About 98% of the population in the United Kingdom is registered with a general practitioner.14
The Health Improvement Network (THIN) is a database of electronic medical records from UK general practices. Participating general practitioners systematically and prospectively retrieve and enter clinical information on patients, including personal data, diagnoses, and prescriptions so that the database provides a longitudinal medical record for each patient. All sections of the population are represented in THIN. The data are collected in a non-interventional way during routine general practice and therefore reflect “real life” clinical care. The information is continually updated. Information from secondary care and other medically related information received by the practice is transcribed and entered retrospectively. Prescribing by general practitioners is particularly well recorded since the computerised entry made by the doctor is also used as the prescription form. Previous comparison with external statistics and other independent studies have shown that both the clinical diagnostic and prescribing information are well recorded and accurate.15
THIN, however, represents data collected from the general practitioner’s medical records and reflects only those events that are deemed to be relevant to the patient’s care and not for research purposes. In most cases a medicine prescribed for the first time is temporally linked with a medical event record (symptom or diagnosis), but no permanent link exists between a prescribed treatment and the reason for prescription. Drugs prescribed by hospital doctors or other specialists do not appear in THIN data unless the treatment is to be continued by the general practitioner. However, due to the constraints of specialist and hospital prescribing budgets many prescriptions issued outside of the general practice are sufficient to cover only the first seven days. After this time the patient usually obtains subsequent prescriptions from their general practitioner. Lastly, some of the drugs used in the United Kingdom are obtained over-the-counter so are not entered in general practitioner records. For this study we used data from 1 January 1990 to 31 December 2008 from 424 general practices and we excluded events that occurred within six months after registration to include only incident cases of outcomes.19
Prescriptions of glucocorticoids
Each prescription of a drug is recorded in THIN as encrypted multilex codes, along with reference to the chapter in the British National Formulary
describing the drug.20
We selected all glucocorticoids prescribed orally or by intramuscular or intravenous injection and this included prednisolone, prednisone, dexamethasone, triamcinolone, betamethasone, methylprednisolone, and deflazacort. We included people aged 18 years and older and identified all who were prescribed such treatment. For multiple consecutive prescriptions we considered that these were part of a unique course of treatment if the gap between two consecutive prescriptions was less than three months. For each course we defined the treatment duration as the time from the first to the last prescription plus the duration of the last prescription. Data on adherence to treatment are not routinely recorded in THIN. To ensure that our samples comprised those people who were most likely taking the drug, we restricted our populations of glucocorticoid users to those who received at least two successive prescriptions for glucocorticoids. We calculated the average daily dosage by multiplying the number of pills prescribed by the dose per pill (calculated in prednisone equivalent) and then divided this by the number of days for which the drug was prescribed. The medical diagnosis recorded on the date of starting glucocorticoids was used as the indication for the prescription. We excluded people taking substitutive glucocorticoids for adrenal insufficiency.
Identification of iatrogenic Cushing’s syndrome
All diagnoses and symptoms are recorded in THIN using the Read classification system.21
Using the method described previously22
we retained four codes to define iatrogenic Cushing’s syndrome: iatrogenic Cushing’s syndrome, drug induced Cushing’s syndrome, steroid facies, and cushingoid facies.
We selected two cohorts from patients recorded in THIN. The first comprised all those prescribed systemic glucocorticoids and the other those who were never prescribed systemic glucocorticoids. Among patients prescribed glucocorticoids, we defined those with a recorded diagnosis of iatrogenic Cushing’s syndrome as the exposed group and those without a recorded diagnosis of iatrogenic Cushing’s syndrome as eligible for inclusion in the first comparison group. Anyone in the second cohort was eligible to be included in the second comparison group, comprising those who had never been prescribed systemic glucocorticoids. When selecting the comparison groups we stratified the samples to ensure the same distribution among groups for sex and age (within 10 year age bands). Additionally, for the first comparison group we stratified the sample to ensure that the distribution for duration of use of glucocorticoids and initial dosage were similar to that of the group prescribed glucocorticoids and with iatrogenic Cushing’s syndrome. Lastly, because of the possible association between indications for treatment with glucocorticoids and cardiovascular disease, in the comparison groups we preferentially selected people with the same underlying diseases as those in the group prescribed glucocorticoids and with iatrogenic Cushing’s syndrome. We selected up to six times as many people in each of the two comparison groups as those in the group prescribed glucocorticoids with iatrogenic Cushing’s syndrome. Patients in both groups were selected at random from the pool of eligible patients. The start of the follow-up period for the analyses was defined by an index date. For those from the group prescribed glucocorticoids with iatrogenic Cushing’s syndrome, this index date was defined as the first record of iatrogenic Cushing’s syndrome. For patients in the comparison groups the index date was randomly selected from within the time they were registered with the general practice. In the case of those prescribed glucocorticoids with no record of iatrogenic Cushing’s syndrome, this period was restricted to that during which they were prescribed the drug.
Covariates of interest
We identified smoking status based on the nearest record before the index date, including data up to five years before this date. For each included patient we extracted clinical data (height, weight, and blood pressure) and biological data (fasting glucose, total cholesterol, and triglycerides levels) from his or her medical records. Data were collected for two periods: in the six months before the index date and in the six months after the index date. Moreover, we searched the drug treatment files for relevant prescriptions. If patients had at least two prescriptions for either diabetes drugs, antihypertensive drugs, cholesterol lowering drugs, oral anticoagulants, or antiplatelet drugs we defined them as being treated with that particular drug. To ascribe the hazard of cardiovascular events to the presence of iatrogenic Cushing’s syndrome, we examined the risk of some diseases which were a priori not related to Cushing’s syndrome. For cardiovascular disorders we searched for valvular heart diseases. For non-cardiovascular events, we defined impacted cerumen in the ear (one of the most common diagnoses recorded in THIN), sprain, and neoplasm (malignant neoplasm of digestive organs, skin, breast, or genitourinary organ) as negative control diseases.
Identification of cardiovascular events
Using the same method as for iatrogenic Cushing’s syndrome, we developed Read code lists to identify recorded diagnoses of either coronary heart disease, heart failure, or ischaemic cerebrovascular events. To exclude a possible but not confirmed diagnosis we only selected stringent codes—for example, acute myocardial infarction, cardiac failure, transient cerebral ischaemia, stroke. Moreover, we also examined anonymised free text associated with records of death to capture fatal events that may not have been recorded by Read codes.
In each group we calculated the incidence of cardiovascular events within the first year after the index date by dividing the number of newly diagnosed cases by the follow-up time up to one year after the index date. If patients had several events we censored them at their first event. We compared the three groups to assess hazard ratios associated with iatrogenic Cushing’s syndrome. We adjusted the estimated hazard ratio for age (continuous variable); sex; underlying disease; smoking status; history of prescription for aspirin, oral anticoagulants, diabetes drugs, antihypertensive drugs, or cholesterol lowering drugs; and initial dosage (continuous variable) and duration (continuous variable) of glucocorticoid use for people prescribed the drug. Using Cox proportional hazards models we accounted for clustering at the general practice level. Proportional hazard assumptions were checked graphically and by analysing Schoenfeld residuals. We checked linearity for continuous variables by comparing two models, one with the linear term and the other with the categories, using the log likelihood ratio test. Continuous variables are presented as medians and 25th to 75th centile values. Categorical variables are presented as proportions, with 95% confidence intervals indicating precision of estimates. Incidence rates are reported per 100 person years at risk. All analyses were done using Stata, version 11.1.