ORIGA is a web platform for clinical data augmentation designed to generate realistic synthetic patients, while ensuring a high level of security and compliance.
It is built around three key features:
Data augmentation with ORIGA involves enhancing your existing datasets by generating new, realistic synthetic data. This technique improves the robustness and performance of artificial intelligence models, particularly for training medical algorithms and conducting clinical trials where available real-world data may be limited.
Data imputation with ORIGA involves completing incomplete or missing datasets using synthetic data that is consistent with your original population. This method enhances the quality of your data and ensures the reliability of analyses, which is particularly critical in the context of clinical trials.
The synthetic patients generated by ORIGA:
Result: you preserve all the richness of clinical data while eliminating the risks associated with its use.
Intensive care anaesthetist, Toulouse University Hospital
The Variational Autoencoders (VAEs) are deep generative models capable of creating realistic synthetic data from learned distributions. By projecting input data into a structured latent space and then resampling from this space, VAEs allow for the generation of new instances that are faithful to the statistical properties of real-world data.
At ORIGA, we’ve developed a range of VAEs tailored to the diversity of medical and industrial data: imaging, tabular data, time-series, image sequences, and multimodal data. Our expertise lies in fine-tuning the latent space and the generation mechanisms based on the type of data.
Option 1: Normalizing Flows – Deterministic evolution in the latent space
The model encodes each observation into a latent variable z_t.
To model the temporal evolution, we don’t impose dynamics on the images themselves, but directly on the latent representations.
With normalizing flows, a smooth and invertible transformation is applied to these latent variables to capture their evolution over time (zₜ₊₁ = fₜ(zₜ)).
Each z_t can then be decoded into an image x̂_t, ensuring temporal continuity and coherence in the generated sequence.
Option 2: Diffusion Process – Stochastic evolution in the latent space
Instead of deterministically transforming the latent representations, the model can learn to diffuse a latent variable over time by gradually adding controlled noise.
We simulate an inverse process where, starting from a noisy state, we learn to move back toward clearer and more structured states.
This mechanism captures a more realistic and stochastic dynamic in the generated sequences, while maintaining the underlying structure of the data.
Thanks to this combination of advanced techniques, our models generate robust, interpretable synthetic data that meet the quality standards required in sensitive contexts such as healthcare or clinical research.
In the biomedical fields, real-world patient data is precious but often scarce, sensitive, or partially observed. Synthetic patients are artificial entities generated from statistical or algorithmic models, designed to replicate the essential properties of real patients while adhering to ethical, legal, and scientific constraints.
Why generate synthetic patients?
Three major model families for generating synthetic patients
What we do at ORIGA
Our generative AI models do not aim to copy real patients, but to create statistically plausible synthetic patients that represent the diversity of clinical profiles. Each generated patient is unique, with no direct correspondence in the original datasets, ensuring confidentiality while preserving analytical value. By integrating probabilistic approaches (VAEs), geometric approaches (Riemannian latent spaces), and temporal approaches (flows and diffusion), our technology produces coherent, faithful artificial cohorts tailored to research, development, and validation needs.
Practical use cases for synthetic patients
At ORIGA, the generation of artificial data is based on principles of scientific rigor, traceability, and responsibility. To ensure the quality of the synthetic data produced, we have implemented a three-pronged validation process: statistical reliability, diversity of generated profiles, and respect for confidentiality. This process combines formal mathematical tests with human business validation.
Statistical Reliability Validation
Our artificial data must be statistically faithful to real data, without reproducing the exact individuals. To achieve this, several tests are systematically applied:
Human Guarantee Committee Validation
The results of the statistical tests are systematically supplemented by a business and medical review. A human guarantee committee, composed of domain experts (clinicians, biostatisticians, industry specialists), analyzes a sample of synthetic data to assess their plausibility, clinical coherence, and operational utility. This hybrid approach allows for a balance between formal rigor and business sense.
Diversity Validation
A good generative model does not simply imitate the original data — it must produce new, realistic, and diverse data that enriches the space of possibilities while respecting statistical and clinical constraints.
To assess this diversification capability, we conduct a collapse test based on a self-regeneration protocol:
Starting with an initial set of artificially generated data, we retrain the model, then generate a new dataset. This process is repeated several times.
Regardless of the model used, this process inevitably leads to a phenomenon called collapse: after a certain number of iterations, the model tends to generate more and more similar profiles, until it produces almost identical data from one pass to the next.
The diversity test measures how many successive iterations can be performed before this collapse occurs significantly.
→ The higher the number of iterations before collapse, the more diverse the initially generated data.
This measure allows us to quantify the richness and exploratory capacity of the model, even from the first generation. It is essential to ensure that the synthetic cohorts produced are not just superficial variations, but truly varied representations, credibly covering the different possible configurations of a phenomenon.
Confidentiality Validation
The generation of artificial data raises an essential question: could the data produced, in any way, allow for the identification of a real individual? To ensure GDPR compliance and non-re-identifiability of patients, we have implemented a formal confidentiality test.
This test is based on a simple idea: comparing the proximity between a real patient and the generated data.
For each real individual, we evaluate if their nearest neighbor among the artificial data is closer than their own reconstructed version (i.e., encoded and then decoded by the model).
We consider confidentiality to be guaranteed when, in at least 50% of cases, the real data is closer to its reconstruction than to any generated individual. This test is a robust indicator of non-memorization by the model. It complements the regularization and training methods we use to avoid overfitting and any risk of information leakage.