Planned Future Activities

Synthetic Data: Examples – Realistic – using AI (SYNDERAI), pronounced /ˈsɪn.də.raɪ/

© HL7 Europe | Main Contributor: Dr. Kai U. Heitmann | Privacy Policy • LGPL-3.0 license

Even better Synthetic Funds — a Problem Description

Analyzing the list of conditions and their frequency generated by Synthea from Mitre that SYNDERAI used from July 2025 on to create the first set of Synthtic data (Packages version 1.0+ and 2.0+), we can observe that the dirtibution is not comaprable to typical population based diseases.

Future data generation processes need to reflect public health conditions of populations in Europe much better than the synthetic data funds used until April 2026. The updated funds used from May 2026 on will be reflected in SYNDEAI packages version 3.0+.

Analysis of the Distribution Problems

1. Administrative & Social Findings Dominate Clinical Reality

The top entries are not clinical diagnoses at all — they are social/administrative findings:

Together these consume ~57% of all records, crowding out clinically meaningful disease burden. These reflect US administrative coding and SDOH (Social Determinants of Health) frameworks that don't map to European health information systems.

2. Major European Burden Diseases Are Severely Underrepresented

Comparing the output to authoritative European sources (WHO Europe, Eurostat, ECDC):

3. US Opioid Crisis Artifacts

Drug overdose appears at 0.47% — a distinctly American epidemiological signature. European drug-related morbidity patterns differ substantially by country and substance.

4. Cancer Is Nearly Absent

Cancer is the second leading cause of death in Europe (after cardiovascular disease), yet it barely registers in the dataset — reflecting Synthea's default US-centric module coverage.

How to Modify Synthea for European Populations

Synthea is modular and highly configurable. Here are the concrete levers to pull:

1. Recalibrate Disease Module Prevalence Parameters

Every Synthea disease is defined in a JSON module under src/main/resources/modules/. Each module contains incidence and prevalence probability tables, often stratified by age and sex. These need to be re-parameterized using European data sources:

Target data sources:

• Eurostat (Health Statistics) — eurostat.ec.europa.eu
• ECDC (European Centre for Disease Prevention and Control)
• WHO Europe Global Health Observatory
• Global Burden of Disease (IHME) — EU-specific country profiles
• National health registries (e.g., NHS Digital for England, RKI for Germany, INSEE for France)

Modules to prioritize for recalibration:

2. Suppress or Reweight US-Centric SDOH Modules

The social_determinants_of_health.json and related modules encode very US-specific SDOH concepts. Options:

• Disable the module entirely by removing it from the module path or setting all transition probabilities to 0 for non-applicable codes.
• Replace SDOH coding with European equivalents — for example, using ICPC-2 or ICD-10-CM equivalents used in EU primary care systems.
• Reduce the "Medication review due" entry — this is generated at extremely high frequency by the care_goals.json or similar administrative modules. Cap it or disable it unless you need it for specific use cases.

3. Adjust the Demographic Input Files

Synthea uses demographic CSV files to drive population age/sex/geographic distributions. SYNDERAI replaced the US Census–based defaults with European equivalents (see Principles).

The European base in SYNDERAI need to get tweaked towards data derived from Eurostat population projections or national census data for your target country/region. This affects:

• Age pyramid (Europe is generally older than the US default)
• Sex ratios
• Urban/rural splits

An older age pyramid will automatically increase the prevalence of age-associated diseases like dementia, COPD, and cardiovascular disease even before you touch the clinical modules.

4. Modify the keep_patients and Simulation Configuration

You can also define the simulation to run against specific age cohorts to over-represent elderly populations (important for EU aging burden).

5. Create New Modules for Europe-Specific Conditions

Some conditions common in Europe may have no Synthea module at all. You'll need to author new JSON modules for:

• Tick-borne encephalitis (Central/Eastern Europe)
• Mediterranean diet–related conditions (Southern Europe)
• Country-specific screening programs (e.g., cervical cancer screening timelines differ significantly from the US)
• Occupational diseases (relevant in industrial regions)
• Alcohol use disorder — recalibrate with WHO Europe alcohol consumption data, which varies enormously by country

6. Recalibrate Mortality and Life Tables

Synthea uses US life tables for mortality. Replace with Eurostat life tables per country:

src/main/resources/cdc_growth_charts.json  ← replace mortality assumptions

European life expectancy patterns and cause-of-death distributions differ from the US — cardiovascular disease dominates earlier, while cancer patterns differ by region.

7. Validate Against Reference Prevalence Targets

After recalibration, validate the output distribution by computing the Kullback-Leibler divergence or a simpler chi-square goodness-of-fit test against your target European prevalence table. You can automate this as part of your data generation pipeline:

python

from scipy.stats import chisquare
import pandas as pd

# expected = European prevalence rates (from WHO/Eurostat)
# observed = Synthea output frequencies
stat, p = chisquare(f_obs=observed_counts, f_exp=expected_counts)

Iterate on module parameters until the KL divergence is minimized across the top 20–30 conditions by burden.

Priority Actions Summary

From April 2026 on the SYNDERAI synthetic funds will be re-generated with specific recalibrated JSON modules for a high-priority condition (e.g., hypertension or depression), or draft a validation script to measure distribution fit against a European reference dataset.

On the way to a solution

We created both deliverables: a recalibrated European Hypertension Synthea module (here) and a Python validation script (here) with statistical fit testing against WHO/Eurostat reference data.

Deliverables

hypertension_europe.json — A drop-in Synthea module replacing the US default. Key changes from the US original: age-stratified onset rates from EHIS 2019 (ranging from 0.45%/yr at age 18–34 to 1.7%/yr at 55–64), ESC/ESH-preferred medications (ramipril → losartan → amlodipine instead of lisinopril/HCTZ), European BP control targets (<130/80), and complication rates calibrated against GBD 2019 Europe. Install by placing it under src/main/resources/modules/.

validate_synthea_europe.py — Run against any conditions.csv with:

python3 validate_synthea_europe.py --input conditions.csv --population 100000 --output results.csv --charts

What the validation found on your actual data

The numbers are striking — only 1 out of 38 reference conditions falls within the acceptable ±20% band:

The three sharpest failure modes surfacing from your data:

1. Respiratory infections are wildly over-represented — Viral sinusitis at 9.78% vs an EU reference of 0.8% (+1,122%), and acute pharyngitis at 5.51% vs 0.7% (+687%). These are acute, self-limiting conditions that accumulate many records per patient lifetime in Synthea's default encounter model, inflating them far beyond population prevalence.

2. Eight conditions are completely absent — COPD, Depression, Anxiety disorder, Osteoarthritis, Osteoporosis, Prostate cancer, Dementia, and Parkinson's disease generate zero records. These represent some of the largest sources of disability-adjusted life years in Europe.

3. Metabolic and cardiovascular burden is severely compressed — Hyperlipidaemia sits at 1.64% vs a 22% EU reference (−93%), Metabolic syndrome at 2.7% vs 25% (−89%), and Hypertension at 4.78% vs 32% (−85%). These are the foundational chronic conditions that drive most downstream morbidity in any European cohort.

The hypertension_europe.json module directly corrects the hypertension gap. The same calibration logic can be applied to the other missing/under-sampled modules following the same pattern.