Privacy-Preserving Active Learning for precision oncology clinical workflows with inverse simulation verification

Description and use

A prototype combines federated DP‑SGD, active learning, and inverse‑simulation verification to train privacy‑preserving, audit‑able models for precision oncology.

Problem

How to train clinically meaningful models on highly sensitive, small, non‑IID oncology datasets without exposing patient data and while providing clinicians a falsifiable audit of model behavior.

Outcome

Active learning combined with client‑level DP‑SGD achieved comparable AUROC to random sampling at a higher privacy budget (ε=3 vs ε=8) on a synthetic oncology benchmark, while the inverse‑simulation verifier provided a falsifiable audit of model decisions.

Workflow

• 1. Deploy a central orchestrator that aggregates encrypted gradients and verification data.
• 2. Each hospital trains a local model with DP‑SGD and per‑feature‑group adaptive clipping.
• 3. Local models send encrypted gradients and acquisition scores to the orchestrator.
• 4. The orchestrator updates the global model and plans active queries for unlabeled cases.
• 5. Hospitals annotate selected cases and send labels back to the orchestrator.
• 6. The inverse‑simulation verifier reconstructs the decision boundary and checks for clinical plausibility.
• 7. The updated global model is redistributed to hospitals for the next round.

Evidence

The hard part was trust — specifically, how do you prove that a model trained on data it never saw actually learned something clinically meaningful?

Source

One interesting finding from my experimentation with client‑level DP: the privacy‑utility curve is flatter when you combine it with active learning.

Source

In my synthetic oncology benchmark, active learning + DP‑SGD at ε=3 matched random sampling + DP‑SGD at ε=8 on AUROC for the actionable‑variant classification task.

Source