ACCEPT 3.0-CPRD: Predicting COPD Exacerbation Risk in UK Primary Care

Overview

accept3_cprd() implements ACCEPT 3.0-CPRD, a recalibration of the ACCEPT 2.0 model specifically for UK primary-care patients. It was derived from the Clinical Practice Research Datalink (CPRD), a large UK primary-care database, and is designed to produce well-calibrated 1-year exacerbation risk predictions in this setting.

The function predicts two outcomes:

Moderate-to-severe exacerbation probability and rate
Severe exacerbation probability and rate

It can also be called via the unified accept() interface using country = "GBR-primary".

When to Use ACCEPT 3.0-CPRD

Model	Population	Setting
`accept2()`	Canadian/international	Hospital or specialist
`accept3()` / `accept3_gbr()`	UK	Secondary care
`accept3_cprd()`	UK	Primary care (GP)

Use accept3_cprd() when your patients are from a UK general practice setting, where disease severity tends to be milder and fewer specialist measurements are routinely available.

Required Identifier

The following column is required but is not a predictor — it is used solely to label rows in the output so predictions can be matched back to patients:

Column	Type	Description
`ID`	character/numeric	Unique patient identifier

Required Predictors

The following columns are the actual model inputs and must be present in your data:

Column	Type	Description
`age`	numeric	Age in years
`male`	integer (0/1)	Sex (1 = male)
`FEV1`	numeric	FEV1 % predicted
`LastYrExacCount`	integer	Total exacerbations in the past year (moderate + severe)
`LastYrSevExacCount`	integer	Severe (hospitalised) exacerbations in the past year
`mMRC` or `SGRQ`	numeric	Dyspnoea score (mMRC 0–4) or SGRQ total score

Note: LastYrSevExacCount must not exceed LastYrExacCount. At least one of mMRC or SGRQ must be provided; if only SGRQ is available it is converted to mMRC internally.

Optional Predictors

These columns are used if available; if missing they are automatically imputed using a CPRD-specific sequential triangular regression model:

Column	Type	Description
`LABA`	integer (0/1)	Long-acting beta-agonist use
`oxygen`	integer (0/1)	Long-term oxygen therapy
`ICS`	integer (0/1)	Inhaled corticosteroid use
`LAMA`	integer (0/1)	Long-acting muscarinic antagonist use
`statin`	integer (0/1)	Statin or CVD medication use
`BMI`	numeric	Body mass index (clamped to 10–60)
`smoker`	integer (0/1)	Current smoker

Imputation proceeds in the order above. Each variable is predicted from the mandatory predictors plus any previously imputed optional predictors.

Basic Usage

library(accept)

# Using the sample patients bundled with the package
results <- accept3_cprd(samplePatients)
head(results)

#> # A tibble: 6 x 4
#>   predicted_exac_probability predicted_exac_rate predicted_severe_exac_probability predicted_severe_exac_rate
#>                        <dbl>               <dbl>                             <dbl>                      <dbl>
#> 1                      0.699               1.20                              0.356                      0.439
#> 2                      0.324               0.392                             0.0415                     0.0424
#> ...

Accessing Results via `accept()`

You can also call accept3_cprd through the unified interface:

results <- accept(
  newdata  = samplePatients,
  version  = "accept3",
  country  = "GBR-primary"
)

Including Predictors in the Output

Set return_predictors = TRUE to include the input columns alongside predictions. This is useful for checking which values were imputed:

results_full <- accept3_cprd(samplePatients, return_predictors = TRUE)
names(results_full)

Imputation Messages

By default, imputation is silent. Set quiet = FALSE to see which predictors were imputed for each patient:

results <- accept3_cprd(samplePatients, quiet = FALSE)
# Example message:
# Patient 3: imputed LABA, BMI, smoker

Handling Missing mMRC / SGRQ

The function requires either mMRC (0–4) or SGRQ (0–100). If only SGRQ is provided, it is converted to mMRC using the published conversion formula before prediction. If neither is available, an error is raised.

# Data with SGRQ instead of mMRC
patients_sgrq <- samplePatients
patients_sgrq$mMRC <- NULL  # remove mMRC
patients_sgrq$SGRQ <- c(45, 30, 60, 20, 55, 10)  # add SGRQ

results <- accept3_cprd(patients_sgrq)

Recalibration Method

ACCEPT 3.0-CPRD applies a Cox-model recalibration to ACCEPT 2.0 predictions:

\[\hat{p}_{UK} = 1 - \exp\!\left(-H_0 \cdot \exp\!\left(\beta \cdot \log(-\log(1-\hat{p}_{2}))\right)\right)\]

where \(\hat{p}_{2}\) is the ACCEPT 2.0 predicted probability, and \(H_0\) and \(\beta\) are optimism-corrected parameters estimated via 200-resample bootstrap from the CPRD dataset:

Outcome	\(H_0\)	\(\beta\)
Moderate-to-severe	0.676	0.986
Severe	0.482	1.124

Comparing Models

library(accept)
library(dplyr)

patients <- samplePatients

a2   <- accept2(patients)
a3   <- accept3_cprd(patients)

comparison <- bind_cols(
  ID        = patients$ID,
  accept2   = a2$predicted_exac_probability,
  accept3uk = a3$predicted_exac_probability
)

print(comparison)

Predictions from ACCEPT 3.0-CPRD are generally lower than ACCEPT 2.0 for moderate-to-severe exacerbations, reflecting the milder disease burden typically seen in UK primary care compared with specialist settings.

Session Info

sessionInfo()