CPRD (Clinical Practice Research Datalink)

CPRD

(Clinical Practice Research Datalink) collects anonymized primary-care electronic health records from NHS general practices in the UK — around 2,400 practices, roughly one in six Britons, with GOLD and Aurum products covering different practice software ecosystems. Records include diagnoses (Read codes in GOLD; SNOMED CT in Aurum), prescriptions issued by the GP, laboratory results, referrals, and immunizations, linkable to HES hospital admissions, ONS death registrations, and disease registries (cancer, congenital anomalies, pregnancy).

Why it matters for RWE

Universal NHS registration means near-complete primary-care capture regardless of payer or employment; patients rarely leave their GP practice, producing record spans measured in decades. Prescription issuance (rather than dispensing) plus complete problem documentation makes exposure and comorbidity ascertainment unusually reliable. Linkage to hospital and mortality data closes much of the out-of-system gap that plagues US EHR sources.

Operational characteristics

• Acceptable-patient flag: practices report data quality ('up to standard'); analyses filter to acceptable patients and practices.
• Prescriptions are issued, not dispensed: assume dispensing unless secondary evidence; duration comes from quantity/dose fields.
• Coding systems: Read codes (GOLD) vs SNOMED CT (Aurum) — phenotype algorithms are not portable between them without mapping.
• Linkage subsets: HES/death linkage covers roughly half of practices and requires separate approval; linkage availability shapes cohort definitions.

Common pitfalls

• Private/secondary care prescribing: specialist-initiated drugs may not appear in primary-care issue records until GP uptake.
• Consultation-driven recording: 'healthy user' gaps mean absence of codes during quiet periods is not confirmation of absence of disease.
• Over-the-counter medications invisible: e.g., NSAIDs and aspirin bought OTC escape exposure measurement.
• UK-specific practice patterns: dosing conventions, formularies, and screening programs limit direct transportability to other systems.

Pros, cons, and trade-offs

• vs US claims: clinical depth (labs, BMI, smoking, lifestyle fields recorded in primary care) vs procedure/cost granularity.
• vs US multi-system EHR (Truveta): population-based registration and decades-long spans vs larger scale but encounter-driven capture.
• Trade-off: primary-care completeness vs specialist/oncology treatment detail that lives in hospital systems (partially recoverable via HES).

When NOT to use

Hospital-only acute outcomes without linkage approval; specialty biologics initiated exclusively in secondary care; US-generalizable cost estimates.

Decision diagram

flowchart LR
  GP[NHS general practices\ndiagnoses - scripts - labs] --> G[GOLD product]
  GP2[EMIS practices] --> A[Aurum product]
  G --> L[CPRD linkage service]
  A --> L
  H[HES admissions] --> L
  O[ONS deaths] --> L
  RG[Disease registries] --> L
  L --> S[Population-based RWE]

Worked example

Scenario

Estimate stroke risk reduction for anticoagulation adherence in atrial fibrillation patients 2010-2020.

Dataset

Adherence-stratified stroke incidence in AF cohort.

Steps

Result

Adjusted HR 1.44 (95% CI 1.25-1.66) for low vs high adherence; results stable across GOLD/Aurum implementations after code-list harmonization.

Trade-offs

Runnable example

\
import pandas as pd

def af_cohort(clinical, therapy, code_lists):
    af = clinical[clinical.medcode.isin(code_lists["af"])]
    index_dx = af.sort_values(["patid","eventdate"]).groupby("patid").first()
    ac = therapy[therapy.prodcode.isin(code_lists["anticoagulant"])]
    first_ac = ac.merge(index_dx.reset_index()[["patid","eventdate"]], on="patid")
    first_ac = first_ac[first_ac.eventdate_y <= first_ac.eventdate_x]
    return first_ac.sort_values(["patid","eventdate_x"]).groupby("patid").first()

def pdc(therapy_rows, start, end):
    # Sum covered days from qty/daily_dose, cap overlaps at calendar span
    span = (end - start).days
    covered = min((therapy_rows.qty / therapy_rows.daily_dose).sum(), span)
    return covered / span

\
library(data.table)

af_cohort <- function(clinical, therapy, af_codes, ac_codes) {
  idx <- clinical[medcode %in% af_codes][order(patid, eventdate)][1:.N, .SD[1], by = patid]
  ac <- therapy[prodcode %in% ac_codes][idx[, .(patid, dx = eventdate)], on = "patid",
                                        allow.cartesian = TRUE]
  ac[eventdate >= dx][order(patid, eventdate)][1:.N, .SD[1], by = patid]
}

pdc <- function(rx_rows, start, end) {
  span <- as.integer(end - start)
  covered <- min(sum(rx_rows$qty / rx_rows$daily_dose), span)
  covered / span
}

\
proc sql;
  create table af_index as
    select c.patid, min(c.eventdate) as dx_date format yymmdd10.
    from clinical c
    where c.medcode in (select medcode from af_codes)
    group by c.patid;

create table first_ac as
    select t.patid, t.eventdate as ac_start, t.qty, t.daily_dose
    from therapy t join af_index i on t.patid = i.patid
    where t.prodcode in (select prodcode from ac_codes)
      and t.eventdate >= i.dx_date
    group by t.patid
    having t.eventdate = min(t.eventdate);
quit;

Citations

1. [1]Herrett E, et al. Data Resource Profile: Clinical Practice Research Datalink (CPRD). International Journal of Epidemiology. 2015.
2. [2]Wolf A, et al. Data resource profile: Clinical Practice Research Datalink (CPRD) Aurum. International Journal of Epidemiology. 2019.

1. [3]NICE. Real-world evidence frameworks referencing CPRD for health technology evaluation.

1. [4]Herrett E, et al. Data Resource Profile: Clinical Practice Research Datalink (CPRD). International Journal of Epidemiology. 2015.