Evidence-Conditioned Completion and Synthesis of Electronic Health Records
Authors: Sicheng Shu, Parker Knight, Huitao Li, Yuan Luo, Yifan Peng, Shawn N. Murphy, Rui Yang, Chuan Hong
Organizations: Department of Biostatistics and Bioinformatics, Duke University School of Medicine, Durham, NC, USA · Department of Biostatistics, University of North Carolina at Chapel Hill, Chapel Hill, NC, USA · Department of Biomedical Informatics and Medical Education, University of Washington, WA, USA · Department of Population Health Sciences, Weill Cornell Medicine, New York, NY, USA · Systems Engineering, Cornell University, Ithaca, NY, USA
Patient information can guide electronic health record (EHR) generation, but how its amount shapes predictive utility and disclosure risk remains poorly characterized. Here we present EHRDiB, which unifies visit completion and synthesis by distinguishing original missingness from withheld inputs and combining a shared multimodal representation with an evidence-conditioned latent diffusion bridge. We compared EHRDiB with three completion and eight synthesis methods on MIMIC-IV and numerical and discretized Duke Health representations. Evaluation covered downstream predictive utility, generative diversity, and distributional fidelity for both tasks, and disclosure risk for synthesis. EHRDiB achieved the smallest utility gaps from real-data-trained classifiers in seven of eight completion comparisons. In synthesis, AUPRC gaps were 45.2% smaller for ten MIMIC-IV principal-diagnosis targets and 53.3% smaller for numerical Duke Health data relative to the best comparator. Synthesis with calibrated thresholding yielded the lowest marginal distribution error across all three datasets, with Coverage above 0.95 for both output rules. Across twelve evidence-fraction intervals in MIMIC-IV synthesis, membership-inference AUROC increased from 0.516 to 0.598 as the mean target evidence fraction rose from 0.20 to 0.50; utility gaps narrowed relative to the lowest evidence configuration but did not decrease monotonically. Our study provides a unified framework and empirical foundation for EHR generation that considers both its utility for clinical analysis and its information disclosure risk.
Figures & tables
Figure 1: Overview of EHRDiB. Availability-gated encoders and clinical context fusion map clinical information from data and evidence views into a shared latent space, with presence and values encoded separately for continuous-valued modalities. A latent diffusion bridge uses an evidence-conditioned denoising network to predict the data endpoint from intermediate latent states, guiding the generation of a visit representation from the evidence representation. The starting distribution is perturbed for synthesis. Parallel decoders map the generated representation to all clinical modalities, with two output rules: Bernoulli sampling and calibrated thresholding (Cali-T). Visual elements adapted from Li et al. Li et al. 2023 ; example EHR image from VA Blue Button U.S. Department of Veterans Affairs .
Utility, set 1
Utility, set 2
Diversity
Fidelity ( ×103 )
Model
Δ AUROC
Δ AUPRC
Δ AUROC
Δ AUPRC
Coverage
APDR
FDD
BCMD
MIMIC-IV
CISS-VAE
0.0048
0.0624
0.0334
0.0848
0.9417
0.8296
16.25
13.34
GAIN
0.0027
0.0504
0.0316
0.0787
0.6076
0.9257
21.69
16.76
CSDI
0.0038
0.0396
0.0396
0.0873
0.9711
0.9671
9.70
12.85
EHRDiB
0.0021
0.0319
0.0314
0.0837
0.9938
0.9434
13.17
11.93
Table 1: Completion results across three datasets. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and completed records. Set 1 comprises ten principal-diagnosis targets in MIMIC-IV and ten diagnosis targets in DH and DH-D; set 2 comprises seven Phecode-defined phenotypes in MIMIC-IV. Bold indicates the best result for each metric within each dataset.
Figure 2: MIMIC-IV synthesis utility for the ten principal-diagnosis targets. Macro-averaged ROC (top) and precision-recall (bottom) curves compare LightGBM classifiers trained on real (TRTR) or synthetic records and evaluated on the same real test set. Columns group variational, adversarial, and diffusion comparators, with EHRDiB (Bernoulli) and EHRDiB (Cali-T) shown in each.
Utility, set 1
Utility, set 2
Diversity
Fidelity ( ×103 )
Model
Δ AUROC
Δ AUPRC
Δ AUROC
Δ AUPRC
Coverage
APDR
FDD
BCMD
MIMIC-IV
MVAE
0.0381
0.2310
0.1272
0.1487
0.8116
0.7442
42.89
25.52
MoPoE-VAE
0.0876
0.3479
0.1913
0.2165
0.5956
0.6548
62.38
32.08
medGAN
0.1440
0.4412
0.1831
0.2386
0.0663
4.8867
382.79
38.86
medBGAN
0.0897
0.3755
0.1474
0.2126
0.0890
5.0766
556.94
33.57
Table 2: Synthesis results across three datasets. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and synthetic records. Set 1 comprises ten principal-diagnosis targets in MIMIC-IV and ten diagnosis targets in DH and DH-D; set 2 comprises seven Phecode-defined phenotypes in MIMIC-IV. Bold indicates the best result for each metric within each dataset.
Figure 3: Marginal distribution fidelity of synthetic records in MIMIC-IV (top), DH (middle), and DH-D (bottom). D, diagnosis; P, procedure (MIMIC-IV only); M, medication; C, clinical measurement. Each modality’s FDD is normalized to the lowest value among models within the same dataset. The reversed logarithmic radial scale runs from 1× at the outer ring to 100× at the center; larger radii indicate lower errors. Columns show variational, adversarial, and diffusion baselines alongside EHRDiB under both output rules.
Figure 4: Absolute Pearson-correlation errors between real and synthetic MIMIC-IV records. All models share the same 120 features, with 30 each from diagnosis (D), procedure (P), medication (M), and clinical measurement (C). Darker colors indicate larger errors on a shared scale. This exploratory subset was selected using test-set error contrasts.
MIMIC-IV
DH
DH-D
Model
MIA
Δ AIA
MIA
Δ AIA
MIA
Δ AIA
MVAE
0.5068
-0.0161
0.4983
-0.0379
0.4990
-0.0291
MoPoE-VAE
0.5010
-0.0325
0.4969
-0.0826
0.4920
-0.0774
medGAN
0.4995
-0.0358
0.5017
-0.0414
0.5018
-0.0511
medBGAN
0.5002
-0.0546
0.5013
-0.0289
0.5013
-0.0370
EMR-WGAN
0.5078
-0.0103
0.5028
-0.0062
0.5031
-0.0095
Table 3: Synthesis disclosure risk across three datasets. Membership inference attack (MIA) performance is measured by AUROC, with chance at 0.5. Attribute inference advantage ( Δ AIA) is the difference in attack AUPRC between a synthetic reference and a real auxiliary reference. Bold indicates MIA AUROC closest to 0.5 and the lowest Δ AIA within each dataset.
Figure 5: Disclosure risk and phenotyping utility in MIMIC-IV synthesis across evidence-fraction intervals. Points show individual intervals at their mean evidence fraction; each line connects the minimum observed value at each mean. (a) MIA AUROC; the gray band shows the comparator range and the dashed line marks random discrimination (0.5). (b) AUPRC gaps for set 1 (ten principal-diagnosis targets) and set 2 (seven Phecode-defined phenotypes); dashed lines indicate the best comparator for each set.
MIMIC-IV
Duke Health
Characteristic
Training
Validation
Test
Training
Validation
Test
Patients, n
95,532
11,946
11,947
96,704
12,073
12,079
Visits, n
196,188
24,292
24,650
750,817
92,220
93,434
Age, years
61 (43-74)
61 (43-74)
61 (44-74)
53.9 (43.7-63.8)
53.9 (43.6-63.8)
54.0 (43.9-63.8)
Female, %
52.5
52.9
52.8
56.7
57.0
57.3
Measurement variables present per visit, median (IQR), %
7.8 (5.2-11.3)
7.8 (4.9-11.3)
7.6 (5.0-11.3)
25.0 (25.0-37.5)
25.0 (25.0-37.5)
25.0 (25.0-37.5)
Table S1: Characteristics of the study populations by data partition. Age is reported as median (IQR), estimated at the first included admission for MIMIC-IV and measured at the PREVENT study baseline for Duke Health. Age and sex summarize unique patients; other distributions summarize visits. Measurement presence uses 462 distinct laboratory items in MIMIC-IV and eight clinical indicators in Duke Health. Target prevalence spans ten principal-diagnosis targets in MIMIC-IV and ten diagnosis targets in Duke Health within each partition. DH and DH-D share the same patients and visits.
Fidelity ( ×103 )
Evidence-fraction interval
FDD
BCMD
[0.1,0.3]
24.01
15.43
[0.1,0.4]
22.92
15.05
[0.1,0.5]
22.38
14.72
[0.1,0.6]
23.21
14.74
[0.1,0.7]
21.87
14.20
Table S2: Distributional fidelity of EHRDiB in MIMIC-IV synthesis across evidence-fraction intervals. FDD, feature distribution distance; BCMD, blockwise correlation matrix distance. Bold indicates the lowest value for each metric.
Evidence-fraction interval
APDR
Coverage
[0.1,0.3]
0.9407
0.9759
[0.1,0.4]
0.9471
0.9786
[0.1,0.5]
0.9548
0.9815
[0.1,0.6]
0.9583
0.9865
[0.1,0.7]
0.9621
0.9883
[0.2,0.4]
0.9587
0.9830
Table S3: Diversity of EHRDiB in MIMIC-IV synthesis across evidence-fraction intervals. Coverage measures the proportion of real-data neighborhoods containing a synthetic record in the shared principal-component space; the average pairwise distance ratio (APDR) compares the dispersion of synthetic and real records. Bold indicates the highest Coverage and APDR closest to 1.
Utility, set 1
Utility, set 2
Evidence-fraction interval
Δ AUROC
Δ AUPRC
Δ AUROC
Δ AUPRC
[0.1,0.3]
0.0054
0.0625
0.0426
0.1018
[0.1,0.4]
0.0043
0.0493
0.0396
0.0996
[0.1,0.5]
0.0042
0.0457
0.0410
0.0969
[0.1,0.6]
0.0037
0.0471
0.0391
0.1005
[0.1,0.7]
0.0034
0.0450
0.0382
0.0970
Table S4: Predictive utility of EHRDiB in MIMIC-IV synthesis across evidence-fraction intervals. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and synthetic records. Set 1 comprises ten principal-diagnosis targets; set 2 comprises seven Phecode-defined phenotypes. Bold indicates the smallest utility gap for each metric and target set.
Evidence-fraction interval
MIA
Δ AIA
[0.1,0.3]
0.5162
-0.0036
[0.1,0.4]
0.5218
-0.0028
[0.1,0.5]
0.5336
-0.0049
[0.1,0.6]
0.5467
-0.0024
[0.1,0.7]
0.5694
-0.0035
[0.2,0.4]
0.5273
-0.0023
Table S5: Disclosure risk of EHRDiB in MIMIC-IV synthesis across evidence-fraction intervals. Membership inference attack (MIA) performance is measured by AUROC, with chance at 0.5. Attribute inference advantage ( Δ AIA) is the difference in attack AUPRC between a synthetic reference and a real auxiliary reference. Bold indicates MIA AUROC closest to 0.5 and the lowest Δ AIA.
Fidelity ( ×103 )
Diversity
Utility
Encoding
FDD
BCMD
Coverage
APDR
Δ AUROC
Δ AUPRC
No explicit encoding
23.11
15.19
0.986
0.958
0.006
0.069
Availability concatenation
22.91
15.41
0.985
0.958
0.007
0.074
Availability embedding
23.34
15.18
0.987
0.957
0.006
0.060
Availability gating
22.11
14.94
0.988
0.958
0.006
0.060
Table S6: Availability-encoding results for EHRDiB in MIMIC-IV synthesis. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and synthetic records for the ten principal-diagnosis targets. Bold indicates the best result for each metric before rounding.
Fidelity ( ×103 )
Diversity
Utility
λKL
FDD
BCMD
Coverage
APDR
Δ AUROC
Δ AUPRC
10−1
44.24
29.69
0.113
0.336
0.361
0.584
10−2
27.37
20.36
0.895
0.849
0.084
0.393
10−3
23.26
15.52
0.987
0.942
0.010
0.107
10−4
22.35
15.11
0.989
0.952
0.006
0.068
10−5
22.11
14.94
0.988
0.958
0.006
0.060
Table S7: KL-regularization results for EHRDiB in MIMIC-IV synthesis. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and synthetic records for the ten principal-diagnosis targets. Bold indicates the best result for each metric before rounding.
Fidelity ( ×103 )
Diversity
Utility
Bridge
FDD
BCMD
Coverage
APDR
Δ AUROC
Δ AUPRC
VE
22.06
14.94
0.989
0.959
0.005
0.060
VP
22.11
14.95
0.989
0.969
0.005
0.059
Linear
21.78
14.63
0.990
0.964
0.006
0.051
Schrödinger
20.91
14.37
0.989
0.974
0.005
0.052
Table S8: Forward-bridge results for EHRDiB in MIMIC-IV synthesis. All paths use the same endpoint-prediction framework. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and synthetic records for the ten principal-diagnosis targets. Bold indicates the best result for each metric before rounding.
Fidelity ( ×103 )
Diversity
Utility
Perturbation, stochasticity
FDD
BCMD
Coverage
APDR
Δ AUROC
Δ AUPRC
0, 0
23.22
14.22
0.990
0.921
0.004
0.053
0, 0.125
20.85
14.41
0.988
0.986
0.005
0.049
0, 0.25
20.94
14.35
0.989
0.978
0.005
0.049
0, 0.33
20.93
14.34
0.989
0.976
0.005
0.050
0, 0.5
20.95
14.36
0.989
0.972
0.005
0.048
Table S9: Hybrid SDE-ODE sampling results for EHRDiB in MIMIC-IV synthesis, varying evidence-endpoint perturbation and stochasticity. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and synthetic records for the ten principal-diagnosis targets. Bold indicates the best result for each metric before rounding.
Fidelity ( ×103 )
Diversity
Utility
Perturbation, stochasticity
FDD
BCMD
Coverage
APDR
Δ AUROC
Δ AUPRC
0, 0
25.50
14.56
0.989
0.873
0.005
0.063
0, 0.3
22.81
14.53
0.991
0.919
0.005
0.051
0, 0.5
22.03
14.51
0.991
0.936
0.005
0.050
0, 0.8
21.57
14.47
0.991
0.951
0.005
0.049
0, 1
21.40
14.44
0.990
0.956
0.005
0.052
Table S10: Stochastic-interpolant sampling results for EHRDiB in MIMIC-IV synthesis, varying evidence-endpoint perturbation and stochasticity. Utility gaps are differences in AUROC and AUPRC between LightGBM classifiers trained on real and synthetic records for the ten principal-diagnosis targets. Bold indicates the best result for each metric before rounding.
Table S11: Default training and sampling settings for the main EHRDiB completion and synthesis experiments.
Figure S1: MIMIC-IV synthesis utility for the seven Phecode-defined phenotypes. Macro-averaged ROC (top) and precision-recall (bottom) curves compare LightGBM classifiers trained on real (TRTR) or synthetic records and evaluated on the same real test set. Columns group variational, adversarial, and diffusion comparators, with EHRDiB (Bernoulli) and EHRDiB (Cali-T) shown in each.
Figure S2: MIMIC-IV synthesis ROC curves for the ten principal-diagnosis targets: (a) sepsis; (b) acute myocardial infarction; (c) major depressive disorder; (d) type 2 diabetes mellitus; (e) cerebral infarction; (f) cholelithiasis and cholecystitis; (g) fracture of femur; (h) alcoholic liver disease; (i) malignant neoplasm of bronchus and lung; (j) respiratory failure. Each panel compares LightGBM classifiers trained on real (TRTR) or synthetic records and evaluated on the same real test set. EHRDiB (Bernoulli) and EHRDiB (Cali-T) are shown alongside the comparators. FPR, false positive rate; TPR, true positive rate.
Figure S3: MIMIC-IV synthesis ROC curves for the seven Phecode-defined phenotypes: (a) type 2 diabetes mellitus; (b) heart failure; (c) cerebrovascular disease; (d) bipolar disorder; (e) suicidal ideation; (f) rheumatoid arthritis; (g) multiple sclerosis. Each panel compares LightGBM classifiers trained on real (TRTR) or synthetic records and evaluated on the same real test set. EHRDiB (Bernoulli) and EHRDiB (Cali-T) are shown alongside the comparators. FPR, false positive rate; TPR, true positive rate.
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