TRUST LONGSYNTH: Differential Privacy Meets Longitudinal Health Data
Synthesizing patient timelines that survive an epsilon-12 budget without collapsing minority cohorts has been the open problem in medical ML for five years.
TRUST LONGSYNTH: Differential Privacy Meets Longitudinal Health Data
Synthesizing patient timelines that survive an epsilon-12 budget without collapsing minority cohorts has been the open problem in medical ML for five years. A September 2026 preprint from the TRUST group (arXiv:2609.22401) puts a number on the trade: 7.5% AUPRC gain over the closest private baseline, canary exposure held to 1.8% instead of 28.8%. The catch is stated up front — the authors call this a protocol, not a clinical claim.
Why irregular records break most DP generators
Longitudinal health records are irregular by construction. Encounters land at uneven intervals, feature vectors have holes, and rare events (adverse reactions, seizures, sepsis) drive most of the downstream utility. Naive differentially private diffusion collapses these signals first because Gaussian noise scales with feature covariance, and rare events sit in the low-density tail. The paper’s fix is architectural: a block-banded temporal covariance, a protected event sampling floor, and score computation performed under zero-concentrated DP accounting rather than the usual moments accountant.
On a 720-patient benchmark at ε=12 and δ=1e-5, the method beat a private diagonal score baseline on calibration, cross-feature correlation, and lag-1 autocorrelation simultaneously — the three axes where DP synthetic data usually pays the biggest tax (NIST SP 800-226 on DP guarantees).
epsilon=12, delta=1e-5
canary_exposure: 1.80 (DP Score baseline: 28.83)
auprc_lift: +7.5%
calibration_ece: -0.041
autocorr_error_lag1: -0.12
protected_event_floor: 0.08The pipeline in one picture
The audit stage matters as much as the generator. Canary exposure is the memorization proxy the community has settled on since 2019, and dropping from 28.8 to 1.8 is the kind of delta a data protection officer can actually read. That maps directly to what GDPR Article 9 demands for health data processed under Recital 26 pseudonymisation logic, and to the anonymisation thresholds ENISA set out in its 2022 advanced pseudonymisation guidance.
Where this sits in the health-data stack
The European Health Data Space regulation makes secondary use of health data the default by 2029, with hospitals expected to release derivative datasets rather than raw records. A DP synthesizer that keeps AUPRC on downstream sepsis models while proving low memorization is exactly the artefact regulators will accept. Our engineering view at CAI Janus is that ε=12 remains too loose for cross-border release — ε≤4 with a fairness gate is the target for anything crossing a border under EHDS Article 33. We build the same instrumentation into our AEGIS privacy audit line and into synthetic-data safety work published under CAI RAG.
The clinical validity claim the authors deliberately avoided is the one your medical device notified body will demand next. Talk to us before you promise a downstream label.