Research map / Image generation / Autoregressive and latent generation
Collaborative Synthesis of Patient Records through Multi-Visit Health State Inference
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 MSIC, a probabilistic graphical model with latent multi-visit health states that synthesizes structured EHR events (symptoms, diagnoses, medications) and generates multi-paragraph medical reports via a multi-generator deliberation framework built on BioBART.
Problem
Synthetic EHR generation methods treat symptoms, diagnoses and medications symmetrically, so generated event combinations may violate medical common sense, and they do not use previous visits well or produce textual descriptions.
Approach
A latent health state per patient, inferred sequentially across visits with a variational prior/posterior, links the event types and conditions both cross-type and longitudinal synthesis; medical reports are generated by three BioBART-based paragraph generators that exchange drafts in a two-phase draft-and-polish decoding.
Key results
- On MIMIC-III, MSIC achieves Unigram 0.846, Bigram 0.701, Seq Bigram 0.645 and DimWise 0.972 for event synthesis, exceeding the best baseline PromptEHR (0.795/0.616/0.603/0.921).
- For report generation on MIMIC-III, MSIC reaches BLEU-2 33.23 and ROUGE-L 27.79, versus 29.80 and 24.39 for BioBART.
- Attribute inference F1 on MIMIC-III is 0.059, below the real data attack baseline of 0.095.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add the latent health state h_t and its associated paths (ablation) | MSIC · MIMIC-III medical event synthesis | Unigram: 0.605 -> 0.846 | Helped |
| add multi-visit sequential modeling (ablation) | MSIC · MIMIC-III medical event synthesis | Unigram: 0.820 -> 0.846 | Helped |
| add the latent health state h_t and its associated paths (ablation) | MSIC · MIMIC-III medical report generation | BLEU-2: 31.17 -> 33.23 | Helped |
| add multi-visit sequential modeling (ablation) | MSIC · MIMIC-III medical report generation | BLEU-2: 32.94 -> 33.23 | Helped |
| add message passing between generators for multi-generator deliberation (ablation) | MSIC · MIMIC-III medical report generation | BLEU-2: 29.51 -> 33.23 | Helped |
| add message passing between generators for multi-generator deliberation (ablation) | MSIC · MIMIC-IV medical report generation | BLEU-2: 26.93 -> 29.64 | Helped |
| use MSIC instead of the strongest event-synthesis baseline, PromptEHR | EHR event synthesis · MIMIC-III | Unigram: 0.795 -> 0.846 | Helped |
| use MSIC instead of the strongest report-generation baseline, BioBART | Medical report generation · MIMIC-III | BLEU-2: 29.80 -> 33.23 | Helped |
Surprising findings
- Removing the health state (MSIC w/o h_t) causes a large drop in sequential bigram score on MIMIC-IV (0.001 vs 0.648), showing the health state is critical for co-occurrence coherence.
Limitations
- Synthetic reports still contain false-positive events not present in real data.
- Report generation evaluated only with n-gram metrics (BLEU, ROUGE).
Experimental setup
Datasets: MIMIC-III MIMIC-IV
Benchmarks: MIMIC-III MIMIC-IV
Metrics: Unigram Bigram Seq Bigram DimWise BLEU ROUGE membership inference F1/accuracy attribute inference F1
Baselines: LSTM+MLP MedGAN MedBGAN MedWGAN EVA MTGAN PromptEHR Transformer GPT-2 BioGPT BART BioBART
Builds on: BioBART deliberation networks Fixed-Prompt LM-Tuning PromptEHR EVA
Closest papers by meaning
- Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing (NeurIPS 2025 · similarity 0.88)
- IGAMT: Privacy-Preserving Electronic Health Record Synthesization with Heterogeneity and Irregularity (AAAI 2024 · similarity 0.88)
- ConSequence: Synthesizing Logically Constrained Sequences for Electronic Health Record Generation (AAAI 2024 · similarity 0.85)
- Context Clues: Evaluating Long Context Models for Clinical Prediction Tasks on EHR Data (ICLR 2025 · similarity 0.84)
- One Loss to Rule Them All: Marked Time-to-Event for Structured EHR Foundation Models (NeurIPS 2026 · similarity 0.84)
- Advancing Medical Multimodal Learning and Data Generation with Diffusion Model and LLM (AAAI 2025 · similarity 0.84)
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.