Revitalizing the General Practice Research Database: plans, challenges, and opportunities

A Two-Stage Decision Making Approach for Safety Studies

For safety studies, two types of hypothesis testing are often considered: detecting a safety signal and ruling out a safety concern. Under the detecting framework, statistical non-significance is often confused with the conclusion that there is no safety concern. Such a conclusion, in the presence of low study power or large variability, is problematic. To overcome the interpretation issue with non-significant results from a detecting hypothesis, we propose a Two-Stage Decision-Making (TSDM) approach for safety studies. It is basically a ruling-out design allowing an interim analysis that applies both detecting and ruling-out criteria at the interim and final stages with a pre-specified alpha spending function. The proposed TSDM approach incorporates both detecting a safety signal and ruling out safety concerns into a single study design to increase the probability of making a definite decision. It is based on the ruling-out framework that utilizes both directions of the confidence interval to make a decision for ruling out unacceptable risk or detecting safety signal at each analysis stage. We assess the proposed TSDM approach by investigating properties such as operational type I error rate, overall study power based on analytical approximations, overall probability of making a decision, and required sample sizes. We conduct Monte Carlo simulations to evaluate such properties regarding various outcome types of confidence intervals and summarize the statistical interpretations and the implications on study design.

Application of multicategory exposure marginal structural models to investigate the association between long-acting beta-agonists and prescribing of oral corticosteroids for asthma exacerbations in the Clinical Practice Research Datalink

OBJECTIVE

To examine the comparative effectiveness of inhaled long-acting beta-agonist (LABA), inhaled corticosteroid (ICS), and ICS/LABA combinations.

METHODS

We used a retrospective cohort design of patients older than 12 years with asthma diagnosis in the Clinical Practice Research Datalink to evaluate asthma-related morbidity measured by oral corticosteroid (OCS) initiation within 12 months of initiating LABAs, ICSs, or ICSs/LABAs. Asthma severity 12 months before drug initiation (use of OCSs, asthma-related hospital or emergency department visits, and number of short-acting beta-agonist prescriptions) and during follow-up (short-acting beta-agonist prescriptions and total number of asthma drug classes) was adjusted as a time-varying variable via marginal structural models.

RESULTS

A total of 51,103 patients with asthma were followed for 12 months after receiving first prescription for study drugs from 1993 to 2010. About 92% initiated ICSs, 1% initiated LABAs, and 7% initiated ICSs/LABAs. Compared with ICSs, LABAs were associated with a 10% increased risk of asthma exacerbations requiring short courses of OCSs (hazard ratio [HR] 1.10; 95% confidence interval [CI] 1.07-1.18). ICS/LABA initiators were 62% less likely than ICS initiators (HR 0.38; 95% CI 0.12-0.66) and 50% less likely than LABA initiators to receive OCS prescriptions for asthma exacerbations (HR 0.50; 95% CI 0.14-0.78).

CONCLUSIONS

In concordance with current asthma management guidelines, inhaled LABAs should not be prescribed as monotherapy to patients with asthma. The findings suggest the presence of time-dependent confounding by asthma severity, which was accounted for by the marginal structural model.

Recent advances in the utility and use of the General Practice Research Database as an example of a UK Primary Care Data resource

Since its inception in the mid-1980s, the General Practice Research Database (GPRD) has undergone many changes but remains the largest validated and most utilised primary care database in the UK. Its use in pharmacoepidemiology stretches back many years with now over 800 original research papers. Administered by the Medicines and Healthcare products Regulatory Agency since 2001, the last 5 years have seen a rebuild of the database processing system enhancing access to the data, and a concomitant push towards broadening the applications of the database. New methodologies including real-world harm-benefit assessment, pharmacogenetic studies and pragmatic randomised controlled trials within the database are being implemented. A substantive and unique linkage program (using a trusted third party) has enabled access to secondary care data and disease-specific registry data as well as socio-economic data and death registration data. The utility of anonymised free text accessed in a safe and appropriate manner is being explored using simple and more complex techniques such as natural language processing.

Determining the predictive value of Read codes to identify congenital cardiac malformations in the UK Clinical Practice Research Datalink

PURPOSES

The purposes of this study were to determine (i) the positive predictive value (PPV) of multiple Read codes used to identify congenital cardiac malformation (CCM) cases in the UK Clinical Practice Research Datalink (CPRD); (ii) the accuracy of the diagnosis date; and (iii) the source of information that the general practitioners (GPs) use for validating the diagnosis suggested by the code.

METHODS

Eight hundred eighty-eight records with Read diagnostic and procedures codes for CCM, between January 1996 and November 2010, were identified from CPRD. Questionnaires were sent to GPs to verify the diagnoses and date of the code-identified events.

RESULTS

A total of 719 questionnaires were returned (81% response rate). The PPV of the CCM codes was 93% (670/719). Thirty-one percent of cases had a different event date than the one recorded in the electronic medical record (EMR); 10% of these differing dates were within 30 days of the code-identified CCM date. GPs used a variety of data sources to confirm CCM diagnoses. Although the EMR was the most frequently used data source (70%), 66% reported using consultation letters, 9% reported using clinical notes or paper charts, and 35% of GPs reported using the hospital record to confirm the CCM diagnosis.

CONCLUSIONS

Clinical Practice Research Datalink Read codes for CCMs have 93% PPV and most likely point to true cases. However, the accuracy of diagnosis dates and the age at diagnosis may not be as reliable. The findings of this study indicate that GPs use information beyond what is available for researchers in the EMR to confirm clinical diagnoses when responding to validation questionnaires. Published 2013. This article is a U.S. Government work and is in the public domain in the USA.

Determining the predictive value of Read/OXMIS codes to identify incident acute myocardial infarction in the General Practice Research Database

PURPOSE

To determine (1) the positive predictive value (PPV) of multiple Read/OXMIS codes to identify incident acute myocardial infarction (AMI) cases in General Practice Research Database (GPRD); (2) the ability to capture the correct timing of the clinical event.

METHODS

A random sample of 238 records (from 155 general practitioner (GP) practices) with AMI codes, between 1 January 1997 and 31 December 2004, was selected from GPRD. Questionnaires were sent to the GPs to verify the diagnosis and timing of code-identified incident AMI events and collect supporting information. We calculated the PPV of the AMI codes as the proportion of code-identified AMIs that the GPs confirmed as AMI cases. Two physicians from Food and Drug administration (FDA), blinded to the GP response, reviewed the supporting hospital records returned by GPs for 98 AMI cases.

RESULTS

A total of 217 questionnaires were completed (91% response rate). The PPV of the AMI codes was 93% (201/217). Thirty one (15%) cases had a different event date than the one recorded in the electronic medical records (EMR); 28 (90%) of the dates were within 15 days. One GP indicated that a patient had a previous AMI, 110 days before the codes that captured the AMI diagnosis. A total of 159 (79%) AMI cases were hospitalized; hospital records were provided for 98 patients. Physician review of the hospital records found that 96% of these records had enough information for classification, but not independent diagnosis, of AMI.

CONCLUSIONS

Information in GPRD is sufficient to identify incident AMI cases and determine the event date with reasonable accuracy.

Obesity and the use of insulin: a study of patients with type 2 diabetes in the UK

OBJECTIVE

This analysis examines how obesity affects the prescribing of insulin for individuals with type 2 diabetes and poor glycemic control.

METHODS

Data were obtained from the UK General Practice Research Database for the years 2000-2004. Patients were eligible if they had been identified as having type 2 diabetes and had undergone at least two valid glycosylated hemoglobin (HbA(1c)) tests. Additionally, patients had to have poor glycemic control on the index date (HbA(1c)>7.4), no use of insulin 6 months prior to the index date, and at least 30 months of data after the index date (N=6468). Descriptive statistics were used to examine unadjusted differences between obese and nonobese patients. A Cox proportional hazards model was applied to examine the relationship between obesity and the relative likelihood of initiation of insulin while controlling for differences in patient characteristics, medication use, and HbA(1c) levels.

RESULTS

Obese individuals were significantly younger (P<.01), significantly more likely to be treated with two oral antidiabetic agents (P=.02), and significantly less likely to be treated with oral monotherapy (P=.02). Controlling for a wide range of factors that may impact receipt of insulin, results revealed that obese individuals had a "hazard" of initiation of insulin significantly lower than that of nonobese patients (hazard ratio=0.814, P=.01).

CONCLUSIONS

Patient age, severity of illness, and prior medication use all affect whether the individual will be initially prescribed insulin. Moreover, the results of this study demonstrate that obesity is an additional critical factor in physicians' decision to begin treatment with insulin.

Adverse event detection in drug development: recommendations and obligations beyond phase 3

Premarketing studies of drugs, although large enough to demonstrate efficacy and detect common adverse events, cannot reliably detect an increased incidence of rare adverse events or events with significant latency. For most drugs, only about 500 to 3000 participants are studied, for relatively short durations, before a drug is marketed. Systems for assessment of postmarketing adverse events include spontaneous reports, computerized claims or medical record databases, and formal postmarketing studies. We briefly review the strengths and limitations of each. Postmarketing surveillance is essential for developing a full understanding of the balance between benefits and adverse effects. More work is needed in analysis of data from spontaneous reports of adverse effects and automated databases, design of ad hoc studies, and design of economically feasible large randomized studies.

Evaluation of the General Practice Research Database congenital heart defects prevalence: comparison to United Kingdom national systems

BACKGROUND

As part of an effort to validate the General Practice Research Database (GPRD) for future studies of medication use in pregnancy, this study examined whether the rates of all, and specific types of, congenital heart defects obtained from the GPRD are similar to those obtained from UK national systems.

METHODS

The prevalence rates of heart defects for 2001-2003 were determined from the GPRD and compared with both the National Congenital Anomaly System (NCAS) and the European Concerted Action of Congenital Anomalies and Twins (EUROCAT). Rate ratios (RRs) and 95% CIs were calculated comparing the prevalence of all congenital heart defects as well as specific types of heart defects in the three data sources. In addition, the effect of the child's age on the frequency of heart defects in the GPRD was determined.

RESULTS

The prevalence of heart defects in the GPRD was more than twice as high as in the NCAS and slightly higher than in the EUROCAT. All differences were statistically significant. The prevalence of specific heart defects varied across the GPRD, NCAS, and EUROCAT. The measured prevalence of congenital heart defects in the GPRD was higher if calculated including children up to age 6.

CONCLUSIONS

The comparisons of the GPRD prevalence rates to national prevalence estimates demonstrate that the GPRD can serve as a more complete source of background prevalence for the most commonly occurring congenital heart defects, which is essential to properly assess possible associations between maternal exposures and congenital heart defects.

Patterns of blood glucose monitoring in relation to glycemic control among patients with type 2 diabetes in the UK

AIMS

The objectives of this study were to compare patterns of blood glucose monitoring among patients with type 2 diabetes initiating therapy with insulin or oral medication and to examine the relationship between the quantity of prescribed monitoring strips and glycemic control.

METHODS

Data were obtained from the UK General Practice Research Database. Patients were eligible if they were identified as having type 2 diabetes, initiated therapy with insulin or an oral agent, and had 12-month postinitiation data. Differences in patient characteristics and number of test strips prescribed between the insulin (n=347) and oral cohorts (n=2436) were examined. Multivariate regressions examined the relationship between quantity of monitoring and glycemic control for a subset of patients (insulin, n=245; oral, n=1795) with available glycosylated hemoglobin (HbA1c) data.

RESULTS

During the 12-month postinitiation period, patients using insulin were prescribed approximately twice as many test strips compared with those patients using oral medication (149 vs. 78, P<.0001). Multivariate regressions revealed that individuals who initiated insulin therapy and were prescribed enough test strips to test at least once per day in the 6 months prior to the HbA1c test date had, on average, a 0.65% lower HbA1c value (P=.02) compared with individuals who were prescribed fewer test strips.

CONCLUSIONS

Results indicate significant differences in the prescription of blood glucose monitoring strips, with patients initiated on insulin prescribed almost twice as many test strips compared with patients initiated on orals. The greater number of blood glucose test strips prescribed was associated with lower HbA1c values for insulin patients.

A review of 10 years of systematic health surveillance in the Army

Health and morbidity reporting has been an important feature of the historical assessment of military campaigns from times of antiquity. Most of these reports have concentrated on hospital admission rates and mortality. In 1994 the British Army introduced a primary care health surveillance reporting system called J94. This provided the first opportunity for the systematic capture and analysis of morbidity data that allowed the identification of disease trends and the audit of remedial action. In parallel with the developments made by the military in the field of health surveillance, a number of initiatives in the NHS tried to develop real time surveillance systems with differing degrees of success. This paper reviews the developments made by military and civilian programs, identifies the problems that have been faced, areas where success has been achieved and the issues that will have to be considered as we prepare for the introduction of the next generation of IT based medical information systems into the military.

The flawed basis for FDA post-marketing safety decisions: the example of anti-depressants and children

The FDA (February 3, 2005) issued a black box warning that all antidepressants increase the risk of suicidal thinking and behavior in children and adolescents that must be cited in all advertising as well as included in the package insert. Following this, there was a sharp decrease in antidepressant prescriptions for children with uncertain public health impact. The current black box does not claim that these medications increase the risk of completed suicides, although this is the clear implication of the term 'suicidality'. The interpretation by the press is unequivocal that lethal outcomes prompted this action. This review concludes the following: (1) Since no suicide occurred in clinical trials of approximately 4400 children, the analyses relied upon 'suicidality' as a surrogate. (2) The classification of adverse events by the Columbia group necessarily relied on inferences, because the available evidence was not prospectively collected for this purpose. (3) The data analysis relied on a composite variable labeled 'suicidality', an unvalidated, inappropriate surrogate. Specific criticisms of analytic procedures and inferences are presented. The failure of the FDA's post-marketing surveillance system is reviewed. The data necessary for objective evaluations of possible post-marketing harm cannot be gathered by the current process. Proper prospective post-marketing surveillance by linked computerized medical records is a crucial issue that deserves major public and political attention and prompt action.

Safety of medications prescribed before and during early pregnancy in a cohort of 81 975 mothers from the UK General Practice Research Database

PURPOSE

To demonstrate a linkage methodology for mother and baby automated medical records, and describe frequency, type, and pregnancy risk level of medications prescribed during pregnancy in a GPRD cohort, 1991-1999.

METHODS

We linked records using a two-stage algorithm and selected pairs with > or = 7 months prenatal records and > or = 2 records in baby's first year of life. Periods of interest were: 90 days prior to a woman's earliest identified pregnancy record (Period I), and this record plus 70 days (Period II, approximate early pregnancy). Medications were classified using the British National Formulary and US Food and Drug Administration Pregnancy Risk Categories.

RESULTS

We achieved over 80% record linkage and defined a cohort of 81,975. Sixty-five per cent of mothers had > or = 1 prescription during both periods combined. Most frequent medications in Period I were anti-bacterial, contraceptive, topical steroid, and bronchodilator. In Period II, they were folic acid, anti-bacterial, antacid, and gynecological anti-infective. In Period I, 4% were FDA category A (considered safest), 34% B, and 49% C and D combined. By Period II, prescription of category A medications increased (folic acid, iron) while other categories declined. Category X medications, with potential teratogenic risk that outweighs maternal benefit, were prescribed to 5714 (7%) women in Period I, and 501 (0.6%) women in Period II (46% progesterone).

CONCLUSIONS

One in every 164 women received a category X prescription in early pregnancy. The visit when pregnancy is first medically recognized represents an opportunity to review prescribed medications in light of contraindication and/or fetal risk.

Economic issues in access to medications: schizophrenia treatment in England

Mental health problems and the services to address them are currently receiving more attention in the UK than ever. Mental health care in England--indeed, across the UK--is experiencing a much needed transformation. It is therefore highly pertinent to examine the patterns of psychotropic medication use, given their intended links to recovery, rehabilitation, and reintegration, as well as to explore the economic and other factors that appear to influence those patterns. These are the aims of this paper. Our attention will be primarily focused on England. What this analysis shows is that given a higher profile by government, including additional funding (although not really benefiting differentially compared to other parts of the health service) and the first national service framework, it is possible to see changes in service patterns, access and (to a degree) outcomes. These changes are occurring at a time when new classes of psychotropic medication are being introduced in a range of therapeutic areas, contributing to the relatively rapid growth of take-up but also raising questions about appropriateness, effectiveness, cost-effectiveness and equity.

Antidiabetic drugs and heart failure risk in patients with type 2 diabetes in the U.K. primary care setting

OBJECTIVE

To assess the effects of antidiabetic drugs on the risk of heart failure in patients with type 2 diabetes.

RESEARCH DESIGN AND METHODS

We conducted a retrospective cohort study with a newly diagnosed diabetes cohort of 25,690 patients registered in the U.K. General Practice Research Database, 1988-1999. We categorized person-time drug exposures to monotherapies in insulin, sulfonylureas (SUs), metformins, and other oral hypoglycemic agents (i.e., acarbose, guar gum) and combination therapy including insulin, combination therapy without insulin, and triple combination therapy with or without insulin. A drug-free time interval served as a reference category. Cox interval-wise (piece-wise) regression analyses were used. The main outcome was incident heart failure.

RESULTS

Among 43,390 drug exposure intervals for 25,690 patients who had a mean follow-up period of 2.5 years, 1,409 patients developed heart failure. Heart failure occurred most frequently in SU monotherapy exposure. After adjusting for duration of diabetes, the timing and order of treatments received, and known risk factors for heart failure, we found no differential effects among type-specific therapies. Patients with any drug use within the first year after diabetes diagnosis had a 4.75-fold higher risk (hazard ratio) for heart failure than those with drug-free status but had no increased risk during subsequent years.

CONCLUSIONS

In conclusion, the use of any pharmacological therapy for type 2 diabetes appears to be associated with an increased risk of heart failure. This risk does not persist beyond the first year after diagnosis of diabetes and does not appear to differ among the types of drug therapy examined. This observation suggests that the severity of diabetes or the preclinical duration of diabetes and the need for drug therapy, and not the therapy itself, is an explanation for heart failure in patients with type 2 diabetes.

Strategies for identifying pregnancies in the automated medical records of the General Practice Research Database

PURPOSE

To develop a method for identifying the beginning and ending records of pregnancies in the automated medical records of the General Practice Research Database (GPRD).

METHODS

Women's records from 1991 to 1999 were searched for codes from 17 pregnancy marker and 7 pregnancy outcome categories. Using the retrieved records, all possible pregnancy marker-outcome combinations were formed per woman. For each combination, the difference in days between record event dates was calculated. Restrictions were applied to select the combination with the earliest pregnancy marker mapped to the first outcome for each pregnancy. Iterations of the algorithm identified multiple pregnancies per woman when present. The algorithm was evaluated by analyzing time between marker and outcome event dates of mapped pregnancies and by analyzing unmapped pregnancy markers and outcomes.

RESULTS

A total of 297,082 pregnancies were identified: 80% by general practitioner (GP) visit codes as the earliest pregnancy marker and 14% by laboratory or procedure codes. Limiting pregnancies to one per woman aged 15-44 years yielded 209,266 pregnancies. Pregnancy mapping success was greater than 80%. Plotting the pregnancies by weeks from earliest pregnancy marker to outcome and by pregnancy marker category showed two peaks in the distribution: 2-3 weeks and 33 weeks.

CONCLUSIONS

Arranging codes and time into algorithms provides a useful tool for pregnancy identification in databases whose size prohibits the audit of printed records. Evaluation of our algorithm confirmed a high degree of mapping success and a sensible time distribution from pregnancy marker to outcome.

Relative cost effectiveness of Depo-Provera, Implanon, and Mirena in reversible long-term hormonal contraception in the UK

OBJECTIVE

To estimate the relative cost effectiveness for women aged > or =30 years, starting long-term hormonal contraception with either levonorgestrel intrauterine system (Mirena), etonogestrel subdermal implant (Implanon) or medroxyprogesterone acetate injection (Depo-Provera).

DESIGN AND SETTING

This was a modelling study, performed from the perspective of the UK NHS, of contraceptive services supplied by a general practitioner. STUDY PARTICIPANTS AND INTERVENTIONS: A dataset was created from the General Practice Research database (GPRD) comprising 16 835 women aged > or =30 years who received levonorgestrel intrauterine system (n = 6080), etonogestrel subdermal implant (n = 277) or medroxyprogesterone acetate injection (n = 10 478) for their long-term contraception between 1997 and 2002.

METHODS

Contraception-related healthcare resource utilisation values and contraception continuation rates were obtained from the GPRD. The incidence of pregnancy associated with each contraceptive was obtained from the published literature. By combining the GPRD dataset with published clinical outcomes, a decision model was constructed. This was used to estimate the expected annualised direct healthcare costs and consequences of the provision of each type of contraception per woman-year in pounds sterling (pound) at 2002/03 prices.

RESULTS

Our model suggests that starting long-term contraception with levonorgestrel intrauterine system or etonogestrel subdermal implant instead of medroxyprogesterone acetate injection is a dominant strategy from the UK NHS perspective. In contrast, starting long-term contraception with etonogestrel subdermal implant instead of levonorgestrel intrauterine system is likely to be the least cost-effective option, since it would lead to an additional cost for each additional avoided pregnancy (pound 21,000).

CONCLUSION

Long-acting reversible hormonal contraception has the benefit of being extremely effective (>99%), and not reliant on patient compliance nor dependent on correct usage. The relative cost effectiveness of using any one contraceptive should be considered in the light of the additional clinical benefits it may confer, user acceptability, QOL, past medical history and the estimated cost of an unintended pregnancy. Choice of contraception is essential to meet diverse user needs and preferences that may change with the user's stage of life. Only by offering choice will the maximum number of women be protected and therefore the greatest savings to the health service be gained.

Using disease registries for pharmacoepidemiological research: a case study of data from a cystic fibrosis registry

BACKGROUND

The Epidemiologic Registry of Cystic Fibrosis (ERCF) was a multicentre, longitudinal follow-up project of cystic fibrosis patients enrolled at some 200 centres in nine European countries between 1994 and 1999.

PURPOSE

We aimed to assess and improve the quality of a subset of data from the ERCF relating to seven English centres (1184 patients), prior to using the data for a long-term cost-effectiveness analysis of dornase alfa (Pulmozyme). Specifically we wanted to assess the completeness and accuracy of the data and the comparability of cases across centres.

METHODS

We used a subset of ERCF data relating to seven UK cystic fibrosis (CF) centres. Following initial data editing, key variable data from a sample of patients from five centres were subjected to a detailed verification of ERCF data against original data sources available in the centres. Disagreements between ERCF reports and original data sources were identified and corrected in the study dataset. In addition, centre staff were questioned about relevant clinical and recording practices.

RESULTS

Thanks to detailed routine data checking procedures on key variables operated by the ERCF, the rates of disagreement between ERCF data and original data as identified in our verification process on the assessed variables are generally low (0.4-3.7%). Some outcome variables (deaths, hospitalisations) seem to be under-reported by some centres. Episodes of pulmonary exacerbation are difficult to identify and also to verify. Twenty-four patients were registered twice (consecutively in two different centres). There were some differences between centres in their interpretation of recording rules.

CONCLUSIONS

Researchers seeking to use disease registry data should consider detailed data quality review processes. Apart from data accuracy, reliable definitions of both critical events as well as their timing are important. The degree of under-reporting, particularly of outcome variables, should be estimated. Information on local clinical and reporting practices is necessary to interpret multi-centre data. Data protection issues may limit the possibilities for detailed data quality assessments of secondary data, as does the accessibility of original data for verification purposes. Our experiences and recommendations may be valuable for those intending to use disease registry data as well as those devising and operating such registries.