Healthcare AI Is Moving Beyond Real-World Data With Big Opportunities and Bigger Questions

Artificial intelligence in healthcare has long been constrained by a paradox: the systems most capable of improving diagnosis, prediction and planning often require access to large volumes of high-quality clinical data that are difficult, expensive or legally sensitive to obtain. Synthetic data promises a way around that bottleneck by creating artificial medical records, images and physiological signals that resemble real clinical information without simply reproducing individual patients.

A new systematic review argues that these technologies are moving rapidly beyond the narrow image of "deepfakes" and toward a broader role as healthcare infrastructure. The study, "From Manipulation to Antidote: Mapping the Computational Capabilities of AI-Generated Synthetic Media to Health-Related Applications and Downstream Benefits," was published in Computers by Wellington Kanyongo and Mampilo Phahlane of the University of South Africa.

The authors reviewed 24 empirical studies published between 2021 and 31 May 2026 after screening 2,214 records across Google Scholar, Scopus, Web of Science and IEEE Xplore. Their synthesis spans synthetic medical images, electronic health records, physiological signals, virtual patients, audiovisual content and deepfake-detection systems, offering one of the clearest attempts yet to connect what synthetic media can technically do with where it may actually create value in healthcare.

Synthetic Data Is Moving From Technical Curiosity to Health Infrastructure

The review finds that synthetic media can now do far more than generate visually convincing medical images. Across the included studies, researchers produced synthetic retinal scans, X-rays, dental radiographs, CT images and MRIs that retained clinically meaningful anatomical and pathological features, while other systems generated artificial ECG signals, cancer registries, intensive-care records and longitudinal electronic health data.

Healthcare AI depends not simply on data volume but on data structure. A synthetic record is useful only if it preserves the correlations, timing, disease patterns and clinical relationships that make real-world data informative. The review reports evidence that synthetic datasets can retain statistical distributions, physiological dependencies and relationships among diagnoses, variables and outcomes rather than behaving like disconnected artificial samples.

Medical imaging still dominates the field: 14 of the 24 reviewed studies used imaging data, while seven involved EHRs, registries or tabular clinical information. GAN-based architectures were the most common computational approach, appearing in 75% of studies, although diffusion models, transformers and other generative systems are beginning to expand the technological landscape.

Synthetic media could become a new layer between sensitive patient data and AI development, allowing researchers to create training environments, simulate rare cases and test algorithms without relying exclusively on unrestricted access to original clinical records. That makes synthetic data less a substitute for real evidence than a potential infrastructure for extending what limited evidence can support.

The Biggest Opportunity Is Where Real Clinical Data Is Scarce

Data augmentation is the most common practical use identified in the study. Synthetic images, ECGs, radiomics data and electronic records are already being used to expand small or imbalanced datasets, particularly where rare conditions or minority classes are underrepresented and where collecting additional real-world cases would be slow, costly or impractical.

It could have particular relevance for developing countries and lower-resource health systems. Many healthcare AI models are trained on datasets concentrated in wealthier institutions and populations, creating concerns about representativeness and generalizability. Synthetic data could help researchers build larger training sets around rare diseases or underserved groups, but only if the source data accurately captures those populations in the first place.

The review also finds applications in diagnostic-model development, including classification, segmentation and prediction across ophthalmology, radiology, dentistry, oncology and neuroimaging. Synthetic EHRs and registry records have additionally supported mortality prediction, risk modelling, patient clustering and health-system analytics, suggesting that the technology could eventually influence not only clinical AI but resource allocation and population-health planning.

Education may become another important use case. Synthetic medical images can expose trainees to uncommon disease presentations without relying solely on available clinical cases, and the review notes that learners trained with synthetic OCT images achieved improvements comparable to those trained with real images in one educational context. Patient-facing applications, including personalized avatars and synthetic video content, have also been explored for communication and anxiety reduction.

Deepfake Risk Means Healthcare Must Verify What AI Creates

The same technology that can solve data shortages can also fabricate clinical evidence. The review highlights manipulation techniques capable of adding or removing tumors, altering disease-related characteristics and transforming medical images while maintaining visual plausibility. In a healthcare setting, that dual-use capability creates risks ranging from insurance fraud and misinformation to unnecessary treatment and diagnostic error.

The study treats deepfake detection as a distinct safety layer rather than merely another generative capability. Researchers are already developing systems to distinguish authentic medical content from GAN- or diffusion-generated images, including detection of manipulated CT scans, fake radiographs and injected or removed pathological features.

Privacy requires equally careful treatment. Synthetic data is often promoted as a privacy-preserving alternative to identifiable patient records, but the review makes an important distinction: data can be privacy-oriented without being demonstrably private. Memorization, reconstruction and similarity to original training records can still create re-identification risk, meaning synthetic datasets should undergo membership-inference testing, reconstruction attacks and other privacy assessments before release.

The existence of artificial data should not create a regulatory shortcut. Hospitals, research institutions and technology companies may need to treat synthetic-data validation as rigorously as they treat cybersecurity or clinical model validation, especially where generated information influences diagnosis, insurance, treatment or patient communication.

The Real Policy Test Is Validation, Not Generation

The authors propose a Computational Capability–Application–Benefit, or CAB, framework. It separates demonstrated capabilities from evaluated applications and then from downstream benefits that remain inferred or require further validation, preventing technical success from being prematurely translated into claims of improved patient outcomes or more efficient health systems.

The framework also introduces maturity levels, ranging from technical feasibility and retrospective validation to clinician or user evaluation and ultimately real-world health-system implementation. Most evidence in the reviewed literature remains concentrated in the earlier stages, while actual implementation at scale is limited. This matters because a model that generates convincing images in a laboratory has not necessarily demonstrated safe or useful performance in routine care.

The review has strengths, including PRISMA-guided screening, independent methodological assessment and coverage across imaging, electronic records, physiological signals and patient-facing media. Twenty of the 24 included studies were rated high quality, but substantial heterogeneity in study design, datasets, evaluation methods and generative architectures limits direct comparison and makes pooled conclusions difficult.

Health regulators and procurement authorities should ask not merely whether synthetic data can be generated, but whether it is clinically plausible, representative, privacy-tested, resistant to manipulation and proven useful for a defined task. Standards for provenance, bias evaluation, privacy attacks and authenticity verification could become as important to synthetic health data as conventional quality standards are to medical devices.

For the Global South health systems, the opportunity is particularly significant but equally conditional. Synthetic data could widen participation in AI research where access to large clinical datasets is constrained, potentially supporting more locally relevant digital-health development. Yet unequal computing capacity, limited validation expertise and poor source datasets could also turn synthetic media into another technology imported without sufficient local evidence.

The authors accordingly call for future work to move beyond technical benchmarks toward clinical trials, patient outcomes and routine healthcare implementation. They also emphasize standardized evaluation of realism, fidelity, fairness, privacy and clinical relevance, with particular attention to rare diseases, underrepresented populations and low-resource settings.