Research map / Federated learning and privacy / Privacy attacks and secure learning
IGAMT: Privacy-Preserving Electronic Health Record Synthesization with Heterogeneity and Irregularity
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes IGAMT, a differentially private seq2seq transformer GAN framework that synthesizes EHRs with heterogeneous temporal/non-temporal features, missing-value masks, and irregular time embeddings, using an Imitator decoder to reduce DP noise.
Problem
Generating synthetic electronic health records that preserve heterogeneous features, structurally missing values, and irregular measurement times, while providing differential privacy with good privacy-utility trade-off.
Approach
Combines a transformer seq2seq autoencoder with mask-based missing-value modeling, time embeddings, and non-temporal embeddings inside three GANs; adds an Imitator decoder that mimics the decoder but only needs gradient perturbation on its shared last layer and discriminator, so the released generator is DP with less noise.
Key results
- On MIMIC-IV-ED clustering, IGAMT (DP) achieves Minkowski distance 3.97 and cosine similarity 0.82 versus DAAE (DP) 5.97 and 0.47.
- DBP classification on synthetic EHR reaches 82.00% test accuracy versus 80.50% on real data.
- Under (epsilon, delta)=(1.5, 1e-5) DP, IGAMT's privacy-utility trade-off is reported to be better than baselines, with the Imitator avoiding utility loss from gradient perturbation.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IGAMT-DP instead of DAAE-DP | EHR synthesis · EHR downstream clustering | Minkowski distance: 5.97 -> 3.97 | Helped |
| Use IGAMT-DP instead of DAAE-DP | EHR synthesis · EHR downstream clustering | Cosine similarity: 0.47 -> 0.82 | Helped |
| Use IGAMT-DP instead of DAAE-DP | EHR synthesis · EHR feature distributions | KL divergence: Feature #0: 0.2 -> 0.2; feature #1: 25 -> 20.78; feature #2: 24.9 -> 7.88; feature #3: 33 -> 3.37; feature #4: 11.01 -> | Mixed |
| Train the DBP classifier on IGAMT synthetic EHR instead of real EHR | CNN-LSTM DBP classifier · DBP classification | test accuracy: 80.50% -> 82.00% | Helped |
| Use IGAMT with missing-value masks and time embeddings to generate EHR | EHR synthesis · MIMIC-IV-ED and Emory Synergy | similarity of missing-value and elapsed-time distributions | Helped |
| Use the Imitator and apply gradient perturbation to it rather than the decoder | IGAMT · MIMIC-IV-ED and Emory Synergy | synthetic EHR utility under differential privacy | Helped |
Surprising findings
- Synthetic data classifier accuracy (82.00%) slightly exceeds that trained on real data (80.50%).
Limitations
- Equal privacy budget allocation across the three DP components is used and not optimized.
- Evaluation relies on PCA, KL divergence, clustering and a single DBP classification task rather than broader downstream benchmarks.
- Uses basic moment accountant composition rather than tighter DP analysis.
Experimental setup
Datasets: MIMIC-IV-ED (vital sign subset) Emory Synergy EHR
Benchmarks: MIMIC-IV-ED Emory Synergy
Metrics: PCA visualization KL divergence of feature distributions Minkowski distance cosine similarity classification accuracy training loss
Baselines: DAAE VAE GAN VAE-GAN AAE
Builds on: DAAE GAN moment accountant medGAN
Closest papers by meaning
- Collaborative Synthesis of Patient Records through Multi-Visit Health State Inference (AAAI 2024 · similarity 0.88)
- Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing (NeurIPS 2025 · similarity 0.84)
- An Investigation of Memorization Risk in Healthcare Foundation Models (NeurIPS 2025 · similarity 0.83)
- Differentially private learners for heterogeneous treatment effects (ICLR 2025 · similarity 0.82)
- Breaking the Barrier of Hard Samples: A Data-Centric Approach to Synthetic Data for Medical Tasks (ICML 2025 · similarity 0.81)
- Advancing Medical Multimodal Learning and Data Generation with Diffusion Model and LLM (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.