Skip to content

Library and CLI for randomly generating medical data like you might get out of an Electronic Health Records (EHR) system

License

Notifications You must be signed in to change notification settings

HicServices/SynthEHR

Repository files navigation

SynthEHR (Previously BadMedicine)

Build Status NuGet Badge

Library and CLI for randomly generating medical data like you might get out of an Electronic Health Records (EHR) system. It is intended for generating data for demos and testing ETL / cohort generation/ data management tools.

SynthEHR differs from other random data generators e.g. Mockaroo, SQL Data Generator etc in that data generated is based on (simple) models generated from live EHR datasets collected for over 30 years in Tayside and Fife (UK). This makes the data generated recognisable (codes used, frequency of codes etc) from a clinical perspective and representative of the problems (ontology mapping etc) that data analysts would encounter working with real medical data.

Datasets generated are not suitable for training AI algorithms etc (See What is Modelled?)

Rename

As of v2.0.0 BadMedicine was renamed to SynthEHR. Previous versions of the software can be found at nuget.org.

Datasets

The following synthetic datasets can be produced.

Dataset Description
Demography Address and patient details as might appear in the CHI register
Biochemistry Lab test codes as might appear in Sci Store lab system extracts
Prescribing Prescription data of prescribed drugs
Carotid Artery Scan Scan results for Carotid Artery
Hospital Admissions ICD9 and ICD10 codes for admission to hospital
Maternity Records of births etc

Usage:

SynthEHR is available as a nuget package for linking as a library

The standalone CLI (SynthEHR.exe) is available in the releases section of Github

Usage is as follows:

SynthEHR.exe c:\temp\

You can change how much data is produced (e.g. 500 patients, 10000 records per dataset):

SynthEHR.exe c:\temp\ 500 10000

Or run only a single dataset:

SynthEHR.exe c:\omg 5000 200000 -l -d CarotidArteryScan

You can seed the generator (Guids generated will still differ)

SynthEHR.exe c:\omg 5000 200000 -l -d CarotidArteryScan -s 5000

Building

Building requires MSBuild 15 or later (or Visual Studio 2017 or later). You will also need to install the DotNetCore 2.2 SDK.

You can build a OS specific binary

First build SynthEHR.csproj

dotnet publish SynthEHR.csproj -r win-x64 --self-contained
cd .\bin\Debug\netcoreapp2.2\win-x64\

Direct to Database

You can generate data directly into a relational database (instead of onto disk).

To turn this mode on rename the file SynthEHR.template.yaml to SynthEHR.yaml and provide the connection strings to your database e.g.:

Database:
  # Set to true to drop and recreate tables described in the Template
  DropTables: false
  # The connection string to your database
  ConnectionString: server=(localdb)\MSSQLLocalDB;Integrated Security=true;
  # Your DBMS provider ('MySql', 'PostgreSql','Oracle' or 'MicrosoftSQLServer')
  DatabaseType: MicrosoftSQLServer
  # Database to create/use on the server
  DatabaseName: SynthEHRTestData

Library Usage

You can generate test data for your program yourself by referencing the nuget package:

//Seed the random generator if you want to always produce the same randomisation
var r = new Random(100);

//Create a new person
var person = new Person(r);

//Create test data for that person
var a = new HospitalAdmissionsRecord(person,person.DateOfBirth,r);

Assert.IsNotNull(a.Person.CHI);
Assert.IsNotNull(a.Person.DateOfBirth);
Assert.IsNotNull(a.Person.Address.Line1);
Assert.IsNotNull(a.Person.Address.Postcode);
Assert.IsNotNull(a.AdmissionDate);
Assert.IsNotNull(a.DischargeDate);
Assert.IsNotNull(a.Condition1);

What is Modelled?

Data generated by SynthEHR is driven by Aggregate distributions of real health data collected in Tayside (UK). This means that codes appear in data with the frequency that match real data. For example in the Hospital Admissions data we can see that ICD9 codes (denoted by dash) cease being recorded in ~1997 in favour of ICD10 codes and we can see the most common admission conditions are sensible:

alt text

ICD 9 and ICD 10 codes in Condition1 (the main condition) upon Hospital Admission

What is not Modelled?

No inter dataset / inter record level randomisation model exists. For example the following would not be modelled:

  • If a patient is on Drug A they are more likely to also be on Drug B
  • Hospitalisations are more likely to be at the beginning/end of a patients life
  • Drug A is likely to be given to patients discharged having been treated for condition Y