Privacy-safe synthetic ICU data and better-balanced training sets for Acute Kidney Injury prediction — how Eindhoven University of Technology (TU/e) and BlueGen.ai made it possible.
Privacy-safe synthetic ICU data for Acute Kidney Injury prediction.
Eindhoven University of Technology (TU/e) is exploring how to give researchers and students hands-on access to medical data from Dutch hospitals. Sharing that data is not off the table, but it typically means NDAs, ethics approval and long procedural delays before anyone can start. Freek Relouw, a PhD researcher building models to predict Acute Kidney Injury (AKI) in ICU patients, ran into exactly that.
The privacy rules for this dataset are very strict. Sensitive patient records stay inside the secure environment and are never handed to students, so the people Freek supervises could not work with the very data their projects were about. He needed another way.
Why the existing options all failed.
Granting students access to the real ICU data means navigating NDAs, lengthy approvals and a cumbersome secure platform, and every exception granted pushes privacy risk back up. Anonymisation does not solve this either: the records can still be traced back to real patients. Dummy data is not a viable alternative, as it fails to imitate the complex behaviour of real ICU patients. As Freek put it, synthetic data is only worth having if it produces the same results as the real data.
The lead use case made it harder.
Freek’s AKI model reads ICU time series data (vital signs, lab values, medication and demographics) over the first 24 hours. That data is irregular and relational, which pushes synthetic-data generation to its limit. The model itself has a separate problem: it beats published baselines on average, but its accuracy collapses on the rare AKI subtypes, where some subgroups hold only 10 to 20 patients; too few to collect more of within a single PhD.
Realistic, privacy safe patient data.
To address these challenges and make the data available for student use while supporting the development of a more robust prediction model, TU/e and Freek adopted synthetic data generated from the original dataset. BlueGen.ai’s synthetic data generator was trained on both the relational and time-series data to produce realistic synthetic patients. The resulting patients not only closely resemble their real counterparts but also replicate their underlying behaviour and statistical characteristics, without revealing or leaking any information from the original dataset.
BlueGen deployed inside myDRE, TU/e's secure research environment.
BlueGen.ai runs entirely inside myDRE, the secure environment where the ICU data already sits. The source data, the model and the synthetic output never leave it, even during training. A generative model was trained conditionally on the Radboud ICU time-series dataset, learning how vital signs, lab values, medication and demographics relate to each other and to the AKI outcome. Because it learns the joint structure rather than each field in isolation, every synthetic record is a coherent virtual patient. Irregular measurement timing is preserved by representing each patient as a sparse time series.
The synthetic data generator was refined over three successive iterations to match Freek’s exact modelling window: first a six-hour slice, then the full first 24 hours (0 to 1440 minutes), and finally a version that also carries the patient’s baseline state from before ICU admission. Because the model is conditioned on the AKI label, it can generate extra patients on demand for under-represented AKI stages. Quality was checked where it counts: a model trained on synthetic patients was compared directly against one trained on the real data for AKI prediction.
From pilot to a shareable, better-balanced dataset
The final model produced three deliverables: a standard synthetic dataset at the natural AKI prevalence, a balanced dataset with equal AKI representation, and an augmentation dataset to append to the real data to even out the rare categories. In downstream testing 500 real and 500 synthetic patients were used and evaluation was done on 800 held-out patients. The synthetic-trained model performed almost identically to the real-data model on AKI prediction. The dataset carries no identifiable individual, so it is safe to share with students and to publish alongside the external-validation study Freek is preparing.
The same model lets Freek oversample the rare AKI subtypes with conditional generation and rebalance the training set where the model was weakest. That matters because a model that performs well on average but falters on rare cases is not one clinicians can rely on at the bedside.
Share ICU data safely
Synthetic patients that match the real cohort but map to no real person – usable by students and collaborators, and publishable.
Lift accuracy on rare cases.
Under-represented AKI subtypes are oversampled with realistic synthetic patients, rebalancing the training set where the model was weakest.
Raw data stays in myDRE
The model trains where the data already lives. Records never leave the secure environment; only the synthetic output is shared.
“For me, the main value is easier and safer student access to the data, and seeing whether synthetic upsampling of underrepresented groups can improve downstream model performance.”
Freek Relouw,
PhD Researcher TU/e