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

PRIME-CVD is a freely available, synthetic medical environment that generates two distinct cardiovascular risk datasets from causal graphs and public statistics to enable safe, reproducible education in data cleaning, causal reasoning, and risk modelling without compromising patient privacy.

The Gist

Imagine you are trying to teach a class of future doctors how to predict heart disease. You want them to practice on real patient records so they can learn how to spot patterns, clean up messy data, and build life-saving models.

But here's the problem: You can't use real patient records.

Just like you wouldn't let a student drive a real car with a real family inside to learn how to parallel park, you can't let students play with real medical data. The privacy laws are strict, the data is locked away, and if a student accidentally "leaks" a patient's identity, it's a disaster.

Enter PRIME-CVD.

Think of PRIME-CVD not as a stolen set of real keys, but as a hyper-realistic, fully functional driving simulator built specifically for learning. It's a massive, digital playground where 50,000 "fake" people live, get sick, and age, all without ever having existed in the real world.

Here is how it works, broken down into simple concepts:

1. The "Recipe" Instead of the "Cake"

Most fake data is made by computers looking at real data and trying to copy it (like a photocopier). But if the copier is too good, it might accidentally print a real person's name.

PRIME-CVD is different. The creators didn't copy real patients. Instead, they wrote a recipe (called a "Directed Acyclic Graph" or DAG).

• They took public statistics (like "people in poorer areas smoke more" or "older people have higher blood pressure").
• They fed these rules into a computer.
• The computer then baked 50,000 new people from scratch, following those rules perfectly.

Because these people were invented from a recipe and not copied from a real person, there is zero risk of privacy leaks. You can't find a real person in the data because they never existed.

2. The Two "Levels" of the Game

The paper offers two versions of this dataset, designed like a video game with two difficulty levels:

Level 1: The "Clean" Version (Data Asset 1)

• The Analogy: Imagine a spreadsheet where every column is perfectly labeled, every number is in the right place, and there are no typos.
• The Use: This is for students who want to learn the math of risk modeling. They can jump straight in and say, "Okay, if I run this equation, does it predict heart attacks correctly?" It's the "textbook" version.

Level 2: The "Messy" Version (Data Asset 2)

• The Analogy: Imagine taking that perfect spreadsheet and throwing it into a blender with a real hospital's filing cabinet.
  • One doctor wrote "High BP," another wrote "Systolic 140," and a third wrote "BP: 140/90."
  • Some patient IDs are scrambled numbers.
  • Some dates are missing.
  • Some data is in different units (like inches vs. centimeters).
• The Use: This is the real-world training. Before a student can do the math, they have to learn to be a detective. They have to clean the data, link the scattered tables, and fix the typos. This teaches them the hardest part of medical informatics: Data Cleaning.

3. Why This Matters

In the past, students had to wait years to get access to real data, or they had to work with data that was so "scrubbed" (cleaned) that it didn't look like real life.

PRIME-CVD solves this by giving them:

• Realism: The fake people have realistic relationships (e.g., if they have diabetes, their blood sugar is high; if they are poor, they are more likely to smoke).
• Safety: No one's privacy is at risk.
• Freedom: Teachers can give these datasets to 1,000 students at once for homework or exams without needing special permission or passwords.

The Bottom Line

PRIME-CVD is a safe, open-source sandbox for the next generation of medical data scientists. It lets them practice on a "digital twin" of a population, learning how to handle the messy, confusing reality of hospital records without ever putting a real patient's privacy at risk. It's the difference between reading a book about driving and actually practicing in a simulator before you ever touch a real steering wheel.

Technical Summary

1. Problem Statement

The field of medical informatics and machine learning relies heavily on openly accessible benchmark datasets for teaching and methodological development. However, patient-level Electronic Medical Record (EMR) data is rarely available for educational purposes due to:

• Privacy and Governance: Strict regulations prevent the release of real patient data.
• Re-identification Risks: Even de-identified data can be vulnerable to re-identification, especially when rare clinical combinations exist.
• Administrative Barriers: Accessing resources like the MIMIC database requires credentialing, creating bottlenecks for large-scale classroom use.
• Limitations of Existing Synthetic Data: Many current synthetic datasets are generated using deep learning models (e.g., GANs, Diffusion models) trained on real patient trajectories. These models retain residual membership-inference risks and often fail to capture the structural "messiness" (heterogeneous coding, missingness, unit inconsistencies) inherent in real-world EMRs.

There is a critical gap for a privacy-preserving, realistic, and structurally complex dataset that allows students to practice data cleaning, causal reasoning, and risk modeling without exposing sensitive information.

2. Methodology

The authors introduce PRIME-CVD, a parametrically rendered environment designed to simulate a cohort of 50,000 adults undergoing primary prevention for cardiovascular disease (CVD). Unlike generative AI approaches, PRIME-CVD is constructed de novo using a Directed Acyclic Graph (DAG) parameterized exclusively by publicly available aggregate statistics (e.g., Australian Bureau of Statistics, AIHW) and published epidemiologic effect estimates.

The methodology involves two distinct data assets:

Data Asset 1: The Clean Cohort

• Generation: A fully specified, epidemiologically coherent cohort generated via a parametric DAG.
• Process:
  1. Exogenous Variables: Socioeconomic status (IRSD quintiles) and age are sampled from target population distributions.
  2. Conditional Generation: Behavioral (smoking), anthropometric (BMI), and clinical variables (Diabetes, CKD, AF, HbA1c, SBP, eGFR) are generated conditionally based on parent nodes in the DAG using logistic and linear models calibrated to published odds ratios and means.
  3. Outcome Simulation: A proportional-hazards model simulates 5-year CVD event times, resulting in an incidence rate of ~4%.
• Characteristics: This asset is "analysis-ready," containing clean, harmonized variables suitable for survival modeling and exploratory analysis.

Data Asset 2: The EMR-Style Relational Database

• Transformation: Data Asset 1 is systematically transformed into a relational, multi-table database to mimic the structural and lexical heterogeneity of real primary-care EMRs.
• Structure: The data is split into three linked tables:
  1. PatientMasterSummary: Demographics, socioeconomic position, and coarse-grained outcomes.
  2. PatientChronicDiseases: A long-form table of diagnoses with heterogeneous free-text labels (e.g., "T2DM," "ICD10: E11") and non-informative dates.
  3. PatientMeasAndPath: Long-form measurements (HbA1c, eGFR, SBP) with mixed units (e.g., % vs. mmol/mol for HbA1c) and varied string descriptions.
• Injected "Messiness": The authors introduce controlled artifacts to simulate real-world challenges:
  • Non-sequential, scrambled patient IDs.
  • Patterned missingness (e.g., 15.66% of non-smokers have missing smoking status).
  • Lexical heterogeneity in diagnosis and measurement labels.
  • Unit inconsistencies and temporally diffused dates.

3. Key Contributions

• Parametric Synthesis without Real Data: PRIME-CVD is the first dataset of its scale generated entirely from aggregate statistics and causal graphs, eliminating the risk of membership inference attacks associated with generative models trained on real patient data.
• Dual-Asset Pedagogical Design: It offers a unique two-tiered learning environment:
  • Asset 1 teaches statistical modeling, causal inference, and survival analysis on clean data.
  • Asset 2 teaches data engineering skills, including table linkage, text harmonization, unit standardization, and cohort reconstruction.
• Realistic Subgroup Imbalance: Despite being synthetic, the dataset preserves realistic risk gradients and subgroup imbalances (e.g., higher CVD risk in lower socioeconomic groups) derived from real epidemiologic evidence.
• Open Accessibility: Released under a Creative Commons Attribution 4.0 license, the data and code are freely available on FigShare and GitHub, removing administrative access barriers.

4. Results and Validation

The paper validates the dataset through three representative educational exercises:

1. Cohort Reconstruction: Students successfully linked the three relational tables in Data Asset 2, harmonized heterogeneous diagnosis labels, and reconstructed mutually exclusive cohorts (CKD-only vs. T2DM-only).
2. Socioeconomic Stratification: Analysis of Data Asset 1 confirmed coherent gradients across IRSD quintiles, showing expected patterns where lower socioeconomic status correlates with higher smoking rates, higher BMI, and increased CVD risk.
3. Multivariable Hazard Modeling: A Cox proportional hazards model fitted to Data Asset 1 yielded hazard ratios closely resembling those found in contemporary Australian CVD studies (e.g., Diabetes HR ~4.15, Age HR ~1.03 per year).

Technical Validation:

• The dataset contains 50,000 individuals with a median age of ~49.6 years.
• Prevalence rates match target statistics: Diabetes (7.4%), CKD (0.7%), AF (~0.7%).
• The 5-year CVD event rate is approximately 4%.
• Correlation structures align with the causal DAG, demonstrating that only explicitly encoded relationships manifest as correlations, ensuring transparency in the data generation process.

5. Significance

PRIME-CVD addresses a fundamental barrier in health data science education: the trade-off between privacy and analytic realism.

• Educational Impact: It enables scalable, hands-on training in data cleaning, causal reasoning, and policy-relevant risk modeling without the need for ethics approval or restrictive data governance agreements.
• Methodological Transparency: By using a parametric DAG rather than a black-box generative model, the dataset offers full transparency regarding the causal assumptions and parameter sources, making it ideal for teaching causal inference.
• Policy Relevance: The dataset supports the development of fair and equitable risk prediction models by allowing learners to explore socioeconomic disparities and subgroup risks that are often suppressed in real-world de-identified datasets to prevent re-identification.

In summary, PRIME-CVD provides a reproducible, safe, and pedagogically robust environment for training the next generation of medical informaticians in cardiovascular risk modeling.