Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction

summary

Video file (mp4)

The gist

The gist The study presents a federated deep learning approach for privacy-preserving cardiovascular disease risk prediction that integrates two population-based cohorts with different

In short

This study used federated deep learning to predict cardiovascular disease risk by training a model across two distinct patient cohorts, Lifelines and the Rotterdam Study. By collaborating without sharing sensitive raw data, the federated approach achieved higher predictive accuracy than local training alone. The results show that combining data from both cohorts improves risk prediction, suggesting larger datasets can compensate for uncertainty in smaller ones.

Key concepts

Federated Learning
A method where a machine learning model is trained across multiple decentralized locations (nodes) holding separate data. Instead of moving patient records to one central server, each location trains a local copy of the model and only sends back the learned updates (model weights), ensuring raw patient data remains private at its source.
DeepSurv Model
This is the specific predictive algorithm used, which combines a deep neural network with the Cox proportional hazards model. It helps estimate how long it takes for a cardiovascular event to occur based on various patient characteristics. The network structure involves input layers and hidden layers designed to process complex risk factors.
Data Harmonization (MyDigiTwin)
This is a pipeline used to standardize data from different sources into a single, common format called HL7 FHIR. It converts raw CSV data into patient-centered files, applies specific mapping rules for each cohort, and flattens the results so they can be used uniformly by the federated learning algorithm.
Federated Averaging (FedAvg)
This is the central aggregation technique in federated learning. After local training on each site finishes, the central server calculates a new global model update by taking a weighted average of all the locally updated model weights. This process allows the entire network to benefit from the knowledge gained across all participating cohorts.

Terminology used across episodes

This episode discusses

The paper

Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction · Read on arXiv

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

Department of Radiology & Nuclear Medicine, Erasmus MC University Medical Center Rotterdam Department of Epidemiology, Erasmus MC University Medical Center Rotterdam Department of Cardiology, Erasmus MC University Medical Center Rotterdam Department of Cardiology, University Medical Center Utrecht

Cardiovascular disease risk prediction models often rely on data from a single institution or centrally pooled datasets. Extending these models across institutions could be limited by privacy regulations and constraints on sharing patient-level data. Federated learning enables collaborative model development without transferring sensitive patient data, but its application in healthcare remains challenging because datasets often differ in size, population characteristics, and outcome definitions. In this study, we present a federated deep learning approach for privacy-preserving cardiovascular disease risk prediction that integrates two population-based cohorts with different characteristics: Lifelines, including 148,230 participants meeting the study inclusion criteria with self-reported outcomes, and the Rotterdam Study, including a smaller cohort of 10,155 participants with digitally linked clinical outcomes. Model performance was primarily evaluated on the Rotterdam Study because of its complete follow-up. Deep survival models trained using federated learning achieved higher predictive performance than models trained locally without federation. For the Rotterdam Study, the C-statistic increased from 0.728 (95% CI: 0.717-0.739) to 0.739 (95% CI: 0.728-0.749). For Lifelines, the C-statistic increased from 0.783 (95% CI: 0.775-0.791) to 0.787 (95% CI: 0.780-0.792). These findings suggest that federated deep learning across heterogeneous cohorts can improve cardiovascular disease risk prediction while preserving the privacy of individual-level patient data.

DOI: 10.3389/fams.2026.1923866

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: We're looking at the paper: "Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction".

Jane: Here's the gist of the paper. The gist The study presents a federated deep learning approach for privacy-preserving cardiovascular disease risk prediction that integrates two population-based cohorts with different characteristics, Lifelines and the Rotterdam Study, achieving higher predictive performance than models trained locally without federation.

Tom: Federated Learning Framework Federated learning enables collaborative model development without transferring sensitive patient data by distributing the training process across several nodes, each holding a separate part of the data Rather than sharing raw records, each node trains a model on its local data and sends only the resulting model parameter updates, such as learned network weights, to a central server Since raw data never leave their source, federated learning has gained attention as a practical means to build predictive models across multiple institutions while protecting data privacy The federated setup follows a client-server architecture implemented on the Vantage6 platform Two data nodes (client 1: Lifelines; client 2: Rotterdam Study) each hold a locally stored, FHIR-harmonized dataset and perform local model training independently Model aggregation followed Federated Averaging (FedAvg) in which the aggregated weight wt+1 is computed as the weighted average of the locally updated weights w k t+1 across all K client nodes.

Jane: Data Harmonization and Model Architecture To enable federated learning across the two cohorts, both datasets were transformed into a standardized, HL7 FHIR-compliant format using the MyDigiTwin data harmonization pipeline This pipeline follows a three-stage workflow: first, raw tabular cohort data in CSV format were converted into patient-centered Common Data Format (CDF) files Second, dataset-specific mapping rules, referred to as pairing rules, were applied to transform each CDF file into FHIR resources compliant with the ZIB profile used in Dutch personal health environments Third, the resulting FHIR resources were flattened into a tabular structure that could be used by the federated learning algorithm.

Tom: The risk prediction model used was DeepSurv (Katzman et al., 2018) which extends the Cox proportional hazards (PH) model by replacing its linear predictor with a feed-forward neural network.

Title and authors: Lu: It’s interesting that they were focused on validating the infrastructure and workflow before trying to prove the predictive benefit across truly different sites >

Tom: Right, so they built this whole system first to make sure the plumbing works before worrying about whether it actually makes a better model >

Jane: And they did that by focusing on how data distribution and local updates behave within the Vantage6 nodes, which is pretty important for anyone building these kinds of systems.

Meng: From an engineering view, validating the infrastructure means we’re checking if the communication between those nodes and the central server is reliable at scale >

Lu: And they specifically looked at how differences in outcome definitions and follow-up information might affect model performance when you combine cohorts with different data profiles >

Tom: Because the two cohorts, Lifelines and Rotterdam Study, have different ways of defining cardiovascular events, so that’s a real challenge for any model trying to merge them.

Jane: And they tackled that by using the MyDigiTwin pipeline to harmonize everything into a standard HL7 FHIR format before training even starts >

Meng: That harmonization step sounds like it adds significant preprocessing overhead, but it seems necessary to make the federated learning algorithm work across those different data types.

Lu: The paper’s structure shows they were trying to move beyond just using a single cohort for evaluation, which is where this study really pushes the idea forward >

Tom: It’s about proving that you can use federation to bridge differences in data quality and follow-up procedures between different studies >

The paper's summary: Jane: Basically, the summary is that they are investigating whether combining a large cohort with self-reported outcomes and a smaller cohort with digitally linked clinical outcomes can enhance the predictive power of a deep learning model.

Tom: They showed that this approach improves discrimination in both datasets when federated training is used compared to when you just train locally on each one >

Lu: Specifically, for the Rotterdam Study, the C-statistic increased from zero point seven two eight with no aggregation to zero point seven three nine with federation >

Meng: So they are saying that the combination of data sources isn't inherently bad; it can actually be beneficial if you use a federated learning mechanism to leverage it >

Jane: But the summary also highlights a crucial caveat: they are extending an infrastructure-focused evaluation to see if federation can improve performance across different cohorts without exchanging any individual-level participant data >

Tom: It’s important to remember that this paper isn't claiming a universal solution for all multi-site data sharing, it’s specifically about the mechanism of privacy preservation through federated learning >

Lu: They are focused on providing a practical way to build predictive models across multiple institutions while respecting strict privacy regulations like GDPR >

The paper's improvements: Jane: They found that incorporating Lifelines into the Rotterdam Study improved its performance, showing that the information from a much larger cohort can partially compensate for issues like uncertainty in self-reported outcomes >

Tom: That’s a direct implication: if you have a large dataset, even if it has some noise, it can help pull the final model prediction closer to what the true risk is >

Lu: This points toward adaptive stopping criteria being beneficial when training across heterogeneous cohorts because the size of each cohort influences how much influence its updates have on the global model >

Meng: From a deployment standpoint, that means we need to be careful about how we weigh those updates; you can't just let the largest dataset dominate everything without checking things >

Jane: And they also highlighted that this setup is consistent with GDPR requirements and data governance agreements, which makes it feasible for real-world use >

Tom: So, the improvement isn't just about getting a higher C-statistic; it’s about showing that you can build these sophisticated models without moving sensitive patient data around >

Conclusion: Jane: It shows that we don't have to rely on a single data source if you want a robust model >

Tom: The key is using federated learning to combine those sources without ever moving the raw patient data between institutions >

Lu: It confirms that the information in the substantially larger Lifelines cohort contributes more strongly to the aggregated model than the Rotterdam Study, which we saw in their performance metrics >

Meng: For us shipping models, it means we need to monitor that imbalance carefully during training because you can't assume a simple uniform contribution from every node >

Jane: It’s a solid paper on how to build these systems for privacy and accuracy simultaneously, and it gives us a good blueprint for future work in this area >

Tom: The paper "Federated Deep Learning for Privacy-Preserving Cardiovascular Disease Risk Prediction" lays out the mechanics of using deepSurv across two distinct cohorts to achieve better risk prediction while maintaining data privacy >

More episodes

← Home