SynEHR: Joint Modeling Inter-visit Temporal Evolution and Intra-visit Clinical Structure for Longitudinal EHR Synthesis

Ximiao Li; Lin Jiang; Rongchao Xu; Dahai Yu; Zhe He; Guang Wang

cs.LGAug 21, 2026

SynEHR: Joint Modeling Inter-visit Temporal Evolution and Intra-visit Clinical Structure for Longitudinal EHR Synthesis

Authors: Ximiao Li, Lin Jiang, Rongchao Xu, Dahai Yu, Zhe He, Guang Wang

Organizations: Florida State University Tallahassee, Florida, United States

Abstract

Longitudinal electronic health records (EHRs) document patients' sequences of clinical visits over time, preserving the temporal evolution of disease progression and care delivery. However, real longitudinal EHRs are difficult to access because they contain large amounts of fine-grained, patient-specific information. Synthetic EHR generation therefore provides a valuable approach for preserving the statistical patterns and clinical structure of patient visit trajectories, enabling broader modeling and analysis when real records are limited. Although recent generative models have made progress in producing future visit sequences, they remain limited in explicitly integrating inter-visit irregular temporal evolution and intra-visit clinical event structures in EHRs, leading to clinically inconsistent and temporally unrealistic visit sequences. In this work, we propose SynEHR, a lightweight adaptive LLM-based framework for longitudinal EHR synthesis. There are two novel designs in SynEHR, i.e., a Temporal State Conditioning Module captures irregular temporal states across visits and a Temporal-Relational Adaptation Module combines these states with patient history to dynamically construct patient-specific relational representations. SynEHR then builds on a parameter-efficient LoRA-adapted language-model generator with next-visit generation capability to train the two modules for temporally and clinically informed generation. Extensive experiments on real-world EHR datasets across fidelity, privacy, and downstream utility evaluations demonstrate that SynEHR outperforms state-of-the-art models by generating more clinically coherent and temporally faithful longitudinal EHR data.

Figures & tables

Figure 1. Illustration of regular and irregular visit patterns.

Panel of Figure 2 — Figure 2. Temporal-regime-specific medication enrichment among visits with Malignant Neoplasm (CCSR NEO072) in the MIMIC-IV dataset.

Figure 3. Overview of the SynEHR framework. The Temporal State Conditioning Module (TSCM) encodes inter-visit intervals into a compact temporal state (Sec. 4.1), and the Temporal-Relational Adaptation Module (TRAM) captures stable and dynamic clinical relations across data types from the patient history (Sec. 4.2). The fused representation is projected into patient-specific soft prefix tokens and injected into the LoRA-adapted backbone to guide autoregressive next-visit generation (Sec. 4.3).

MIMIC-III		MIMIC-IV
Total Patients	1,795	Total Patients	56,028
Total Visits	7,274	Total Visits	334,391
Diagnosis	3,366	Diagnosis	24,062
Medications	2,363	Medications	4,479
Lab Items	633	Lab Items	893
Procedure	1,019	Procedure	12,218

Table 1. Statistics of used datasets.

Dataset	Data Type	Metric	MLP	BEHRT	medGAN	SynTEG	EVA	TWIN	TabDDPM	MedDiff	ScoEHR	EHRPD	PromptEHR	HALO	HiSGT	EHR2Path	SynEHR-L3.1	SynEHR-Q2.5	SynEHR-Q3
MIMIC-III	Diagnosis	LPL	318.74	96.42	236.48	35.76	28.94	26.11	111.27	651.38	672.44	16.42	72.84	88.57	39.86	15.73	16.18	15.61	15.34
	Diagnosis	MPL	305.86	91.35	224.17	33.94	27.48	24.96	103.84	629.57	648.63	15.51	68.71	81.42	37.24	15.63	15.88	15.54	15.46
	Medication	LPL	541.28	148.73	394.67	74.68	41.52	39.36	182.54	918.63	921.46	20.46	88.34	96.18	63.92	18.96	19.48	18.81	18.39
	Medication	MPL	508.46	142.51	376.42	71.93	39.68	38.12	169.87	887.14	896.85	18.74	81.57	89.33	60.85	18.91	19.02	18.83	18.79
	Lab Item	LPL	162.47	48.37	79.84	24.96	18.74	16.91	56.48	401.92	419.35	15.08	39.46	84.72	22.41	14.51	14.88	14.63	14.57
	Lab Item	MPL	154.32	45.82	74.19	23.41	17.82	16.36	51.27	388.16	402.84	14.22	35.78	76.39	21.08	14.37	14.58	14.31	14.41

Table 2. Code-level fidelity comparison. Lower is better. SynEHR-L3.1,SynEHR-Q2.5, and SynEHR-Q3 denote SynEHR with LLaMA-3.1, Qwen2.5, and Qwen3 backbones, respectively. Best values are shown in bold, and second-best values are underlined.

Dataset	Statistic	MLP	BEHRT	medGAN	SynTEG	EVA	TWIN	TabDDPM	MedDiff	ScoEHR	EHRPD	PromptEHR	HALO	HiSGT	EHR2Path	SynEHR-L3.1	SynEHR-Q2.5	SynEHR-Q3
MIMIC-III	Diagnosis Code Freq.	0.346	0.186	0.337	0.378	0.274	0.290	0.152	0.134	0.180	0.228	0.224	0.361	0.229	0.042	0.079	0.146	0.129
	Medication Code Freq.	0.379	0.201	0.415	0.291	0.334	0.241	0.192	0.190	0.231	0.255	0.229	0.329	0.234	0.138	0.176	0.187	0.181
	Lab Item Code Freq.	0.038	0.095	0.047	0.047	0.014	0.152	0.113	0.201	0.241	0.248	0.162	0.057	0.065	0.024	0.013	0.021	0.023
	Procedure Code Freq.	0.430	0.301	0.298	0.494	0.283	0.360	0.148	0.199	0.241	0.262	0.258	0.401	0.331	0.111	0.106	0.168	0.141
	Visit Length	0.193	0.266	0.662	0.632	0.116	0.268	0.046	0.425	0.371	0.559	0.224	0.662	0.040	0.640	0.017	0.034	0.029
	Inter-visit Time	0.250	0.227	0.274	0.332	0.371	0.171	0.055	0.071	0.090	0.193	0.243	0.274	0.159	0.271	0.049	0.068	0.061

Table 3. JSD between real and synthetic data distributions. Lower is better.

Figure 4. Privacy assessment on MIMIC-III. Lower is better.

Approach	10%	20%	35%	50%
MLP	13.47 [13.43, 13.48]	13.29 [13.17, 13.33]	13.08 [12.96, 13.28]	13.12 [12.97, 13.26]
BEHRT	13.58 [13.55, 13.62]	13.44 [13.25, 13.62]	13.19 [13.09, 13.34]	13.27 [13.17, 13.36]
medGAN	16.94 [16.78, 17.04]	17.11 [16.92, 17.13]	17.43 [17.33, 17.55]	17.67 [17.47, 17.73]
SynTEG	13.18 [13.02, 13.36]	12.98 [12.78, 13.00]	12.76 [12.68, 12.89]	12.81 [12.75, 12.97]
EVA	13.31 [13.21, 13.33]	13.12 [13.11, 13.17]	12.88 [12.84, 12.99]	13.28 [13.22, 13.39]
TWIN	13.29 [13.21, 13.45]	13.14 [13.02, 13.29]	12.86 [12.72, 13.05]	12.91 [12.74, 13.07]

Table 4. PD Comparisons.

Model	ARF	Shock	Mortality	Avg.
MLP	0.690 / 0.318	0.652 / 0.236	0.701 / 0.207	0.681 / 0.254
BEHRT	0.716 / 0.347	0.674 / 0.258	0.732 / 0.230	0.707 / 0.278
medGAN	0.648 / 0.283	0.615 / 0.198	0.666 / 0.178	0.643 / 0.220
SynTEG	0.673 / 0.302	0.637 / 0.216	0.688 / 0.190	0.666 / 0.236
EVA	0.692 / 0.324	0.655 / 0.236	0.708 / 0.215	0.685 / 0.258
TWIN	0.706 / 0.337	0.671 / 0.258	0.723 / 0.225	0.700 / 0.273

Table 5. Risk prediction results using multiple data types. Each entry is reported as AUROC / AUPRC. The Avg. column averages the three task-specific scores under the same metric.

Model	Macro-F1 $\uparrow$	MAE $\downarrow$	NLL $\downarrow$	Brier $\downarrow$	ECE $\downarrow$
MLP	0.287	178.6	1.070	0.558	0.148
BEHRT	0.319	169.4	1.018	0.523	0.126
medGAN	0.254	191.8	1.128	0.606	0.173
SynTEG	0.276	184.7	1.091	0.582	0.157
EVA	0.291	177.2	1.058	0.552	0.139
TWIN	0.303	173.5	1.039	0.538	0.132

Table 6. Time interval prediction results under the train-on-synthetic, test-on-real protocol. Higher is better for Macro-F1, while lower is better for MAE, NLL, Brier score, and ECE.

Variant	Avg. LPL $\downarrow$	Avg. MPL $\downarrow$	Inter-visit Time JSD $\downarrow$	ARF AUROC $\uparrow$	Shock AUROC $\uparrow$	Mortality AUROC $\uparrow$
SynEHR	16.03	16.01	0.059	0.827	0.791	0.838
SynEHR w/o TSCM	17.64	16.94	0.113	0.798	0.753	0.812
SynEHR w/o TRAM	17.02	18.25	0.078	0.791	0.745	0.805
SynEHR w/o Static Branch	16.70	16.78	0.071	0.815	0.781	0.823
SynEHR w/o Dynamic Branch	17.14	17.24	0.096	0.806	0.762	0.819
SynEHR w/o Confidence-Aware Fusion	16.44	16.48	0.068	0.819	0.784	0.831

Table 7. Ablation results on MIMIC-III. We report representative fidelity and utility metrics.

Synthetic Data Generation

Synthetic data generation creates artificial examples, simulations, and datasets to train, test, or protect machine-learning systems.

See all 620 papers

Mar 24, 2026

CDMT-EHR: A Continuous-Time Diffusion Framework for Generating Mixed-Type Time-Series Electronic Health Records

Electronic health records (EHRs) are invaluable for clinical research, yet privacy concerns severely restrict data sharing. Synthetic data generation offers a promising solution, but EHRs present unique challenges: they contain both numerical and categorical features that evolve over time. While diffusion models have demonstrated strong performance in EHR synthesis, existing approaches predominantly rely on discrete-time formulations, which suffer from finite-step approximation errors and coupled training-sampling step counts. We propose a continuous-time diffusion framework for generating mixed-type time-series EHRs with three contributions: (1) continuous-time diffusion with a bidirectional gated recurrent unit backbone for capturing temporal dependencies, (2) unified Gaussian diffusion via learnable continuous embeddings for categorical variables, enabling joint cross-feature modeling, and (3) a factorized learnable noise schedule that adapts to per-feature-per-timestep learning difficulties. Experiments on two large-scale intensive care unit datasets demonstrate that our method outperforms existing approaches in downstream task performance, distribution fidelity, and discriminability, while requiring only 50 sampling steps compared to 1,000 for baseline methods. Classifier-free guidance further enables effective conditional generation for class-imbalanced clinical scenarios.

Shaonan Liu, Yuichiro Iwashita, Soichiro Nakako +2

Department of Core Informatics, Graduate School of Informatics, Osaka Metropolitan University, Sakai, Osaka 599-8531, Japan · Smart Data & Knowledge Services Department, German Research Center for Artificial Intelligence (DFKI GmbH), 67663 Kaiserslautern, Germany · Department of Computer Science, RPTU University Kaiserslautern-Landau, 67663 Kaiserslautern, Germany +1

Synthetic Data Generation Diffusion Models Electronic Health Records Longitudinal EHR Modeling Diffusion Timestep Scheduling

Ask PDF

May 27, 2026

FedEHR-Gen: Federated Synthetic Time-Series EHR Generation via Latent Space Alignment and Distribution-Aware Aggregation

Synthetic Electronic Health Record (EHR) generation provides a promising avenue for data augmentation and cross-hospital modeling in privacy-constrained healthcare settings. However, most existing EHR generative models are centralized and require pooling data across hospitals, which is often infeasible when real-world data sharing is restricted. While federated EHR generation offers a natural solution, direct federated modeling often collapses or diverges due to the high dimensionality, sparsity, and cross-hospital heterogeneity of EHR data. In this work, we propose FedEHR-Gen, the first federated framework for synthetic time-series EHR generation across distributed hospitals. FedEHR-Gen uses a two-stage learning paradigm. First, we introduce a federated autoencoder that projects high-dimensional and sparse EHR features onto a compact latent space. To ensure semantic consistency across hospitals, we develop a layer-wise matching aggregation mechanism that aligns local encoders into a unified global latent space. Second, operating on this aligned latent space, we train a federated temporal conditional variational autoencoder (TCVAE) with distribution-aware aggregation, enabling stable temporal generative modeling under severe cross-hospital heterogeneity. Extensive experiments on the eICU and MIMIC-III datasets demonstrate that FedEHR-Gen achieves generation fidelity, downstream utility, and privacy risk comparable to centralized training, while consistently outperforming the standard federated baseline.

Jun Bai, Ziyang Song, Yue Li

McGill University Mila - Quebec AI Institute Montreal, Canada · Ohio University Athens, United States

Synthetic Data Generation Conditional Variational Autoencoder Longitudinal EHR Modeling Electronic Health Records Heterogeneous Federated Learning

Ask PDF

Sep 24, 2026

Synthetic Hospital: An Open, Verifiable, Physician-Validated Longitudinal EHR Benchmark

Frontier language models are rarely used in clinical workflows because the realistic, longitudinal benchmarks needed to develop them are scarce. Real electronic health record (EHR) data cannot be openly shared due to privacy, ethics or data use issues and it does not contain verifiable ground truth since the chart records only reflect what clinicians documented. We introduce Synthetic Hospital, an open, fully synthetic, fact-grounded longitudinal EHR benchmark that resolves the open sharing and verifiable ground truth barriers. Built entirely from public medical-education material with no protected health information, it comprises 1,268 longitudinal patients and 5,602 encounters, where every diagnosis, finding, and temporal relation is grounded in standard ontologies (ICD-10-CM, SNOMED CT, LOINC) and with a complete provenance chain back to its source medical education material. Synthetic Hospital is served through a simulated hospital record system that mirrors real EHR infrastructure (standard interoperability APIs, role-based access and function-calling interface). In a blinded review, physicians distinguished its records from real patient charts at near-chance rates (53%). Across 10 frontier and open models, none approaches ceiling: the best model reconstructs a patient's longitudinal problem list with a severity-weighted F1 of 0.73, level with the mean of seven physicians on a matched subset but well below the best of them (0.89), and misses roughly half of clinically relevant findings when summarizing a chart. Overall, these results highlight that Synthetic Hospital is a difficult and realistic test of clinical AI performance.

Christine Park, Valerie Chen, Tim Dettmers

Carnegie Mellon University

Synthetic Benchmark Generation Clinical Language Model Evaluation Electronic Health Records Synthetic Data Evaluation

Ask PDF

Abstract

Synthetic Data Generation

Explore similar work

CDMT-EHR: A Continuous-Time Diffusion Framework for Generating Mixed-Type Time-Series Electronic Health Records

FedEHR-Gen: Federated Synthetic Time-Series EHR Generation via Latent Space Alignment and Distribution-Aware Aggregation

Synthetic Hospital: An Open, Verifiable, Physician-Validated Longitudinal EHR Benchmark