x larger), outcome precision (self-reported vs. clinical), and population demographics, the federated approach successfully improved discrimination in both cohorts. The larger improvement in the smaller Rotterdam Study suggests it benefited significantly from the information provided by the larger Lifelines cohort.

Significance and Claims
The paper claims that federated deep learning can effectively improve CVD risk prediction across heterogeneous cohorts while strictly preserving individual-level data privacy. Key contributions include:

Real-World Validation: Moving beyond infrastructure proofs-of-concept using pseudo-splits to demonstrate the feasibility of FL across two independent, real-world cohorts with distinct characteristics and outcome definitions.
Value of Imperfect Data: The study suggests that large cohorts with less precise, self-reported outcomes (Lifelines) can still contribute valuable information to improve risk prediction in smaller cohorts with high-precision clinical data (Rotterdam Study) when used in a federated setting.
Privacy Compliance: The framework successfully operated within existing institutional privacy agreements (GDPR compliance), sharing only aggregated model weights and ensuring no individual-level data left the local environments.
Convergence Dynamics: The authors note that convergence behavior in real-world heterogeneous settings differs from pseudo-split scenarios, with the larger cohort exerting a dominant influence on the aggregated model, potentially affecting the performance trajectory of smaller cohorts in later iterations.

The authors conclude that while limitations exist (such as the two-node setup and the handling of competing risks), the findings support the potential for large-scale, multi-institutional collaborative CVD risk prediction without compromising data privacy.

Technical Summary: Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction

More like this