Researchers at the US Food and Drug Administration (FDA) reported in a 2024 review that synthetic radiology data can supplement scarce or unbalanced patient datasets, but uncertainty, bias and fidelity still require validation. The technique can reduce reliance on directly acquired records for some tasks, yet it does not turn a weak dataset into reliable clinical evidence.

A separate 2024 study in Nature found that repeatedly training generative models on their predecessors' output can make rare parts of the original distribution disappear. The study did not test medical imaging, so its model-collapse findings are a warning about recursive data pipelines rather than proof that every medical synthetic-data project will fail.

What synthetic medical data are

Synthetic data are samples created partly or entirely by computation rather than direct acquisition from a person. In medicine, they can include tabular records produced by a statistical model, images created by a generative adversarial network or diffusion model, and scans simulated from models of anatomy and imaging hardware. The 2024 radiology review groups these methods into statistical generative, physics-based and hybrid approaches.

That differs from de-identification, which transforms existing records, and from federated learning, which changes where training happens. Federated learning keeps source records at participating sites and exchanges model updates, while synthetic-data systems create new examples for training, testing or quality assurance. The approaches can be combined, but they address different constraints.

Server racks and network cables in a hospital data center (illustrative image)

Why hospitals and developers consider it

The US FDA says representative medical datasets can be difficult to obtain because of acquisition costs, safety limits, privacy restrictions and low disease prevalence, and it is studying synthetic data as a way to supplement patient datasets. Generated examples can add labeled cases for development or provide standardized inputs for quality-assurance tests. They can also model imaging conditions that would be costly or unsafe to reproduce by repeatedly scanning patients.

Volume is not the same as coverage. The radiology review says developers still need to measure uncertainty, bias, population representation and generalizability across clinical settings, patient groups and devices. A generator trained on a narrow source cannot be assumed to represent cases the source failed to capture.

Synthetic does not mean private by default

Synthetic rows or images are newly generated, but the model that creates them may have learned from sensitive records. The US National Institute of Standards and Technology (NIST) says many synthetic-data methods satisfy neither differential privacy nor another formal privacy property. Calling an output synthetic therefore does not establish that it cannot reveal information about the source dataset.

Formal privacy protection can also reduce utility. NIST says constructing accurate, differentially private synthetic data is difficult because adding privacy-preserving noise can weaken the correlations analysts need. Rules governing the original records and the generated output remain country-specific; APPI News could not verify a single international rule that classifies all synthetic health data as non-personal.

Legal documents and file folders represent rules governing health data use (illustrative image)

Clinical fidelity needs a separate test

The FDA researchers' review says evidence supporting synthetic data should reflect how heavily a medical-device evaluation relies on it and may require comparisons with patient data, cross-validation between generation methods or analysis of distribution gaps. It adds that primary evidence of safety and effectiveness would continue to come from a robust test set of real patients. Those statements are conclusions from a research review, not binding US FDA guidance.

Visual realism alone is insufficient. A synthetic image can appear plausible while omitting a signal needed for the task or reproducing an artifact associated with one scanner. Developers need task-level evaluation on sequestered real data drawn from the populations, clinical settings and equipment covered by the intended use.

Model collapse is a narrower warning

The 2024 Nature paper defined model collapse as a process in which generated data enter the next model's training set, errors compound and low-probability parts of the original distribution disappear first. The researchers demonstrated the effect with Gaussian mixture models, variational autoencoders and an OPT-125m language-model setup. This was recursive generative-model training, not a trial of a diagnostic device.

In the language-model experiment, generation nine veered from a passage about medieval architecture to a list of colored jackrabbits, while retaining a random 10 percent of original data produced only minor degradation in that setup. Neither finding establishes a safe real-data percentage for medical imaging. It also does not show that a single round of synthetic augmentation causes clinical failure.

Digital noise and distorted patterns represent declining data quality (illustrative image)

What evidence should accompany the dataset

Producers should document the source population, generation method, intended use, disclosure testing, representation of relevant groups and any overlap with training records. Evaluation should compare performance on sequestered real data from the patient populations and devices in scope. Synthetic training and test sets also need separate provenance to avoid circular validation.

Synthetic data can serve as augmentation, simulation or quality-assurance material when its limits are measured. It does not by itself resolve access, privacy, bias or clinical validation, and its role must be judged against the laws and evidentiary standards in each country. The deciding factors are whether the dataset protects the source records and remains faithful enough for the specific task it is meant to support.