Motivation

The Clinical Practice Research Datalink (CPRD) is a UK database of anonymised primary care electronic health records. It is widely used to study the effectiveness and safety of medications. However, prescription data from CPRD are often messy, with systemic issues such as missing information on prescription end dates and medication quantities, and prescribing instructions are typically provided as unstructured free text. In pharmaco-epidemiology studies, the data preparation steps used to determine drug exposure are rarely fully reported, yet assumptions made during this stage can have considerable implications for risk attribution of possible adverse events.

We have previously developed a framework¹ for dealing with missing information on prescription end dates and other issues such as overlapping prescription periods. The framework was implemented using software. Besides being only available for users, the earlier algorithm did not deal with the problem of free-text prescriptions. The current R package, drugprepr (formerly drugprepCPRD), builds upon the earlier algorithm and aims to make the framework available to a wider audience through its implementation as free, open-source software.

This vignette describes how to use the drugprepr package to transform CPRD drug data contained in a typical therapy.txt file into information on individuals’ drug use over time. We will walk through all the steps needed to perform the transformation. We assume the user is already familiar with CPRD data and has basic knowledge of R software.

CPRD data

The CPRD Gold data follows the CPRD Gold specification. It is made up of several tables containing information related to the patients. One of the CPRD tables, called ‘Therapy’, contains the prescription information for a patient. This table can be linked to the ‘common dosages’ lookup table to get the text instructions for the prescribed product (i.e. a drug). Table , based on the fictional dataset cprd that is bundled with this package, presents hypothetical prescription data for two individuals.

The event_date in the table is often used as the start date of exposure, but the stop date of the exposure is not available per se and we have to infer it from the available information. CPRD provides two options, namely numdays (number of days) and dose_duration, which can be used to determine the prescription stop date. However, these values are often missing, and even when present do not offer flexibility for the researcher in computing the stop dates in the case of prescriptions with a variable dose frequency (number of times the prescription is taken per day) and dose number (e.g. the number of tablets to take at a time).

As described below, the drugprepr package extracts the dose frequency and dose number from the text instructions and provide a series of data processing steps with multiple options to define start and stop dates of a given prescription. The package works at the prodcode level to give granularity for each medication.

The algorithm

The package drugprepr is made up of the following functions that must be executed sequentially.

Compute numerical daily dose (ndd)

Extract dose frequency (number of times the prescription is to be taken per day) and dose number (number of units of drug to take at a time) from the free text, in order to compute the numerical daily dose.

Define implausible values

Given “plausible” values (based on prescribing guidelines and clinical experience), this stage identifies those values outside the plausible range, which may a result of data input or processing errors.

Decision 1: Handle implausible quantities

Values outside the “plausible” range may be [1a] ignored, [1b] set to missing, or imputed. Imputation options include: [1c1] set to the mean value for that patient for that product code, [1c2] set to the mean value for that practice for that product code, [1c3] set to the mean value for the whole cohort for that product code, etc. See the package manual for all possible options.

Decision 2: Handle missing quantities

Options for missing qty are: [2a] leave as missing, [2b1] set to the mean value for that patient for that product code, [2b2] set to the mean value for that practice for that product code, [2c] set to the mean value for the whole cohort for that product code, etc,.

Decision 3: Handle implausible numerical daily doses

Options for implausible numerical daily doses are the same as for decision 1.

Decision 4: Handle missing numerical daily doses

Options for missing numerical daily doses are the same as for decision 2.

Decision 5: Clean duration

cleans implausibly high values for each of the three available duration variables (numdays, dose_duration, and qty/ndd). Options for cleaning each duration variable are: [5a] make no changes, [5b(X)] set to missing if duration is greater than X months, or [5c(X)] set to X if duration is greater than X months. X is 6, 12, or 24.

Decision 6: Select stop date

defines a stop date for each prescription. calculate stop date as prescription start date + one of the following duration definitions: [6a] numdays, [6b] dose_duration, [6c] qty/ndd, or [6d(X)]

Decision 7: Handle missing stop date

if stop date is missing: [7a] keep as missing, [7b] set to the mean value for that product code for that patient, [7c] set to the mean value for that product code for the whole cohort, [7d] set to the mean value for that product code for that patient, otherwise set to the mean value for that product code for the whole cohort.

Decision 8: Handle multiple prescriptions

for multiple prescriptions for the same product code on the same day, but with different stop dates, options are: [8a] do nothing, ; [8b] calculate the mean duration of prescriptions and drop redundant records; [8c] keep the record with the shortest duration; [8d] keep the record with with the longest duration; [8e] sum the durations and drop redundant records.

Decision 9: Handle overlapping prescriptions

for consecutive records with overlapping start and stop dates, options are [9a] to ignore the overlap but sum ndds; [9b] move the overlapping time.

Decision 10: Handle short gaps between prescriptions

handles small gaps between consecutive prescriptions by either allowing these gaps to remain classified as unexposed or reclassifying the gaps as exposed when the gap is less than a specified number of days. options include [10a] do nothing – the gap remains classified as unexposed; [10b(X)] move the stop date of the preceding prescription to “fill in” the gap, reclassifying the time as exposed, if the gap between consecutive prescriptions is less than X days. X is 15, 30, or 60

R package drugprepr

You can install the latest development version from GitHub with the following R commands:

install.packages('remotes')
remotes::install_github("belayb/drugprepr")

Once the package is installed, we load the drugprepr package into the working environment as follows.

library(drugprepr)

data_ndd <- compute_ndd(cprd, min, min)

result <- drug_prep(data_ndd,
                    min_max_dat,
                    decisions = c('b', 'b1', 'b', 'b1', 'b_6',
                                  'c', 'a', 'd', 'a', 'b_15'))

decisions <- make_decisions('ignore',
                            'mean population',
                            'missing',
                            'mean practice',
                            'truncate 6',
                            'qty / ndd',
                            'mean individual',
                            'mean',
                            'allow',
                            'close 15')
decisions
#>      implausible_qty          missing_qty      implausible_ndd 
#>                  "a"                 "b3"                  "b" 
#>          missing_ndd implausible_duration   calculate_duration 
#>                 "b2"                "c_6"                  "c" 
#>     missing_duration          clash_start          overlapping 
#>                  "b"                  "b"                  "a" 
#>           small_gaps 
#>               "b_15"
# drug_prep(example_therapy, plausible_values, decisions)