OFH Synthetic Data Generator

This repository introduces the ofhsyn R library that creates synthetic datasets, aiming to mimic the structure of multiple linked health data sources (participant, questionnaire, clinic measures, HES-like events, primary care medicines, deaths, and geography) available in the Our Future Health cohort. It is designed so researchers can develop and test workflows before running them on the much larger real data.

This project is designed to generate synthetic data from rule-based and stochastic generators, schema dictionaries, and user-supplied code lists, rather than from participant-level OFH source records.

The output data is a general approximation of the schema intended for code development, not for statistical inference. It applies rule-based data generation based on public schemas and does not take any kind of real partcipant data as input.

This is an unofficial tool and is not affiliated with Our Future Health.

Note on Official Tooling: In June 2026, OFH released their official Python-based Synthetic Data Generator (currently in Beta), which generates participant and questionnaire tables. These will be much more representative than ofhsyn, and it is highly recommended to use their official resource where possible. However, ofhsyn can still support code development by providing R-native schemas across a broad range of data entities (such as geography, clinic measures, deaths, and HES/primary care events) that are not yet covered by the official Beta.

The generator produces CSV files for a synthetic cohort, including:

• participant data
• questionnaire data
• clinic measurements
• outpatient, inpatient, emergency, and deaths HES data
• primary care medication data
• country/region data

All datasets are linked by pid.

The sample size, diagnostic codes, and BNF drug codes are determined by the user. These codes are then sampled and distributed using configurable probabilities (for cohort coverage, record counts, and field-level missingness) across the inpatient, outpatient, A&E, and primary care medicines datasets.

Minimum:

• R 4.2+

install.packages('ofhsyn')

install.packages('ofhsyn')
library(ofhsyn)

syn <- OFHCohortSynthesizer$new(project_root = ".", seed = 123)

out <- syn$run_all(n = 1000)

You control which codes appear in outputs, and how frequently records are generated.

• Set code pools with icd10, opcs4, and bnf_codes (or *_file inputs).
• Set dataset-level coverage and record volume with proportions and record_multipliers.
  • E.g., setting proportions = list(nhse_outpat = 0.25) and record_multipliers = list(nhse_outpat = 1.2) for n = 1000 gives about 250 unique outpatient participants and about 1,200 outpatient rows (including repeat visits).
• Set within-dataset code proportionality (relative code frequencies) with code_config weights, for example icd10_weights, opcs4_weights, primary_icd10_weights, underlying_icd10_weights, ae_specific_weights, and read_weights.
  • E.g., setting code_config = list(nhse_outpat_data = list(icd10_weights = c(I210 = 5, I500 = 1))) gives roughly a 5:1 ratio of I210 to I500 among sampled outpatient ICD-10 codes (subject to normal random variation).

ICD10 code files should use this structure (CSV columns, or tab-separated TXT with same columns):

OPCS4 code files should use this structure (CSV columns, or tab-separated TXT with same columns):

BNF codes should use this structure:

library(ofhsyn)

out <- generate_ofh_cohort(
	n = 1000,
	seed = 123,
	icd10 = c(
		I210 = "STEMI of anterolateral wall",
		I500 = "Congestive heart failure"
	),
	opcs4 = c(
		K401 = "Percutaneous transluminal balloon angioplasty of coronary artery"
	),
	bnf_codes = data.frame(
		BNFCode = c("0212000B0", "0601023A0"),
		BNFName = c("Atorvastatin 20 mg tablets", "Metformin 500 mg tablets"),
		Formulation = c("tablets", "tablets"),
		Strength = c("20 mg", "500 mg"),
		stringsAsFactors = FALSE
	)
)

# or load code lists from files
# - ICD10/OPCS4: include both code and description
#   - TXT: tab-separated columns (code\tdescription)
#   - CSV: 'code,description' columns
# - BNF: use a CSV with columns BNFCode, BNFName, Formulation (optional Strength)
out <- generate_ofh_cohort(
	n = 1000,
	seed = 123,
	icd10_file = "icd10_codes.txt",
	opcs4_file = "opcs4_codes.csv",
	bnf_codes_file = "bnf_medications.csv"
)

# return objects only in the R environment (no CSV files written)
out_objects_only <- generate_ofh_cohort(
	n = 1000,
	seed = 123,
	save_csv = FALSE,
	return_objects = TRUE
)

# write CSV files only (no returned R objects)
generate_ofh_cohort(
	n = 1000,
	seed = 123,
	save_csv = TRUE,
	return_objects = FALSE,
	output_dir = "example"
)

# customize generation probabilities
out_custom_probs <- generate_ofh_cohort(
	n = 1000,
	seed = 123,
	proportions = list(
		nhse_outpat = 0.25,
		nhse_inpat = 0.20,
		nhse_ed = 0.30,
		nhse_primcare_meds = 0.75
	),
	record_multipliers = list(
		nhse_outpat = 1.2,
		nhse_inpat = 1.1,
		nhse_ed = 1.3
	),
	code_config = list(
		nhse_outpat_data = list(diag_4_02_missing_prob = 0.70),
		nhse_inpat_data = list(single_diag_prob = 0.85)
	),
	save_csv = FALSE,
	return_objects = TRUE
)