CPRD has generated a number of synthetic datasets that can be used for training purposes or to improve algorithms or machine learning workflows.
Two high-fidelity synthetic datasets are being made available with a nominal administrative fee:
- CPRD cardiovascular disease synthetic dataset
- CPRD COVID-19 symptoms and risk factors synthetic dataset
An additional fee will apply for an annual teaching licence for the high-fidelity synthetic datasets.
CPRD has also developed the CPRD Aurum sample dataset, a medium-fidelity synthetic dataset that resembles the real world CPRD Aurum. Pricing information is available from firstname.lastname@example.org.
The three synthetic datasets will need a data sharing agreement (DSA) with the applicant’s organisation for access in line with advice received from the Information Commissioner’s Office (ICO) Innovation Hub in response to a formal query by the MHRA.
For access to these datasets please submit the CPRD Synthetic data access request form to email@example.com including ‘Synthetic data access request’ in the email subject header. Applicants from organisations that are not existing CPRD clients will also need to submit a new client request form.
Existing multi-study licence (MSL) clients will not need a separate DSA to access the CPRD Aurum sample dataset as this will be added to their MSL agreement. MSL clients do not need to submit a synthetic data access request form and can apply to access to the CPRD Aurum sample dataset at no additional cost by emailing firstname.lastname@example.org.
Download: CPRD synthetic data access request form (Word, 410KB, 3 pages)
CPRD has generated high-fidelity synthetic datasets using a synthetic data generation and evaluation framework that was developed under a grant from the Regulators’ Pioneer Fund launched by The Department for Business, Energy and Industrial Strategy (BEIS) and managed by Innovate UK. The synthetic data generation and evaluation framework used to generate this synthetic dataset and the synthetic datasets are owned by the Medicines and Healthcare products Regulatory Agency (MHRA).
These high-fidelity synthetic datasets replicate the complex clinical relationships in real primary care patient data while protecting patient privacy as they are wholly synthetic. They can be used instead of real patient data for complex statistical analyses as well as machine learning and artificial intelligence (AI) research applications.
The high-fidelity synthetic datasets are based on data derived from CPRD Aurum and as such, are not suitable if you wish to understand the raw structure of the CPRD Aurum database. For this purpose, please request the medium-fidelity CPRD Aurum sample dataset.
This synthetic dataset is based on anonymised real primary care patient data extracted from the CPRD Aurum database. The dataset focuses on cardiovascular disease risk factors and was a proof-of-concept dataset developed as part of a project funded by the Regulators’ Pioneer Fund launched by The Department for Business, Energy and Industrial Strategy (BEIS) and managed by Innovate UK.
Download: CPRD Synthetic Cardiovascular Disease Data Specification (PDF, 194KB, 5 pages)
This synthetic dataset is based on anonymised real primary care patient data extracted from the CPRD Aurum database. The dataset focuses on patients presenting to primary care with symptoms indicative of COVID-19 (confirmed/suspected COVID-19) and control patients with negative COVID-19 test results. The dataset includes data on sociodemographic and clinical risk factors.
The development of this dataset was funded by NHSX using the synthetic data generation and evaluation framework developed under a grant from the Regulators’ Pioneer Fund launched by The Department for Business, Energy and Industrial Strategy (BEIS) and managed by Innovate UK.
Further information and publications
Download: Tucker et al preprint
Publication: de Benedetti, J., Oues, N., Wang, Z., Myles, P., Tucker, A. (2020). Practical lessons from Generating Synthetic Healthcare Data with Bayesian Networks. In: Koprinska I. et al. (eds) ECML PKDD 2020 Workshops. ECML PKDD 2020. Communications in Computer and Information Science, vol 1323. Springer, Cham. https://doi.org/10.1007/978-3-030-65965-3_3
Publication: Tucker, A., Wang, Z., Rotalinti, Y. et al. Generating high-fidelity synthetic patient data for assessing machine learning healthcare software. npj Digit. Med. 3, 147 (2020). https://doi.org/10.1038/s41746-020-00353-9
Publication: Wang, Z., Myles, P., Tucker, A. Generating and evaluating cross-sectional synthetic electronic healthcare data: Preserving data utility and patient privacy. Computational Intelligence. 2021; 1– 33. https://doi.org/10.1111/coin.12427
Myles, P et al. Synthetic data and the innovation, assessment, and regulation of AI Medical devices. RF Quarterly. 2021; 1(2): 48-53. © 2021. Regulatory Affairs Professional Society. (A preprint version of this article can be found at Myles et al preprint (PDF, 303KB, 6 pages))
Publication: Wang, Z et al. Evaluating a Longitudinal Data Generator using Real World Data. 2021 IEEE 34th International Symposium on Computer-Based Medical Systems (CBMS). https://doi.org/10.1109/CBMS52027.2021.00074
Download: Wang et al preprint (PDF, 1MB, 6 pages)
Download: Draghi et al. preprint (Bayesboost: Identifying and Handling Bias Using Synthetic Data Generators): http://dx.doi.org/10.2139/ssrn.4052302
Our R code for generating high-fidelity synthetic data is available via Github (https://github.com/zhenchenwang/latent_model). The R package bnlearn (v4.6.1) is used for all Bayesian network inference. The R function FCI is used, which is part of the pcalg package (v2.6–11), to identify latent variables. Kmmd is implemented using the R Package kernlab (v0.9–29). This code can be used with the high-fidelity cardiovascular synthetic dataset as a ground truth dataset, for training purposes and can be adapted for use with other similar datasets.
The CPRD Aurum sample dataset is a medium-fidelity synthetic dataset that resembles the real world CPRD Aurum with respect to the data types, data values, data formats, data structure and table relationships.
This synthetic dataset can be used for multiple purposes including as a sample dataset to understand the structure and utility of the anonymised CPRD Aurum database, to use as a data management teaching/training resource, to develop/validate/test analytics tools for use with CPRD Aurum data, to improve bespoke CPRD Aurum application interfaces/algorithms, e.g. a bespoke cohort selection tool, or to develop machine learning workflows that can be applied to anonymised CPRD Aurum data. The development of this dataset was funded by NHSX.
For pricing please contact email@example.com.
Download: CPRD Aurum sample dataset release notes (PDF, 128KB, 2 pages)