PRIME-CVD: A Parametrically Rendered Informatics Medical Environment for Education in Cardiovascular Risk Modelling

Nicholas I-Hsien Kuo; Marzia Hoque Tania; Blanca Gallego; Louisa Jorm

cs.LGMar 11, 2026

PRIME-CVD: A Parametrically Rendered Informatics Medical Environment for Education in Cardiovascular Risk Modelling

Authors: Nicholas I-Hsien Kuo, Marzia Hoque Tania, Blanca Gallego, Louisa Jorm

Organizations: Centre for Big Data Research in Health (CBDRH), The University of New South Wales, Sydney, Australia

Abstract

In recent years, progress in medical informatics and machine learning has been accelerated by the availability of openly accessible benchmark datasets. However, patient-level electronic medical record (EMR) data are rarely available for teaching or methodological development due to privacy, governance, and re-identification risks. This has limited reproducibility, transparency, and hands-on training in cardiovascular risk modelling. Here we introduce PRIME-CVD, a parametrically rendered informatics medical environment designed explicitly for medical education. PRIME-CVD comprises two openly accessible synthetic data assets representing a cohort of 50,000 adults undergoing primary prevention for cardiovascular disease. The datasets are generated entirely from a user-specified causal directed acyclic graph parameterised using publicly available Australian population statistics and published epidemiologic effect estimates, rather than from patient-level EMR data or trained generative models. Data Asset 1 provides a clean, analysis-ready cohort suitable for exploratory analysis, stratification, and survival modelling, while Data Asset 2 restructures the same cohort into a relational, EMR-style database with realistic structural and lexical heterogeneity. Together, these assets enable instruction in data cleaning, harmonisation, causal reasoning, and policy-relevant risk modelling without exposing sensitive information. Because all individuals and events are generated de novo, PRIME-CVD preserves realistic subgroup imbalance and risk gradients while ensuring negligible disclosure risk. PRIME-CVD is released under a Creative Commons Attribution 4.0 licence to support reproducible research and scalable medical education.

Figures & tables

Synthetic Cohort Characteristics (N = 50,000)
Age (years)	Mean (SD): 49.71 (12.37)
	Median [IQR]: 49.63 [41.33, 58.09]
	Range: 18.0–90.0
IRSD Quintile Distribution	Q1: 21.28% (Most disadvantaged)
	Q2: 16.11%
	Q3: 23.88%

Table 1: Baseline characteristics of the PRIME-CVD cohort used for cardiovascular risk simulation.

Figure 1: Baseline characteristics of the PRIME-CVD cohort used for cardiovascular risk simulation.

Figure 2: Directed acyclic graph representing the causal structure embedded in PRIME-CVD. Circular and rectangular nodes denote numeric and categorical variables, respectively; blue, orange, and red represent demographic/lifestyle determinants, chronic diseases, and anthropometric/physiological measurements.

Table	Column Name	Notes
PatientMasterSummary	Patient_ID	Patient identifier: Unique, non-sequential integer key derived from the cohort row index of Data Asset 1.

PatientMasterSummary	Age_At_2024	Approximate patient age in 2024: Continuous, two-decimal precision, of baseline simulated age exacted at 2024.

PatientMasterSummary	IRSD_Quintile	Australian socioeconomic disadvantage quintile: Integer in {1,…,5}, where 1 = most disadvantaged and 5 = least disadvantaged.

Table 2: Unified data dictionary for the PRIME-CVD relational dataset [PatientEMR] .

Parent → Child	Notes	References
IRSD $\rightarrow$ BMI	Higher deprivation slightly increases BMI.	[ 31 ] † [ 32 ] †

IRSD $\rightarrow$ Smoking	More disadvantaged areas have higher smoking rates.	[ 33 ] † [ 34 ] ⋆

BMI $\rightarrow$ Diabetes	Higher BMI increases diabetes risk.	[ 35 ] † [ 36 ] ⋆

Table 3: Simplified Parent–Child Relationships in the PRIME-CVD DAG

Figure 3: Constructing the PRIME-CVD EMR-style Data Asset 2. The clean DAG-generated cohort is split into three relational tables and augmented with realistic EMR artefacts including, missingness, ID scrambling, heterogeneous terminology, and mixed units.

IRSD	Age	Smoking	BMI	Diabetes	CKD	HbA1c	eGFR	SBP	AF	CVD_Event	CVD_Time
4	50.40	non	28.17	0	0	4.32	83.08	118.67	0	1	2.96
3	39.23	ex	16.83	0	0	4.70	81.49	123.72	0	0	4.41
5	55.49	non	23.52	0	0	3.67	86.78	126.65	0	0	4.49
4	51.91	non	31.98	0	0	4.49	94.70	113.89	0	0	4.89
1	47.09	ex	25.35	0	0	4.32	86.34	125.91	0	0	4.49

Table 4: First 5 rows from Data Asset 1.

Patient_ID	Age_At_2024	Smoking	IRSD	CVD_Event	CVD_Time
269	57.40	non	4	1	2019-12
272	46.23	ex	3	0	2022-12
281	62.49	non	5	0	2022-12
296	58.91	non	4	0	2022-12
317	54.09	ex	1	0	2022-12

Table 5: First 5 rows from Data Asset 2’s PatientMasterSummary.csv .

Patient_ID	Category	Date
1338163469	Diabetes	2012-05
150591944	Diabetes	2014-12
545940569	ICD10: E11	2016-10
7134075944	Diabetes	2013-02
5959722392	Diabetes	2014-09

Table 6: First 5 rows from Data Asset 2’s PatientChronicDiseases.csv .

Patient_ID	Measure	Value	Description	Date	Unit
5386159421	eGFR	77.24	eGFR	2014-12	mL/min/1.73m 2
5790852944	HbA1c	4.50	HBA1C	2013-02	%
3141314912	HbA1c	4.16	Hb A1c	2016-09	%
5573606096	eGFR	89.52	eGFR	2013-10	mL/min/1.73m 2
3311104121	SBP	131.55	SBP	2013-11	mmHg

Table 7: First 5 rows from Data Asset 2’s PatientMeasAndPath.csv .

Figure 4: Socioeconomic distributions for mutually exclusive CKD and T2DM cohorts reconstructed from the relational Data Asset 2. Bar plots summarise the prevalence of IRSD quintiles within each cohort following linkage and harmonisation of diagnosis records, enabling comparison of socioeconomic profiles across conditions.

IRSD	N	Age median [IQR]	Diabetes (%)	CKD (%)	AF (%)	BMI mean (SD)	SBP mean (SD)	eGFR mean (SD)	HbA1c mean (SD)	Event (%)
1	10,640	49.53 [41.3–58.09]	8.44	0.77	0.81	29.50 (4.95)	124.98 (16.12)	82.72 (6.22)	4.81 (0.97)	4.28

2	8,055	49.69 [41.44–58.26]	7.78	0.72	0.72	28.99 (4.96)	124.23 (16.11)	82.75 (6.10)	4.80 (0.92)	4.38

3	11,941	49.54 [41.35–58.00]	7.53	0.80	0.69	28.39 (4.95)	123.32 (16.21)	82.76 (6.14)	4.79 (0.93)	4.34

Table 8: Characteristics of Data Asset 1 stratified by IRSD quintile.

Figure 5: IRSD-stratified distributions of variables in Data Asset 1.

Covariate	HR	95% CI	p -value
Age (per year)	1.03	[1.03, 1.03]	$<0.001$
Atrial fibrillation	2.90	[2.24, 3.75]	$<0.001$
Chronic kidney disease	0.97	[0.75, 1.26]	0.84
Diabetes mellitus	4.15	[3.77, 4.57]	$<0.001$
HbA1c (%)	1.37	[1.33, 1.41]	$<0.001$
Body mass index	1.01	[1.01, 1.02]	$<0.001$

Table 9: Hazard ratios from a multivariable Cox proportional hazards model fitted to Data Asset 1.

Figure 6: Hazard ratios derived from a Cox model fitted to Data Asset 1.

Abbreviation	Definition
ABS	Australian Bureau of Statistics
AF	Atrial Fibrillation
AIHW	Australian Institute of Health and Welfare
BMI	Body Mass Index
CKD	Chronic Kidney Disease
CKM	Cardio–Kidney–Metabolic (domain including CVD, diabetes, and CKD)

Table 16

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: Directed acyclic graph representing the causal structure embedded in PRIME-CVD: circular nodes: numeric vairables; rectangular nodes: binary/categorical variables; blue: demographic/lifestyle determinants; orange: chronic diseases; red: anthropometric/physiological measurements.

From Node	To Node	Notes
IRSD	SmokingStatus	– Each IRSD quintile uses a different probability vector over ["non", "current", "ex"] . – Higher deprivation (lower IRSD) has higher probability of being a current smoker. – Implemented in sample_smoking_given_irsd() via rng.choice .

IRSD	BMI	– BMI mean shifts linearly with $(Q_{\text{center}}-q)$ , slope $\Delta_{\mathrm{BMI}}=0.6$ . – Central mean chosen so IRSD-weighted mean equals population BMI mean (28.29). – Sampled from Normal $(\mu_{q},\,\mathrm{SD}=4.98)$ with truncation to [15, 60].

Age	Diabetes	– Diabetes log-odds uses $(\text{Age}-\text{AGE\_MEAN})/10$ . – Age coefficient $\log(1.8)$ encodes OR 1.8 per 10 years. – Older age increases Bernoulli probability in sample_diabetes_given_age_bmi() .

Appendix

Table 10: PRIME-CVD DAG Edges – Part 1

From Node	To Node	Notes
CKD	eGFR	– CKD adds fixed shift $\Delta_{\mathrm{CKD}}=-31$ . – Strong downward deviation in filtration rate for CKD. – Final eGFR $\sim$ Normal(mean, SD=5).

Age	SBP	– SBP mean includes age-centered term. – Coefficient $0.38$ mmHg per year ( $\approx 3.8$ per decade). – Older age increases expected SBP.

SmokingStatus	SBP	– Ex-smokers: $+1.7$ mmHg; Current smokers: $+3.8$ mmHg. – Added through shift_smoke vector. – Non-smokers are reference with shift=0.

Appendix

Table 11: PRIME-CVD DAG Edges – Part 2

Figure 8: Constructing the PRIME-CVD EMR-style data asset. The clean DAG-generated cohort is split into three relational tables (red) and augmented with realistic EMR artefacts (grey), including missingness, ID scrambling, heterogeneous terminology, and mixed units to create Data Asset 2.

Target	Sampling Rules	Description of Injected Heterogeneity
Smoking Status Missingness	Bernoulli( $p=0.1566$ ) applied to non-smokers	15.66% of patients coded as “non” have their value set to NaN , producing patterned lifestyle missingness consistent with real GP EMRs. No missingness is applied to “ex” or “current” smokers.

Diabetes Terms	Multinomial over 6 labels $[0.40,0.20,0.15,$ $0.15,0.05,0.05]$	Six heterogeneous term ( ‘‘Diabetes’’ , ‘‘T2DM’’ , ICD10/ICD9 strings , lay terms) are sampled for each diabetic patient. Ensures realistic variation in coding systems across time and clinicians.

AF Terms	Multinomial over 6 labels $[0.45,0.30,0.075,$ $0.075,0.05,0.05]$	Blends abbreviations, long-form names, and ICD codes ( e.g., ‘‘AF’’ , ‘‘Atrial fibrillation’’ , ‘‘A-fib’’ ) to generate lexical diversity typical of EMR problem lists.

Appendix

Table 12: Sampling distributions governing injected messiness in the relational data asset.

Location	Irregularity	Notes
Patient_ID	Non-sequential identifier	– Computed as $(i^{2}-77)\times 3+500$ to obscure ordering. – Mirrors de-identified EMR IDs with no intuitive structure. – Keeps deterministic linkage across tables.

Age_At_2024	Shifted age reference year	– Age stored as “Age at 2024”, computed as Age + 7. – Reflects EMRs where age corresponds to extraction year, not event time. – Adds mild temporal misalignment relative to baseline.

SMOKING_STATUS	Patterned missingness	– Starts from true smoking status in clean cohort. – 15.66% of non-smokers set to NaN .

Appendix

Table 13: Injected Irregularities in [PatientEMR].[MasterSummary]

Location	Irregularity	Notes
Patient_ID	Presence-only condition table	– Contains rows only for patients with diabetes, CKD, or AF. – Same nonlinear ID transformation used as in other tables.

Category (Diabetes)	Diabetes label variability	– Canonical “Diabetes” replaced by multinomial labels. – Includes ICD codes, abbreviations, and lay labels. – Introduces semantic variability requiring harmonisation.

Category (AF)	AF label variability	– Similar randomness to diabetes labels.

Appendix

Table 14: Injected Irregularities in [PatientEMR].[ChronicDiseases]

Location	Irregularity	Notes
Patient_ID		– Same nonlinear ID transformation used as in other tables.

Measure	Stacked long-form biomarkers	– HbA1c, eGFR, SBP concatenated into one table. – Requires pivoting or filtering to build analytic datasets.

Description (HbA1c)	Heterogeneous labels	– Includes “HbA1c”, “HBA1C”, “A1C”, “Glycated hemoglobin”, etc. – Probability-weighted distribution.

Appendix

Table 15: Injected Irregularities in [PatientEMR].[MeasAndPath]

Figure 9: IRSD-stratified distributions of variables in the PRIME-CVD cohort.

Figure 10: Lower-triangular Pearson correlation matrix for key variables in the PRIME-CVD cohort.

Figure 11: Age-stratified distributions of variables in the PRIME-CVD cohort.

Age	N	IRSD (%)	Diabetes (%)	CKD (%)	AF (%)	BMI mean (SD)	SBP mean (SD)	eGFR mean (SD)	HbA1c mean (SD)	Event (%)
N	2,878	1: 22.03 2: 15.71 3: 23.00 4: 17.23 5: 22.03	1.74	0.28	0.10	28.28 (4.97)	113.41 (14.86)	87.39 (5.28)	4.62 (0.70)	0.21

A	7,976	1: 20.84 2: 16.12 3: 23.92 4: 16.76 5: 22.35	2.95	0.29	0.18	28.34 (5.11)	117.82 (15.24)	85.31 (5.33)	4.67 (0.75)	0.56

B	14,746	1: 21.66 2: 16.00 3: 24.40 4: 16.64 5: 21.31	4.90	0.41	0.49	28.36 (5.02)	121.42 (15.17)	83.71 (5.41)	4.72 (0.82)	1.49

Appendix

Table 16: Characteristics of the PRIME-CVD cohort stratified by age group (N: 18–29, A: 30–39, B: 40–49, C: 50–59, D: 60–74, K:

\geq

75).

Figure 12: Suggested Reconstruction Pipeline

Medical Education

See all 8 papers

Synthetic but Not Realistic: The Evaluation Challenge in Generative Modelling for Structured Electronic Medical Records
Jun 8, 2026Nicholas I-Hsien Kuo, Blanca Gallego, Louisa JormSynthetic Tabular Data Generation Electronic Health Records
Centre for Big Data Research in Health, the University of New South Wales, Sydney, Australia
LLMs for Cardiovascular Risk Prediction from Structured Clinical Data
Apr 21, 2026Jeba Maliha, Md Rafiul KabirMedical LLMs Clinical Risk Prediction
Central Michigan University, Mount Pleasant, MI, USA
Interpretable Synthetic Medical Tabular Data Generation for Clinical Decision Support Using Fuzzy Cognitive Maps
Sep 30, 2026Michael Vasilakakis, Dimitris K. IakovidisSynthetic Data Generation Synthetic Tabular Data Generation
Department of Computer Science and Biomedical Informatics, University of Thessaly, Lamia, Greece

Jun 8, 2026J/K move · Enter open · S save

Synthetic but Not Realistic: The Evaluation Challenge in Generative Modelling for Structured Electronic Medical Records

Nicholas I-Hsien Kuo, Blanca Gallego, Louisa Jorm

Centre for Big Data Research in Health, the University of New South Wales, Sydney, Australia

Synthetic Tabular Data Generation Electronic Health Records Synthetic Data Evaluation Synthetic Clinical Data Generation

Open paper AskPDF arXiv

Synthetic healthcare data are widely proposed as privacy-preserving substitutes for real patient data, yet their evaluation remains dominated by statistical similarity and predictive performance that do not reflect clinical validity. We introduce a multi-dimensional evaluation framework grounded in epidemiology, assessing descriptive fidelity, clinical utility, and structural validity, corresponding to descriptive, predictive, and causal questions. We evaluate four representative generative paradigms - GAN-based, VAE-boosted, diffusion-based, and masked modelling - using PRIME-CVD, a 50,000-person cohort with known ground-truth structure. While all models reproduce marginal distributions, none simultaneously preserve subgroup structure, effect estimates, and dependency structure. Notably, models with strong distributional fidelity can exhibit poor calibration and distorted relationships, leading to unreliable inference. These results show that current evaluation practices can overestimate synthetic data quality and motivate domain-informed assessment based on the ability to support valid clinical and scientific conclusions.

Abstract

Medical Education

Explore similar work

Synthetic but Not Realistic: The Evaluation Challenge in Generative Modelling for Structured Electronic Medical Records

LLMs for Cardiovascular Risk Prediction from Structured Clinical Data

Interpretable Synthetic Medical Tabular Data Generation for Clinical Decision Support Using Fuzzy Cognitive Maps