Evaluation of synthetic data generation for longitudinal 6-year follow-up subclinical cardiovascular disease data using GAN models
This study evaluates the performance of TGAN and CTGAN models in generating synthetic 6-year longitudinal cardiovascular disease data, finding that while CTGAN generally offers superior fidelity and privacy compared to TGAN and a perturbation baseline, it still requires refinement to better preserve temporal characteristics and performance across younger age groups.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
Heart disease remains the leading cause of death around the world, but much of it is preventable through changes in lifestyle and early intervention. To design better ways to stop these diseases before they start, researchers need vast amounts of personal health information. They look at how blood pressure, cholesterol, and daily habits change over many years to understand the long-term path to illness. However, sharing this kind of detailed, private data is difficult. Strict laws protect patient identities, and people are often reluctant to let even doctors see their full medical history due to fears of privacy leaks. This creates a bottleneck: scientists cannot build the advanced tools needed to prevent heart disease if they cannot access the data required to train them.
To solve this, researchers have turned to a technique called synthetic data generation. Imagine creating a completely new set of patient records that look and behave exactly like real ones, but where every single person is entirely fictional. These fake records preserve the complex relationships between different health factors—like how age, diet, and genetics interact—without containing any actual private information about real people. While this idea has been tested on snapshots of data taken at a single moment, it has not been widely tried on long-term records that track people over years. This gap matters because preventing heart disease is a long-term goal, and tools built on short-term data might miss the slow, subtle changes that lead to illness.
A team of researchers in Sweden and Italy recently set out to test whether current artificial intelligence tools could handle this long-term challenge. They focused on a specific type of AI model known as a generative adversarial network, which works like a pair of competing computers. One computer tries to create fake health records, while the other tries to spot the difference between the fake ones and the real ones. Through this competition, the creator learns to make increasingly realistic data. The team tested two versions of this technology, one called CTGAN and an earlier version called TGAN, using a real-world dataset from a six-year study of heart disease prevention. This study followed thousands of people in Sweden, checking their health at the start, after three years, and again after six years. The researchers wanted to see if the AI could learn the patterns of these changing health records well enough to create a new, fake dataset that scientists could use for research without ever seeing a real patient's name.
The results showed that the newer model, CTGAN, was generally better at creating realistic data than its older counterpart. When the researchers compared the fake records to the real ones, the CTGAN model successfully preserved the statistical patterns of the group as a whole. It was particularly good at handling categories of data, such as marital status or medication types. In terms of privacy, both AI models performed very well, creating data that was far less likely to accidentally reveal the identity of a real person than a simpler, older method of just adding random noise to the numbers. This suggests that using these advanced AI models could allow researchers to share sensitive health data more freely, accelerating the development of new prevention tools.
However, the study also revealed important limitations that researchers must address before these tools are ready for widespread use. While the AI did well with the group as a whole, it struggled more with specific subgroups, particularly younger participants. The data for people in their forties was less accurately recreated than the data for those in their sixties. This is a significant finding because early prevention is most critical for younger adults. The AI also had more trouble accurately recreating the smooth, continuous changes in numbers like blood sugar or cholesterol levels over time compared to the simpler noise-addition method. The researchers noted that this might be because the AI models were originally designed for data taken at a single point in time, not for the complex, shifting patterns of a six-year journey.
Ultimately, the study suggests that while these AI tools are a powerful step forward for medical research, they are not yet perfect. They can generate useful, privacy-safe data for studying heart disease prevention, but they need refinement to handle the specific challenges of long-term tracking and diverse age groups. The authors conclude that this is the first time such a test has been done on a long-term dataset, and their findings provide a necessary benchmark. They indicate that future work must focus on improving how these models handle the passage of time and ensure they treat all age groups fairly, so that the synthetic data can truly serve as a reliable stand-in for real patients in the fight against heart disease.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.