Research map / Image generation / Fast diffusion sampling
Advancing Medical Multimodal Learning and Data Generation with Diffusion Model and LLM
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.
A doctoral-consortium style overview summarizing two diffusion-based EHR models (MedDiffusion and EHRPD) and proposing future work toward a generalized multimodal EHR generation model.
Problem
Synthesizing electronic health records is needed to address data scarcity, bias, and fairness, but EHR data are multimodal and sequential, and existing generative models (GANs, VAEs, language models, diffusion models) capture these complexities only partially.
Approach
Summarizes the author's prior diffusion-based EHR work: MedDiffusion uses synthetic patient data augmentation with step-wise attention for health risk prediction, and EHRPD uses time-aware visit embedding and a predictive DDPM with temporal interval estimation; proposes future multimodal generation across codes, notes, images and sensor readings.
Key results
- MedDiffusion reportedly significantly outperforms 14 baselines across multiple datasets on health risk prediction (as summarized from prior work).
- EHRPD reportedly significantly outperforms existing approaches in fidelity, privacy, and utility (as summarized from prior work).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use MedDiffusion to augment training with synthetic patient data and model visit relationships with step-wise attention | MedDiffusion · Multiple datasets (not named) | predictive performance | Helped |
| Add temporal interval estimation and time-aware visit embeddings to EHR generation | EHRPD · EHR generation (dataset not named) | data quality and diversity | Helped |
Limitations
- Current EHR generation is limited to one-to-one or two-to-one modality generation.
- General-domain multimodal generation models are not directly transferable to medical data due to subtle cross-modal relationships.
Experimental setup
Baselines: MedGAN ehrGAN EVA TWIN PromptEHR HALO TabDDPM MedDiffusion EHRPD
Builds on: DDPM TabDDPM MedGAN PromptEHR
Closest papers by meaning
- Addressing Asynchronicity in Clinical Multimodal Fusion via Individualized Chest X-ray Generation (NeurIPS 2024 · similarity 0.84)
- MedSegDiff-V2: Diffusion-Based Medical Image Segmentation with Transformer (AAAI 2024 · similarity 0.84)
- Collaborative Synthesis of Patient Records through Multi-Visit Health State Inference (AAAI 2024 · similarity 0.84)
- Generating Multi-Table Time Series EHR from Latent Space with Minimal Preprocessing (NeurIPS 2025 · similarity 0.84)
- Are Images Indistinguishable to Humans Also Indistinguishable to Classifiers? (CVPR 2025 · similarity 0.83)
- Latent Drifting in Diffusion Models for Counterfactual Medical Image Synthesis (CVPR 2025 · similarity 0.83)
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.