Synthetic Clinical Data with Mixtures of Experts

Real clinical data is scarce, sensitive, and hard to share. This research project asks whether a Mixture-of-Experts architecture can generate synthetic patient data that is faithful enough to train on, yet carries no real patient inside it.

It is a group project of roughly fifteen students. I own two phases. The first is frugal adaptation: parameter-efficient fine-tuning (LoRA and related PEFT methods) so a large foundation model can be specialised on limited clinical data without retraining from scratch.

The second is evaluation, which is where synthetic data projects usually fall down, and where I was made coordinator. The framework scores generated data on three axes (statistical fidelity, clinical plausibility, and downstream utility via train-on-synthetic-test-on-real) and aggregates them by weakest link rather than by average, so a model cannot pass by being excellent on one axis and useless on another. It covers 24 MIMIC-IV feature definitions and three downstream classifiers.

Two teammates independently built their own evaluation tooling; mine is the implementation that was reviewed and merged to main, and another student has since run it on a different machine and operating system.

What it has not done yet is produce a result. The generation side emits a different key format, so a format adapter is still needed, and the labels required for the utility axis are missing from the dataset as delivered. The thresholds in the framework are design targets, not measurements.